Best AI Monitoring Platforms 2026 — Expert Guide

Why You Can Trust GetAIGovernance + Our Research

Every vendor on this page was evaluated against the same criteria using public documentation, funding disclosures, product announcements, customer evidence, and independent industry recognition. No vendor paid to be included. Vendor selection reflects our independent editorial assessment of each platform's fit, depth, and differentiation within the AI monitoring category. All sources are listed at the bottom of this article.

⚠ BE AWARE: THE NUMBER RANKINGS "#1, #2..." DO NOT MEAN ONE COMPANY IS BETTER THAN ANOTHER. COMPANIES ARE LISTED IN ALPHABETICAL ORDER. ONE PLATFORM IS NOT BETTER BECAUSE OF FUNDING SIZE OR YEARS IN OPERATION. EACH PLATFORM ADDRESSES A SPECIFIC SIGNAL CATEGORY — THE RIGHT CHOICE DEPENDS ON THE PROBLEM YOU ACTUALLY NEED TO SOLVE.

Most organizations deploying AI in production think they have a monitoring program. They have dashboards. They have alert configurations. What they rarely have is a clear answer to who is supposed to act on what those dashboards surface, within what timeframe, and where the evidence of that action goes.

That's a governance problem, but it starts with a signal problem. Different parts of an AI system produce fundamentally different kinds of signals, and no single platform covers all of them with equal depth. A platform built to detect feature drift in traditional ML models has nothing useful to say about whether an autonomous agent invoked a tool it was never designed to invoke. A platform that tracks token spend won't catch a hallucination. A platform that scores output quality won't tell you whether your ingestion pipeline dropped upstream records before the model ever saw them.

Buyers who treat AI monitoring as a single category end up comparing platforms that address completely different problems. They select one that covers their most obvious gap and find out too late that the signal categories they actually needed to close were elsewhere. This guide organizes the leading AI monitoring platforms by the specific signal layer they address — aligned to the signal framework in GAIG's AI Monitoring Signals Explained. The goal is to show which platform addresses which problem, so procurement decisions are based on actual coverage rather than vendor marketing.

Three platforms in this guide have been acquired: Langfuse by ClickHouse (January 2026), W&B Weave by CoreWeave (May 2025), and SUPERWISE by Blattner Technologies (January 2023). All three continue operating as distinct products under their respective parent companies, with unchanged licensing, unchanged product roadmaps, and active development. Their acquisitions are disclosed in their individual entries. This differs from the Galileo situation, where Cisco absorbed the platform directly into Splunk Observability Cloud — which is why Galileo does not appear in this article.

What AI Monitoring Platforms Actually Do

AI monitoring platforms collect telemetry from deployed AI systems and surface signals that indicate whether those systems are performing as expected. They tell you when a model's outputs start degrading, when a user interaction produces something unusual, when a prompt pattern looks like an injection attempt, when an agent makes a tool call it wasn't authorized for. What they don't do is write policies, run approval workflows, or enforce access controls — those functions belong to the governance and security layers respectively.

The fourteen signal categories in GAIG's framework span the full production AI stack, from the technical performance of models and infrastructure to the behavior of autonomous agents and the audit trails regulators examine. Most platforms in this market cover a meaningful subset of those categories well and leave others to adjacent tools. Understanding which platform covers which signal categories is the primary decision an enterprise AI team needs to make before selecting a monitoring vendor — and the purpose of organizing this guide the way it's organized.

How We Evaluated These Platforms

1. Signal Category Depth: Does the platform's core product actually monitor the signals the category requires, or is the capability a secondary feature added onto a different primary function?

2. Production Evidence: Has the platform been deployed at meaningful scale in enterprise environments, with named customers and specific outcomes to show for it?

3. Independent Validation: What analyst recognition, named enterprise customers, or third-party research validates the platform's claims?

4. Agentic Coverage: Does the platform handle autonomous agent monitoring, or was it built for traditional ML models and extended to cover agents in name only?

5. Buyer Fit: What team inside an organization evaluates and operates this platform day to day?

The AI Monitoring Platforms: A Quick Overview

Several platforms in this guide appear across more than one signal category. Arize AI covers three categories (Performance + Drift, Model Behavioral Drift, Input/Prompt Signals). Braintrust covers two (User Behavior, Feedback). Coralogix covers two (Anomaly, Infrastructure/Pipeline). Fiddler AI covers three (Output Quality, Business Impact, Governance Evidence). Levo.ai covers two (Anomaly, Infrastructure/Pipeline). W&B Weave covers two (Cost and Resource, Governance Evidence). These overlaps reflect genuine product depth across multiple signal types rather than padding. Platforms are listed alphabetically, not by rank.

Arize AI — Best for Production LLM and Agent Observability at Enterprise Scale

The Enterprise Standard for OpenTelemetry-Native AI Monitoring

Choose Arize AI if: you have LLMs, agents, or traditional ML models running in production and you need a monitoring platform that tracks performance, drift, and prompt-level signals at scale, with an open-source path for teams that want self-hosted tracing and an enterprise tier for teams that need compliance, collaboration, and production reliability.

Founded: 2020

HQ: Berkeley, CA

Employees: 196 (as of May 2026)

Funding: $131M total (Series C, February 2025)

Recognition: Named customers include DoorDash, Instacart, Reddit, Uber, Booking.com, Roblox, Air Canada, Cohere, Microsoft, Siemens, TripAdvisor, Flipkart, the U.S. Navy, and PepsiCo. Processes 1 trillion spans and runs 50+ million evaluations monthly across its customer base.

Arize AI maintains two products that address different parts of the monitoring stack. Arize Phoenix is the open-source observability library, built on OpenTelemetry and self-hostable on a local machine, in Docker, or in the cloud, with over 10,200 GitHub stars and 26 million SDK installs per month. Arize AX is the enterprise production monitoring platform, which extends Phoenix's tracing and evaluation capabilities to fleet-scale deployments with role-based access control, compliance certifications (SOC 2, GDPR, HIPAA), real-time alerting through PagerDuty, OpsGenie, and Slack, and Alyx, an AI debugging assistant that answers natural language questions about agent performance directly inside the platform. The January 2026 Phoenix CLI release extended terminal access to traces, datasets, and experiments for teams using Claude Code and Cursor, reflecting how Arize's development has kept pace with how working engineers actually interact with AI systems.

Arize covers three signal categories in this guide. For Performance + Drift, the platform tracks session, trace, and span-level signals across LLMs, agents, and traditional ML models simultaneously — detecting latency degradation, accuracy drops, and behavioral changes before they reach business-visible thresholds. For Model Behavioral Drift, AX applies statistical drift detection with configurable sensitivity thresholds and root cause surfacing that identifies which features, prompts, or data changes drove a degradation. For Input/Prompt Signals, the platform's trace-level instrumentation captures incoming prompt structure, token volume, and input anomalies across production traffic, surfacing patterns that indicate prompt injection attempts, distribution shifts, or unauthorized use patterns. The Evaluator Hub, introduced in early 2026, is a centralized system for creating, versioning, and reusing evaluators across tasks with commit-level version control — which means the criteria used to evaluate a model's outputs in development can be applied exactly to production traffic without manual reconfiguration.

The platform's OpenTelemetry foundation is its primary architectural advantage. Arize's instrumentation follows open standards rather than proprietary formats, which means traces can be routed to other observability backends if needed and organizations don't build a dependency on Arize-specific data formats. The OpenInference instrumentation library supports LangChain, LlamaIndex, OpenAI Agents SDK, CrewAI, NVIDIA NeMo Agent Toolkit, PydanticAI, and dozens of other frameworks without requiring teams to restructure their agent code. AWS and Azure marketplace availability, plus AX's Azure-native deployment option, puts the platform inside the procurement infrastructure most large enterprises already use.

The U.S. Navy and PepsiCo as named production customers, alongside DoorDash, Instacart, Reddit, and Uber, provide a credibility baseline that few AI monitoring platforms can match. The $131M in funding from a February 2025 Series C gives Arize the runway to maintain and expand the integration footprint that makes Phoenix's open-source adoption valuable to paying enterprise customers — which is the dynamic the whole open-core model depends on.

✓ What We Like

• Genuine open-source commitment: Phoenix has over 10,200 GitHub stars and 26 million SDK installs per month, which reflects real adoption rather than a token open-source marketing gesture.

• OpenTelemetry-native architecture: Vendor-agnostic tracing prevents lock-in and lets teams route data to other backends if their needs change.

• Three signal categories in one platform: Performance + Drift, Model Behavioral Drift, and Input/Prompt Signals covered with genuine depth rather than surface-level feature checkboxes.

• Evaluator Hub: Version-controlled evaluators that apply the same criteria from development through production, closing the gap between testing and monitoring.

• Named enterprise customer breadth: The U.S. Navy, PepsiCo, DoorDash, Reddit, Uber, and Microsoft on the same customer list signals the platform works across very different organizational contexts.

• CLI integration for AI coding environments: The January 2026 terminal access release for Claude Code and Cursor reflects where working engineers actually operate.

⚠ What to Know

• AX starts around $50,000 per year, which means the enterprise tier is a meaningful budget commitment for smaller teams who may be better served starting with Phoenix's self-hosted option.

• Arize monitors AI systems — it doesn't build them. Teams need separate platforms for agent design, knowledge base management, and deployment orchestration.

• The depth of AX's compliance and collaboration features requires configuration investment; teams expecting to plug in and get value immediately should plan for an implementation phase.

• Alyx, the AI debugging assistant, is a newer capability and production feedback on its accuracy in complex multi-agent debugging scenarios is still accumulating.

Signal Categories Covered

Performance + Drift
Model Behavioral Drift
Input/Prompt Signals
Agent Behavior Tracing
Evaluation and Scoring

Regulatory Frameworks

SOC 2
GDPR
HIPAA
NIST AI RMF

Best For

• ML engineering and platform teams: Organizations running LLMs and agents at production scale who need OpenTelemetry-native tracing with genuine enterprise reliability and compliance certifications.

• Teams starting with open-source: Data science and engineering teams that want a self-hosted monitoring foundation with the option to graduate to enterprise managed infrastructure as production scale grows.

• Organizations across AI types: Companies running traditional ML models, generative AI applications, and autonomous agents simultaneously, where a unified monitoring platform reduces the overhead of maintaining separate tools for each.

Pricing: Phoenix is free and open-source. AX starts around $50,000/year; contact Arize directly for enterprise pricing or request a match through GetAIGovernance.net.

Arthur AI — Best for Agent Operational Health and Regression Detection

The Platform That Caught a GPT-5 Regression Before It Reached Users

Choose Arthur AI if: you have autonomous agents deployed in production and need a monitoring layer that tracks tool call accuracy, decision quality, and behavioral consistency across the full agent development lifecycle — with the ability to detect regressions before they reach end users.

Founded: 2018

HQ: New York, NY

Employees: ~51

Funding: $63M total (Series B, September 2022, led by Acrew Capital and Greycroft with Index Ventures and Work-Bench)

Recognition: Named customers include Upsolve (personal finance) and Expel (cybersecurity, 50% reduction in ML monitoring time). AWS Marketplace and Google Cloud Marketplace availability. ADLC Methodology introduced and adopted across enterprise customers in 2025.

Arthur AI's Agent Discovery and Governance (ADG) platform, launched in December 2025 on AWS Marketplace and extended to Google Cloud Marketplace in January 2026, is the operational monitoring layer for agentic AI that the company built after spending years on traditional ML observability. The ADG platform treats agents as first-class production systems requiring continuous evaluation: it discovers all agents running in an environment, tracks their behavior through the full agent development lifecycle (which Arthur calls the ADLC), applies automated evaluations across prompt quality, tool selection accuracy, tool invocation parameters, and output groundedness, and surfaces behavioral anomalies before they compound into user-visible failures. The April 2026 platform update added unified evaluators, automated 24-hour compliance checks, configurable trace retention policies, and an Engine Chatbot AI assistant for debugging agent behavior in plain language.

The Upsolve case study is the most specific production evidence in this article. Upsolve, a nonprofit that provides free legal tools for debt relief, deployed autonomous agents in a financial services context where incorrect agent outputs carry direct consequences for users — incorrect debt advice delivered at scale. Arthur detected a critical GPT-5 regression in Upsolve's deployment before it reached any end users. The regression involved behavioral drift in how the agent handled specific financial scenarios, caught through Arthur's continuous ADLC evaluation layer rather than discovered through user complaints. For a platform positioned specifically at agent operational monitoring, that case study is precisely the proof that matters.

Arthur's ADLC methodology — which treats agent governance as a continuous lifecycle from development through post-deployment monitoring rather than a pre-launch checklist — produces evaluation checkpoints at every stage. Automated evals run continuously across production traffic using LLM-as-a-judge templates that score outputs for hallucination, relevance, and tool call accuracy. Teams can trigger re-evaluation when a model update, prompt change, or new tool integration happens, comparing current agent behavior against a verified baseline before and after the change. The platform's open-source Arthur Bench library handles offline LLM evaluation, while the production ADG platform handles live operational monitoring — the two are designed to work together as part of the same ADLC workflow rather than as separate tools requiring separate configuration.

The Expel deployment cut ML monitoring time by 50 percent while improving coverage, which speaks to the platform's architecture: Arthur monitors continuously in the background rather than requiring manual evaluation runs, and its alert routing through Slack and PagerDuty means teams don't check a dashboard to learn something is wrong. Arthur's limitation in this guide is scope — the ADG platform is strongest at operational health monitoring for agents specifically and does not cover the cost attribution, governance evidence, or identity monitoring functions that other platforms in this article address.

✓ What We Like

• Documented regression detection in production: The Upsolve GPT-5 case study is specific, verifiable, and directly relevant to what agent operational monitoring is supposed to accomplish.

• ADLC methodology: A structured lifecycle framework for agentic AI that produces evaluation checkpoints at every stage rather than treating monitoring as a post-deployment activity.

• Automated ADLC evaluations: Continuous evals across tool selection, tool invocation parameters, and output quality without requiring manual test runs after each deployment change.

• Cloud marketplace availability: AWS and Google Cloud Marketplace deployments simplify procurement for organizations already committed to those environments.

• Open-source Arthur Bench: Offline LLM evaluation library on GitHub gives teams independent access to evaluation tooling without committing to the enterprise platform.

• 50% monitoring time reduction at Expel: A named security customer with a specific and credible efficiency outcome.

⚠ What to Know

• Arthur is an agent operational monitoring platform, not a full AI monitoring suite. Cost tracking, governance evidence generation, and identity monitoring require separate platforms.

• The ADG platform launched in December 2025, so enterprise deployment evidence is still developing relative to Arthur's longer-standing ML monitoring capabilities.

• At 51 employees and $63M raised, Arthur is smaller than several platforms in this guide — vendor stability is a relevant consideration for organizations planning multi-year monitoring infrastructure.

• Deep evaluation configuration requires ML engineering involvement; teams without dedicated AI engineering resources should plan for that during implementation.

Signal Categories Covered

Agent Operational Monitoring
Agent Behavior Tracing
Performance + Drift
Output Quality Evaluation

Regulatory Frameworks

EU AI Act (Annex III)
NIST AI RMF
OWASP Top 10 for LLM Applications

Best For

• Organizations with agents in financial, legal, or healthcare contexts: Teams where incorrect agent behavior has direct consequences for users, and where catching regressions before they reach production is a legal or reputational requirement.

• ML and AI engineering teams: Organizations with dedicated engineering capacity to configure and operate continuous evaluation workflows across the agent development lifecycle.

• Cloud-native AI deployments: Teams on AWS or Google Cloud who want marketplace-integrated deployment without standing up separate monitoring infrastructure.

Pricing: Not publicly listed. Contact Arthur AI directly or request a match through GetAIGovernance.net.

Braintrust — Best for User Behavior Monitoring and Evaluation Feedback Loops

The Observability Platform Built for AI Products That Have to Stay Reliable

Choose Braintrust if: you're running AI features inside a product that real users depend on — and you need to monitor how users interact with those features, score the quality of responses, and feed those signals back into continuous evaluation cycles without rebuilding your data infrastructure to do it.

Founded: 2023

HQ: San Francisco, CA

Funding: $121M total — $80M Series B closed February 17, 2026, led by ICONIQ at $800M valuation, with Andreessen Horowitz, Greylock, Elad Gil, and Basecase Capital returning

Recognition: Named customers include Notion, Replit, Cloudflare, Ramp, Dropbox, Stripe, Zapier, Airtable, and Instacart.

Braintrust's observability platform is organized around a specific problem: AI products in production change constantly — prompts get updated, models get swapped, retrieval logic gets adjusted — and without a way to track those changes against real user interaction data, teams discover quality problems through support tickets and churn rather than through monitoring. The platform collects and stores every trace from every user interaction through Brainstore, a purpose-built OLAP database that Braintrust built specifically because existing databases hit performance bottlenecks at the scale of AI trace data. Brainstore handles hundreds of megabytes per agent interaction across production deployments at Notion and Cloudflare without the query slowdowns that sent earlier customers away from general-purpose time-series databases.

User Behavior monitoring in Braintrust means tracking how users interact with AI features over time — which prompts they send, which responses they accept or correct, where interaction patterns shift in ways that indicate user friction or model quality degradation. The platform's trace viewer captures session-level and interaction-level detail, letting teams drill into specific user sessions where something went wrong and work backward through the chain of prompts, retrievals, and model calls that produced the outcome. Feedback signals feed into the same system: when users correct a model output, rate a response, or abandon an interaction, those signals get logged as labeled data that teams can use to build evaluation datasets for the next model or prompt change.

The Loop AI assistant analyzes production traces at scale to suggest prompt improvements, identify hallucination patterns, and create evaluation datasets from real production failures — automating the part of the feedback cycle that usually requires a data scientist to manually triage logs and extract meaningful examples. This closes the loop Braintrust CEO Ankur Goyal described when explaining why he built the company: the painful process of building evaluation tooling from scratch at every company that runs AI in production, because no general-purpose monitoring tool captures the right signals for AI quality.

Braintrust covers both User Behavior and Feedback signal categories because the platform was built around the idea that those two things are inseparable. User behavior data is only meaningful if you can act on it, and acting on it requires feeding it back into evaluation cycles and prompt engineering. The $80M Series B at an $800M valuation from ICONIQ, a firm that backed Salesforce, Snowflake, and Datadog in their growth phases, reflects a bet that Braintrust occupies the same infrastructure position in AI that those platforms occupy in their respective categories.

✓ What We Like

• Brainstore database: Purpose-built OLAP architecture handles the scale of AI trace data that general-purpose databases struggle with, making production-scale user behavior monitoring actually fast.

• Named customer breadth: Notion, Cloudflare, Ramp, Dropbox, Stripe, and Instacart represent a diverse set of AI-native production deployments across different product types and scale profiles.

• Feedback-to-evaluation loop: User correction signals flow directly into evaluation datasets, closing the gap between monitoring what happened and improving what happens next.

• Loop AI assistant: Automated pattern detection and prompt suggestion across production traces reduces the manual triage work that otherwise falls to data scientists after every quality incident.

• $800M valuation at Series B: ICONIQ's investment signals the kind of institutional confidence that enterprise procurement teams look for when evaluating vendor stability.

⚠ What to Know

• Braintrust was founded in 2023, making it one of the youngest platforms in this guide despite the substantial funding round.

• The platform is strongest for teams running AI features inside consumer or enterprise products — traditional ML model monitoring for batch inference or tabular models is less central to Braintrust's current architecture than LLM and agent observability.

• Maximum value requires integrating Braintrust tracing into existing product code; the benefits are proportional to how completely teams instrument their AI interactions.

• Post-Series B product plans were announced at the Trace user conference but not yet fully detailed in public documentation at the time of this guide's publication.

Signal Categories Covered

User Behavior
Feedback
Output Quality Evaluation
Agent Trace Collection

Regulatory Frameworks

SOC 2
GDPR

Best For

• AI-native product companies: Teams where AI features are central to the product experience and user behavior monitoring feeds directly into product quality decisions.

• Engineering teams running high-volume LLM interactions: Organizations where traditional monitoring databases have already shown performance limits on AI trace data at production scale.

• Teams with active prompt engineering cycles: Companies that update prompts, models, and retrieval logic frequently and need to know immediately how those changes affect real user interactions.

Pricing: Free tier available. Pro and Enterprise pricing is custom. Contact Braintrust directly or request a match through GetAIGovernance.net.

Coralogix — Best for Unified Infrastructure and AI Anomaly Detection

Enterprise Observability Extended Across AI Workloads and Traditional Infrastructure

Choose Coralogix if: you want AI workload monitoring inside the same observability platform that already handles your logs, metrics, and traces for traditional infrastructure — with an AI investigator agent that can answer production questions in plain language rather than requiring engineers to build dashboards.

Founded: 2014

HQ: San Francisco, CA (founded in Israel)

Funding: $550M total — $200M Series F closed June 3, 2026, co-led by Advent, CPPIB, and Greenfield Partners at a $1.6B valuation

Recognition: 5,000+ customers including IBM, Tradeweb, and JFrog. Acquired Aporia (AI observability and guardrails) in December 2024. FedRAMP Moderate authorization in progress. AWS strategic partnership.

Coralogix's observability platform reached enterprise AI monitoring through the December 2024 acquisition of Aporia, an AI observability and guardrails company that had previously counted Lemonade, DoorDash, MunichRe, Bosch, and Sixt among its customers. The acquisition created Coralogix AI, a dedicated research center led by Aporia's co-founders Liran Hason and Alon Gubkin, with an explicit mandate to solve AI monitoring problems including transparency, security, monitoring, governance, and control. Olly, Coralogix's AI investigator agent launched in general availability in December 2025, operates on top of this unified data foundation: it identifies root causes, surfaces anomalies, detects performance degradations, and recommends remediation steps in response to natural language questions from engineers and operations teams who don't want to build dashboards to get answers.

For Anomaly signals, Coralogix's schema-free telemetry data lake ingests logs, metrics, traces, and AI workload data simultaneously, which means anomaly detection runs across the full production stack rather than only within AI-specific monitoring silos. When an LLM starts producing unusual output patterns at the same time that infrastructure latency spikes, Coralogix connects those signals rather than presenting them as two separate incidents in two separate tools. For Infrastructure and Pipeline signals, the platform tracks data pipeline health, compute resource behavior, model serving infrastructure, and the dependencies between AI workloads and the infrastructure they run on — the layer where problems often originate before they appear in model output quality or user experience.

The June 3, 2026 Series F, which closed 27 days before this article's publication, is the most recent major funding event among all twelve platforms in this guide. The $200M raise at a $1.6B valuation, with Advent International as lead investor, reflects continued institutional confidence in the observability-as-intelligence thesis that CEO Ariel Assaraf articulated directly: AI systems are becoming operational participants in observability, not just subjects of it, and Coralogix's architecture was built for exactly that transition before the AI agent market arrived to validate the bet. The 5,000 customer base, FedRAMP Moderate authorization in progress, and AWS strategic partnership give enterprise procurement teams the scale evidence and compliance trajectory they need for multi-year commitments.

Coralogix's limitation in this guide is specificity: the platform is strongest at broad infrastructure and AI workload observability. Teams whose primary need is deep LLM-specific evaluation — scoring model outputs for hallucination, tracking prompt-level quality metrics, or building evaluation datasets from production traces — will find the Aporia-derived AI capabilities useful but less specialized than purpose-built LLM evaluation platforms like Arize or Braintrust. Coralogix's value is the unification of AI and traditional infrastructure monitoring under one platform, not the depth of any single AI-specific signal type.

✓ What We Like

• Olly AI investigator: Natural language production queries in a general availability product, not a beta feature — engineers can ask what changed and get an answer rather than building a dashboard to find out.

• Fresh $200M Series F: The June 3, 2026 raise at $1.6B valuation is the most recent major funding event in this entire guide and signals strong investor confidence at an important stage.

• Aporia AI capabilities: The December 2024 acquisition brought AI guardrails, hallucination detection, and AI-SPM into the platform with a dedicated research team behind them.

• 5,000+ enterprise customers: IBM, Tradeweb, and JFrog as named customers provide credibility across financial services, technology, and enterprise infrastructure categories.

• Schema-free telemetry architecture: Full-fidelity ingestion across logs, metrics, traces, and AI workload data without the storage tradeoffs that constrain other observability platforms.

• FedRAMP Moderate authorization in progress: Public sector and regulated industry buyers already have a path to procurement without waiting for authorization.

⚠ What to Know

• Coralogix's AI monitoring depth comes from the Aporia acquisition; teams that need specialized LLM evaluation capabilities — scoring, evaluation datasets, prompt management — should evaluate whether the Aporia-derived features meet their requirements or whether they need a purpose-built LLM evaluation platform alongside Coralogix.

• Enterprise pricing scales with data volume ingested; teams with high-volume AI workloads should model cost carefully before committing.

• Olly reached general availability in December 2025; production evidence on AI-specific root cause analysis accuracy is still accumulating relative to the platform's longer infrastructure monitoring history.

• The platform is genuinely large and feature-rich; getting full value requires implementation investment and ongoing configuration, not a lightweight plug-in deployment.

Signal Categories Covered

Anomaly Detection
Infrastructure / Pipeline
Performance Monitoring
AI Guardrails (Aporia)AI-SPM

Regulatory Frameworks

SOC 2
GDPR
HIPAA
FedRAMP Moderate (in progress)

Best For

• Large enterprises with existing observability investments: Organizations that want AI workload monitoring integrated into the same platform handling traditional infrastructure, rather than maintaining separate AI-specific monitoring alongside existing tools.

• Operations and SRE teams: Engineers who need production answers at machine speed without building dashboards for every new AI workload they're responsible for.

• Organizations scaling beyond startup-tier monitoring: Companies growing past the point where open-source tools plus custom dashboards are sufficient, into environments where unified telemetry across AI and infrastructure is an operational necessity.

Pricing: Custom enterprise pricing. Contact Coralogix directly or request a match through GetAIGovernance.net.

Evidently AI — Best Open-Source Option for Model Drift Detection and ML Monitoring

The Self-Hosted Choice for Teams That Want Full Control Over Drift Methodology

Choose Evidently AI if: your team needs open-source, self-hosted drift detection and ML monitoring that you can extend and customize for your specific models, features, and data distributions — without paying for an enterprise monitoring platform and without building your own statistical evaluation tooling from scratch.

Founded: 2020

HQ: San Francisco, CA

Employees: ~6

Funding: $1.15M (Y Combinator, Fly Ventures, Nauta Capital, Davidovs Venture Collective, PLF)

Evidently is an open-source Python library for evaluating, testing, and monitoring machine learning models in production. It generates visual drift reports and test suites from model predictions and reference data, covering data drift, prediction drift, target drift, data quality checks, and model performance metrics across a range of statistical tests configurable by the user. The library runs locally, in Jupyter notebooks, Docker containers, or cloud environments, produces reports in HTML and JSON, and integrates with Grafana for dashboard-based monitoring. Cloud-hosted Evidently is available at $500 per month for teams that want the evaluation infrastructure managed rather than self-deployed.

Evidently's place in this guide is as the strongest open-source second option at the Performance + Drift signal category alongside Arize AI. The distinction matters for a specific buyer profile: data science teams that want to run drift detection on their own terms, with full visibility into and control over the statistical tests being applied, without depending on a vendor's proprietary evaluation methodology. Evidently's tests are configurable and transparent — teams can inspect exactly what statistical test ran, against what threshold, with what result — which makes the platform useful for teams that need to explain their drift detection methodology to auditors or regulators.

The funding figure — $1.15 million from Y Combinator and European venture firms — reflects a lean team (six employees by PitchBook's count) operating an open-source project with a commercial cloud tier alongside it. This is genuinely a small company, and procurement teams weighing vendor stability should account for that. The open-source library's independence from the commercial product mitigates some of that risk: the codebase is public, permissively licensed, and in principle can be maintained by the community even if the commercial entity changes. Evidently is the right choice for teams whose primary need is the open-source library itself and whose budget or vendor stability requirements don't demand an enterprise platform with dedicated support.

✓ What We Like

• Full statistical transparency: Every drift test is configurable and explainable — teams can see exactly what ran, with what parameters, against what reference dataset.

• Y Combinator backing: Institutional validation from the accelerator that funded Stripe, Airbnb, and Dropbox gives the project credibility beyond what the funding figure alone suggests.

• Self-hosted flexibility: Runs locally, in notebooks, in Docker, or in the cloud — no mandatory vendor dependency for the core monitoring function.

• Open-source codebase: Public and permissively licensed, which gives teams a long-term option to maintain or fork the library if the commercial trajectory changes.

• $500/month cloud entry point: Accessible pricing for the managed cloud tier compared to enterprise platforms that start at tens of thousands of dollars annually.

⚠ What to Know

• Six employees and $1.15M raised — the smallest funded company in this guide by a significant margin. Vendor stability is a genuine procurement consideration for any multi-year monitoring commitment.

• Evidently covers model drift and data quality monitoring; it doesn't cover LLM-specific evaluation, agent monitoring, cost tracking, or governance evidence generation.

• Self-hosted deployments require engineering effort to maintain and update; teams expecting low operational overhead should consider whether the managed cloud tier or a larger platform better fits their staffing.

• Independent third-party analyst recognition is limited in public documentation; enterprise procurement teams should conduct direct technical evaluation rather than relying on analyst market placement.

Signal Categories Covered

Performance + Drift
Model Behavioral Drift
Data Quality Monitoring

Best For

• Data science teams with engineering capacity: Organizations that can self-host and maintain monitoring infrastructure and want full control over drift detection methodology without paying for an enterprise platform.

• Teams needing explainable drift detection: Organizations that have to justify their drift monitoring approach to auditors, regulators, or internal governance functions and need statistical tests that are transparent and configurable.

• Organizations evaluating before committing: Teams that want to run drift monitoring on real data before committing budget to an enterprise platform.

Pricing: Open-source library is free. Cloud tier starts at $500/month. Contact Evidently AI directly or request a match through GetAIGovernance.net.

Fiddler AI — Best for Output Quality Monitoring and Governance Evidence Generation

The Control Plane for AI Agents and the Output Quality Layer Below It

Choose Fiddler AI if: you need a platform that monitors the quality of AI outputs in production — scoring them for hallucination, bias, toxicity, and accuracy — while also generating the auditable governance evidence that regulated industries and EU AI Act compliance requirements demand.

Founded: 2018

HQ: Palo Alto, CA

Funding: $100M total (Series C, January 27, 2026, led by RPS Ventures with Lightspeed, Lux Capital, Insight Partners, Mozilla Ventures)

Recognition: Named #1 in AI Agent Security and Risk Management by CB Insights. AWS Pattern Partners status. 4x revenue growth in the last 18 months. Fortune 500 customers across healthcare, financial services, and insurance. Nielsen CEO Karthik Rao publicly cited Fiddler as "fundamental to our AI strategy."

Fiddler AI's Control Plane is the company's core architectural bet: that as AI systems become compound — agents chaining model calls, tool calls, and retrieval operations across a single user interaction — the monitoring and governance infrastructure needs to sit above all of those individual components and provide visibility across the full chain. The Control Plane delivers standardized telemetry, continuous output quality evaluation, enforceable policy, and auditable governance documentation across traditional ML models, generative AI applications, and autonomous agents simultaneously. The Fiddler Trust Service handles real-time output quality control: proprietary trust models score model outputs for accuracy, safety, and compliance in under 10 milliseconds, without routing evaluation through third-party LLMs, which matters both for latency and for keeping sensitive data inside enterprise boundaries.

Output Quality monitoring is where Fiddler has built the deepest independent validation in this guide. The platform evaluates hallucination, bias, toxicity, groundedness, relevance, and custom quality metrics across production traffic at the pace at which that traffic actually arrives. Karthik Rao's public statement as Nielsen CEO — that Fiddler is fundamental to the company's AI strategy and that he's looking forward to the control plane rollout — is the kind of on-record customer endorsement that carries real weight in enterprise procurement. The CB Insights ranking as the number one company in AI Agent Security and Risk Management, combined with AWS Pattern Partners status, provides two independent third-party signals that the Control Plane meets enterprise quality thresholds rather than only marketing claims.

Business Impact signals represent the most underserved category in this entire guide. No platform in the market currently provides a mature, dedicated Business Impact monitoring function that directly ties AI output quality to measurable commercial outcomes — task success rates, conversion attribution, and revenue impact tied to specific agent actions. Fiddler's Control Plane comes closest, through its ability to connect output quality scoring to downstream business metrics within a single observability layer, and through the Trust Service's capacity to track when quality degradation begins affecting user outcomes. The write-up is honest about the limitation: Fiddler addresses Business Impact as a downstream consequence of output quality monitoring, not as a dedicated primary signal with the same depth as its output quality capabilities.

Governance Evidence is the third category Fiddler addresses in this guide, and it's the one where the Control Plane's architecture is most directly relevant. The platform generates immutable audit trails of model behavior, evaluation scores, policy enforcement decisions, and governance actions across the AI lifecycle. For organizations subject to EU AI Act Article 13 transparency obligations, SR 26-2 ongoing monitoring requirements, or internal audit requirements, these records provide the evidence chain that a governance review needs: what the system did, how it was evaluated, what policy applied, and what the outcome was. Fiddler's ADLC (AI Delivery Lifecycle) documentation maps directly to the evidence types those frameworks require.

✓ What We Like

• Fiddler Trust Service: Proprietary trust models evaluate output quality in under 10 milliseconds without routing data through third-party LLMs — fast enough for production use without data sovereignty compromise.

• Nielsen CEO quote: Public, on-record enterprise customer endorsement from a Fortune 500 CEO is the kind of validation most monitoring platforms never produce.

• CB Insights #1 ranking: Independent analyst recognition specifically for AI Agent Security and Risk Management.

• 4x revenue growth in 18 months: Commercial momentum that reflects actual market adoption rather than funded growth alone.

• Control Plane architecture: Unified observability across ML models, GenAI, and agents in one system rather than three separate monitoring tools.

• AWS Pattern Partners status: Enterprise procurement validation from AWS's most rigorous partner tier.

⚠ What to Know

• Fiddler covers three signal categories in this guide (Output Quality, Business Impact, Governance Evidence); organizations with primary needs in only one of those areas should evaluate whether the full platform's cost and complexity are justified.

• Business Impact monitoring in Fiddler is downstream of output quality scoring rather than a standalone signal category with dedicated monitoring infrastructure — organizations needing dedicated business impact attribution should evaluate whether that distinction matters for their use case.

• The Control Plane's full capabilities are still rolling out as of early 2026; some features described in the January Series C announcement are on the near-term roadmap rather than fully GA.

• Custom enterprise pricing only; no public pricing tiers for self-serve evaluation.

Signal Categories Covered

Output Quality
Business Impact
Governance Evidence
Performance + Drift
Bias Detection
Hallucination Scoring

Regulatory Frameworks

EU AI Act (Article 13)
SR 26-2
HIPAA
NIST AI RMF
SOC 2

Best For

• Regulated enterprises: Healthcare, financial services, and insurance organizations where output quality monitoring and governance evidence generation are both compliance requirements and operational necessities.

• Organizations with agentic production deployments: Teams where agents are making consequential decisions across chained tool calls and where monitoring a single model's output is insufficient.

• Enterprises preparing for EU AI Act enforcement: Organizations that need audit-ready governance evidence tied to production AI behavior rather than assembled manually before an audit.

Pricing: Not publicly listed. Contact Fiddler AI directly or request a match through GetAIGovernance.net.

Langfuse — Best for Open-Source LLM Cost Tracking and Prompt Management

The Most Widely Deployed Open-Source LLM Observability Platform

Choose Langfuse if: you need to track LLM costs, token usage, and latency across production interactions, manage prompt versions and experiments, and do all of it through an open-source platform you can self-host or run in the cloud — without committing to a vendor that controls your data.

Founded: 2023

HQ: Berlin, Germany (San Francisco office)

Funding: Acquired by ClickHouse (January 16, 2026), which simultaneously raised $400M at a $15B valuation. Langfuse operates as an independent product within ClickHouse with unchanged MIT licensing and unchanged product roadmap.

Recognition: 20,000+ GitHub stars. 26 million SDK installs per month. Customers include 19 of the Fortune 50, 63 Fortune 500 companies overall, Merck, Intuit, Twilio, Khan Academy, and Dropbox.

Acquisition disclosure: ClickHouse acquired Langfuse on January 16, 2026. Langfuse's MIT license is unchanged and self-hosting remains a first-class deployment path. The full Langfuse team joined ClickHouse and continues building Langfuse with the same roadmap and the same product interfaces. Langfuse Cloud operates unchanged. This differs from acquisitions where a product is absorbed into the parent's platform — ClickHouse's business model depends on Langfuse remaining viable as an independent product, and the technical integration (Langfuse already ran on ClickHouse's database before the acquisition) makes the combination a natural deepening of an existing relationship.

Langfuse is the open-source LLM observability and evaluation platform that handles cost and resource monitoring more completely than any other open-source tool in the market. The platform tracks token usage per prompt, per session, and per user across every supported LLM provider — OpenAI, Anthropic, Google, AWS Bedrock, Azure OpenAI, and others — attributing cost and latency to specific experiments, prompt versions, and agent configurations rather than rolling everything into aggregate spend figures that don't help teams understand what's expensive or why. Merck's Chief Data and AI Officer described the platform directly: "Langfuse enables us to track every prompt, response, cost, and latency in real time, turning black-box models into auditable, optimizable assets."

The cost and resource signal category maps precisely to what Langfuse was built to solve. When Arthur AI's LinkedIn page described Uber burning through its 2026 AI coding budget in four months and Microsoft canceling Claude Code licenses over unexpected agent token consumption, those outcomes are the result of organizations deploying agents without production cost monitoring. A single agent can consume a thousand times more tokens than a one-shot query; multiplied across thousands of agents running continuously, that compounds into costs nobody budgeted for. Langfuse's version-controlled prompt management, combined with its cost attribution down to individual experiment runs, gives teams the data they need to trace an unexpected cost spike back to the prompt change, model update, or agent configuration that caused it.

The 26 million SDK installs per month and 20,000+ GitHub stars reflect the kind of community adoption that happens when a tool solves a real problem well enough that developers recommend it to each other rather than requiring a sales motion. The ClickHouse acquisition adds financial backing and deeper performance infrastructure (Langfuse already ran on ClickHouse's database in v3, migrating from Postgres specifically because of performance bottlenecks at the scale of LLM trace data), while keeping Langfuse's open-source identity intact. The Pro tier at $199 per month provides SOC 2 and ISO 27001 compliance, which matters for enterprise procurement. The Khan Academy engineering team's on-record description — that Langfuse gives developers "extremely fast feedback" on AI implementation performance and is "fundamental to how our developers understand their AI implementations" — reflects what the platform actually delivers.

✓ What We Like

• 26 million SDK installs per month: That scale of community adoption reflects genuine utility rather than marketing.

• MIT license, unchanged post-acquisition: Self-hosting remains a full-featured option without license risk.

• Version-controlled prompt management: Cost attribution that traces spending back to the exact prompt version, model, and experiment configuration that caused it.

• 19 Fortune 50 companies: Enterprise adoption at the highest market cap tier validates production-scale reliability.

• Named quotes from Merck CDAO and Khan Academy engineers: Specific, on-record customer descriptions of what the platform actually does for them.

• ClickHouse backing: Financial resources and database performance infrastructure from a $15B company, with a business model that depends on keeping Langfuse good.

⚠ What to Know

• Langfuse was acquired by ClickHouse in January 2026; while licensing and roadmap are unchanged, the long-term independence of the product depends on ClickHouse's continued commitment to the open-source model.

• Langfuse is strongest for LLM cost tracking, prompt management, and evaluation. It doesn't cover infrastructure anomaly detection, agent identity monitoring, or output quality scoring with the same depth as dedicated platforms for those categories.

• The Pro tier at $199/month includes SOC 2 and ISO 27001; Enterprise pricing for custom deployments with additional compliance requirements is custom and requires direct engagement.

• Self-hosted v3 runs on ClickHouse; teams migrating from Langfuse v2 (Postgres-based) should plan for that infrastructure change.

Signal Categories Covered

Cost and Resource
LLM Tracing and Observability
Prompt Management
Evaluation and Scoring

Regulatory Frameworks

SOC 2
ISO 27001
GDPR

Best For

• Teams prioritizing open-source and self-hosting: Organizations that need a full-featured LLM observability platform without mandatory vendor dependency, built on a codebase they can inspect, fork, and maintain.

• Engineering teams managing LLM spend: Companies where unexpected token costs from agents or high-volume LLM features are a real operational risk and prompt-level cost attribution is a necessary diagnostic tool.

• AI-native product companies at scale: Organizations like those among the 19 Fortune 50 customers that need LLM tracing infrastructure proven at massive volume.

Pricing: Open-source self-hosted: free. Cloud Pro: $199/month (SOC 2, ISO 27001). Enterprise: custom. Contact Langfuse directly or request a match through GetAIGovernance.net.

Levo.ai — Best for Kernel-Level AI and API Traffic Visibility

eBPF-Powered Pipeline Monitoring Without Proxy Deployment

Choose Levo.ai if: you need visibility into AI agent, MCP server, and API traffic at the kernel level — discovering what's actually flowing through your infrastructure without deploying proxies, modifying application code, or adding network infrastructure changes.

Founded: 2021

HQ: San Francisco, CA

Employees: ~39

Funding: $4M total (Cota Capital, Engineering Capital, Foundation Capital, Streamlined Ventures, Think + Ventures)

Recognition: Recognized by industry practitioner James Berthoty for eBPF innovation. Winner of Most Innovative Startup at FINSEC2025. Gartner Peer Insights featured reviews from enterprise deployments in fintech environments.

Note on category placement: Levo.ai straddles AI security and AI monitoring. Their core product is eBPF-powered API and AI security — discovery, testing, runtime protection, and governance enforcement across APIs, agents, MCP servers, LLMs, and vector stores. This entry evaluates their pipeline monitoring and anomaly detection capabilities specifically. Organizations evaluating Levo for security enforcement rather than monitoring should review the AI Security Platforms 2026 guide.

Levo.ai's unified AI and API security platform uses eBPF (extended Berkeley Packet Filter) technology to observe traffic at the Linux kernel level without modifying application code, deploying SDK agents, or changing network architecture. The eBPF sensor captures API endpoints, request and response schemas, authentication methods, data types, agent tool calls, and MCP server interactions by reading network traffic at the kernel, which means it sees everything that flows through the infrastructure regardless of protocol, framework, or language. In the context of AI monitoring, this produces Infrastructure and Pipeline signals that other monitoring tools miss: actual MCP server traffic, agent-to-tool calls captured at the network layer, and data flows between AI components that application-layer instrumentation doesn't reach because it wasn't installed on the systems producing those flows.

For Anomaly signals, Levo's 2026 Launch Week added MCP Discovery (continuous inventory of MCP servers across laptops, cloud, and remote environments, with risk scoring by exposure and behavior), AI Firewall (inline blocking of malicious traffic while letting legitimate AI interactions continue), and MCP Security Testing (validating MCP servers against chain-of-thought conversations that trigger tool calls, checking for token mismanagement, privilege escalation, and prompt injection). The Anomaly detection function operates through behavioral baseline learning: Levo observes normal traffic patterns at the kernel level and flags deviations from those baselines, including tool call sequences that fall outside established patterns, data flows that cross boundary conditions, and agent behavior that changes after a model update or prompt change.

The fintech Gartner Peer Insights review — describing Levo as "a way to scale API security in a fintech environment without slowing down our developers" — reflects the platform's practical deployment reality: it adds monitoring and security without requiring changes to how engineers build or deploy applications. For infrastructure and platform teams responsible for AI pipeline monitoring across environments they don't fully control (vendor APIs, third-party MCP servers, agent frameworks from multiple sources), that agentless architecture is a meaningful operational advantage over tools that require code instrumentation or proxy deployment to see anything.

Levo's limitation in this guide is the same one flagged in the note above: the platform is primarily a security product with strong monitoring capabilities rather than a pure monitoring platform. Teams whose primary requirement is Infrastructure and Pipeline monitoring specifically, without the security enforcement components, should evaluate whether they want the full Levo stack or a monitoring-first platform like Coralogix alongside a separate security tool.

✓ What We Like

• eBPF kernel-level visibility: Sees all traffic regardless of protocol, framework, or language — no agent deployment, no proxy, no code changes required.

• MCP server monitoring: Discovers and inventories MCP servers across distributed environments, a specific and increasingly important signal category that most monitoring tools don't reach.

• Agentless discovery: The fully agentless API inventory builder added in February 2026 extends the same no-deployment-overhead approach to organizations that can't install agents in regulated environments.

• Behavioral baseline learning: Anomaly detection through deviation from observed normal patterns rather than static rule sets, which adapts as agent behavior evolves.

• FINSEC2025 Innovation Award: Financial services sector recognition for a platform with explicit fintech deployment evidence.

⚠ What to Know

• Levo is primarily an AI and API security platform; teams evaluating it specifically for monitoring should scope the security capabilities separately from the monitoring ones they actually need.

• $4M in total funding and ~39 employees is a lean company profile for enterprise infrastructure monitoring; vendor stability is a legitimate consideration for multi-year deployments.

• eBPF operates at the Linux kernel level, which means the platform requires Linux environments; Windows coverage added PCAP-based sensors, but the core eBPF advantage is Linux-native.

• MCP Discovery and AI Firewall launched during 2026 Launch Week — these are newer capabilities with shorter production deployment histories than Levo's core eBPF API monitoring functions.

Signal Categories Covered

Infrastructure / Pipeline
Anomaly Detection
MCP Server Monitoring
Agent Traffic Visibility

Regulatory Frameworks

PCI DSS
HIPAA
SOC 2
GDPR

Best For

Security and infrastructure teams in regulated industries: Organizations in fintech, healthcare, and government environments where deploying monitoring agents or proxies requires approval processes that slow down visibility.

Teams managing distributed AI infrastructure: Organizations with AI agents, MCP servers, and APIs spread across cloud, on-premises, and edge environments that need unified visibility without modifying each component.

Organizations monitoring third-party AI components: Teams responsible for AI supply chain visibility where third-party agents, tools, and MCP servers generate traffic they can't instrument at the application layer.

Pricing: Not publicly listed. Contact Levo.ai directly or request a match through GetAIGovernance.net.

Opal Security — Best for Agent Identity Monitoring and Just-in-Time Access Governance

The Strongest Production Evidence for Non-Human Identity Monitoring

Choose Opal Security if: you have AI agents operating in production under credentials that were provisioned once and never reviewed, and you need a monitoring and governance layer that tracks what those agents are actually accessing, enforces just-in-time scoping, and surfaces over-privileged access before it becomes an incident.

Founded: Not publicly disclosed

HQ: San Francisco, CA and New York, NY

CEO: Howard Ting (joined December 2025)

Funding: $59M total — $23M round closed June 4, 2026, led by Greylock and Battery Ventures, with Box Group, SVCI, and Cambium Capital

Recognition: Notable Capital Rising in Cyber 2026 — named among the 30 most promising private cybersecurity startups as selected by 150 leading CISOs. Named customers include Databricks (86,000 just-in-time access requests through the platform; Sharon Aby, Senior Manager of IAM at Databricks, on record: "Once our access data was centralized in Opal, generating user access reviews became effortless"), Notion, Cloudflare, Scale AI, CoreWeave, SpaceXAI, and Superhuman. Chronosphere cut standing access by 88 percent using the platform.

Opal Security's AI-native access governance platform, anchored by Paladin — Opal's AI engine for access governance — and the Agents in Opal feature set launched at Identiverse in June 2026, addresses a monitoring problem that most security teams discovered only after agents were already running in production: the credentials agents use to access enterprise systems are almost never reviewed after initial provisioning. Most organizations deploy AI agents with broad access scoped during development, and that access persists indefinitely — there is no operational trigger that prompts a review, so nobody does one. Agents frequently inherit the standing, over-scoped credentials of the people who created them, which compounds the exposure every time a new agent gets deployed.

Opal brings AI agents into the same access graph, reviews, ownership records, and policy-as-code as every other identity in the environment — human employees, service accounts, and API tokens included. Paladin is the engine underneath that governance layer: it reasons over access requests, approves routine grants automatically based on policy, and escalates only the decisions that need a human. Teams set how much authority Paladin handles on a dial rather than a binary switch, and a human signs off on every certification, which is the posture security leaders and auditors can actually defend. Agents in Opal extends that into four specific workflow areas: AI-guided access reviews (where Paladin reduces the roughly 100-hour manual review cycle that most teams still run), access request fulfillment, OpalQuery for plain-English access intelligence questions, and AI-assisted OpalScript for automated remediation actions.

The Databricks figure is the most specific, checkable production evidence in this entire guide for any platform: 86,000 time-bound access requests processed through Opal at a single enterprise deployment. Sharon Aby, Senior Manager of IAM at Databricks, is on record describing the experience directly. Chronosphere cut standing access by 88 percent. At that scale and with those specific outcomes attributed by name, Opal is operating as genuine infrastructure rather than a pilot or proof of concept. The $23M round closed June 4, 2026, 26 days before this article's publication, with Greylock and Battery Ventures leading — making it the most recently funded platform in this guide. More than 60 percent of Opal's team has joined since the start of 2026, which reflects a company growing fast enough to match the market moving under it.

Opal sits at the boundary between AI monitoring and AI security — identity governance for agents is both a monitoring function (what are agents doing with their access in practice?) and an enforcement function (what should they be allowed to do?). This guide evaluates Opal at the Agent Identity and Authorization monitoring layer specifically. Organizations evaluating Opal for identity enforcement more broadly should also review the AI Security Platforms 2026 guide.

✓ What We Like

• Named, on-record Databricks IAM lead: Sharon Aby's quote is specific, attributed, and describes an operational outcome — a different category of evidence than vendor-published case study language.

• 86,000 JIT access requests at Databricks: A checkable production number that describes infrastructure-scale deployment, not a pilot.

• Chronosphere 88% standing access reduction: A second named customer with a specific, measurable outcome, not a general endorsement.

• Paladin access dial, not a binary switch: The architecture keeps a human accountable for the decisions that matter while agents handle routine access governance at machine speed.

• Notable Capital Rising in Cyber 2026: Selected by 150 working CISOs — peer recognition that reflects operational credibility.

• $23M round on June 4, 2026: Greylock and Battery Ventures leading the most recent funding event in this guide, 26 days before publication.

⚠ What to Know

• Agents in Opal launched at Identiverse in June 2026 — agent-specific deployment evidence is accumulating, while the platform's core access governance capabilities have a longer production track record.

• Custom enterprise pricing only; no self-serve evaluation tier or published starting price.

• Maximum value comes from deploying Opal across the full identity estate — agents alongside human users and service accounts — rather than as a standalone agent-only monitoring layer.

• Company founding date and total employee count are not publicly disclosed, which limits third-party assessment of organizational scale beyond what the funding and customer evidence reveals.

Signal Categories Covered

Agent Identity and Authorization
Non-Human Identity Monitoring
Access Pattern Anomaly
Credential Activity Tracking

Regulatory Frameworks

SOC 2
NIST AI RMF
EU AI Act

Best For

• High-growth AI-native companies: Organizations like Databricks, Notion, and Cloudflare where AI agent adoption is accelerating faster than security teams can track credentials manually and where the access review backlog is already a real operational problem.

• Security and IAM teams: Identity and access management functions that need to bring agents into the same governance infrastructure as human and service account identities without standing up a separate agent-specific system.

• Organizations with standing access debt: Teams where agents, service accounts, or users have accumulated over-scoped permissions that nobody has reviewed and where automated remediation through OpalScript is more realistic than manual cleanup.

Pricing: Not publicly listed. Contact Opal Security directly at opal.dev or request a match through GetAIGovernance.net.

SUPERWISE — Best for Regulated Industry AI Monitoring with Built-In Governance Controls

Production AI Monitoring for Banking, Healthcare, and Insurance

Choose SUPERWISE if: you're running AI in a regulated industry — banking, healthcare, or insurance — and you need production monitoring that combines model observability with built-in guardrails, audit trails, and governance controls in a single platform, without assembling those functions from separate tools.

Founded: 2019 (acquired by Blattner Technologies, January 2023; continues operating as SUPERWISE® with independent product development)

HQ: Tel Aviv, Israel

Funding: $4.5M raised pre-acquisition. Post-acquisition financials not publicly disclosed.

Recognition: Named a Sample Vendor for AI Observability in Banking in the Gartner Hype Cycle for Generative AI in Banking 2025. Featured in seven 2025 AI Governance Hype Cycles. Named customers include King, Klarna, Riskified, monday.com, Bridgestone, Firestone AG, and Home Depot. Monitors over 10 billion inferences across the customer base.

Acquisition disclosure: SUPERWISE was acquired by Blattner Technologies in January 2023. The platform continues operating under the SUPERWISE® brand with active product development, including the SUPERWISE AMP (Agentic Management Platform) launch, Gartner Hype Cycle recognition, and new enterprise customer acquisitions post-2023. Pricing tiers are publicly available (Starter free, Pro $299/month, Enterprise custom), which is a sign of active commercial operation rather than a sunset product.

SUPERWISE's Agentic Management Platform (AMP), described by the company as a control plane for AI operations in regulated industries, combines observability, risk management, and operational oversight for AI models and agents with guardrails, explainability, and compliance features in a single system. The platform monitors production AI behavior in real time, tracks model drift and data quality degradation, fires automated alerts before quality issues reach business impact, and enforces governance guardrails that stop harmful outputs before they reach end users. Runtime protection operates in under 10 milliseconds, meaning guardrail enforcement doesn't add meaningful latency to production AI interactions.

SUPERWISE's place in the User Behavior and Feedback signal categories reflects a specific production pattern that banking and insurance customers face: AI models making ongoing decisions on underwriting, fraud detection, lead scoring, and customer lifetime value predictions, where the feedback signal that a model has gone wrong is often a lagging business metric rather than an obvious technical failure. Paz Aviv, Product Lead at one of SUPERWISE's customers, described monitoring over 6,000 metrics in real time across their AI estate — giving the team "control over our dynamic business and peace of mind that we'll always know about unwanted issues." That scale of metric tracking, combined with the AMP's automated incident grouping that reduces alert noise, addresses the User Behavior monitoring problem from a model-operation angle: watching how models behave when users interact with them in production, surfacing degradations in accuracy or fairness that would only appear in business KPIs weeks later without monitoring.

The Gartner recognition specifically for AI Observability in Banking is the most relevant third-party validation for SUPERWISE's regulated-industry positioning. Gartner's definition in the 2025 report — that AI observability in banking "extends beyond traditional IT monitoring by specifically tracking real-time AI behaviors, inputs, outputs to improve decision accuracy and enhance trust in banking" — describes exactly the problem SUPERWISE's customer base in financial services faces and the platform's monitoring architecture addresses. The Klarna, Riskified, and monday.com customer references add commercial breadth across fintech and consumer technology alongside the banking and insurance vertical depth.

✓ What We Like

• Gartner Hype Cycle for Generative AI in Banking 2025: Named vendor recognition specifically for the regulated financial services AI observability category where SUPERWISE's depth is most differentiated.

• 7 AI Governance Hype Cycle appearances in 2025: Repeated independent analyst recognition across multiple governance and observability frameworks.

• AMP runtime protection under 10ms: Guardrail enforcement fast enough for production AI interaction volumes without latency impact.

• 10 billion+ inferences monitored: Production scale evidence that the platform handles real enterprise AI deployment volumes.

• Free Starter tier: Teams evaluating regulated-industry AI monitoring can run a proof of concept without committing budget first.

• Combined observability and guardrails: Monitoring and enforcement in one system reduces the integration overhead of connecting separate monitoring, alerting, and intervention tools.

⚠ What to Know

• Acquired by Blattner Technologies in January 2023; post-acquisition financials are not publicly disclosed, and the parent company's primary domain is agricultural technology rather than enterprise software.

• G2 profile activity has been limited in recent periods, suggesting the team has deprioritized public review platforms; direct customer reference calls are a more reliable evaluation signal than public review volume.

• A January 2026 technical evaluation scored the platform low on dataset management and data versioning; organizations with complex data lineage requirements should verify those capabilities directly.

• SUPERWISE's strongest positioning is in regulated industries — organizations outside banking, healthcare, and insurance may find more suitable primary options in this guide.

Signal Categories Covered

User Behavior
Feedback
Performance + Drift
Bias Detection
Governance Evidence

Regulatory Frameworks

SR 26-2
EU AI Act
SOC 2
HIPAA
GDPR

Best For

Banks and financial institutions: Organizations running AI on underwriting, fraud detection, and credit decisions where model drift has direct regulatory consequences and Gartner banking-specific recognition matters for procurement justification.

Healthcare and insurance AI teams: Regulated environments where monitoring, guardrails, and compliance evidence need to come from one platform rather than being assembled from separate tools.

Organizations needing a free evaluation tier: Teams in regulated industries that want to run a proof of concept on real production data before committing budget to an enterprise monitoring platform.

Pricing: Starter: free (5 agents, 10,000 API calls). Pro: $299/month. Enterprise: custom. Contact SUPERWISE directly or request a match through GetAIGovernance.net.

W&B Weave — Best for Experiment-Linked LLM Cost Attribution and Change Governance

LLM Observability Tied to the Experiment Tracking Standard Used by a Million Developers

Choose W&B Weave if: your team already uses Weights & Biases for ML experiment tracking, and you want LLM tracing and cost attribution that connects directly to the experiment and model version history you already maintain there — without adopting a separate platform for AI monitoring.

Founded: 2018 (Weights & Biases, acquired by CoreWeave May 2025 for ~$1.4-1.7B; W&B Weave continues as an active CoreWeave product)

HQ: San Francisco, CA

Funding: $250M raised pre-acquisition. CoreWeave is publicly listed (NASDAQ: CRWE).

Recognition: 1 million+ active users. 40% of Fortune 500 use W&B for ML experiment tracking. Customers include OpenAI, Meta, NVIDIA, Toyota, Samsung, and Salesforce. CoreWeave launched ARIA, an agentic research assistant built on W&B Weave, on June 29, 2026.

Acquisition disclosure: CoreWeave acquired Weights & Biases on May 5, 2025 for approximately $1.4 to 1.7 billion. W&B Weave continues as an actively developed CoreWeave product; ARIA, the agentic research assistant built on Weave, launched on June 29, 2026 (one day before this article's publication). Pricing is publicly available and unchanged. CoreWeave is a publicly listed GPU cloud company; W&B operates within CoreWeave's AI infrastructure portfolio rather than being dissolved into a specific product line.

W&B Weave is the LLM observability and evaluation layer within the Weights & Biases platform, which has been the standard tool for ML experiment tracking since 2018. Weave tracks every prompt, response, cost, token count, and latency across production LLM interactions, attributing those signals back to the experiment runs, model versions, and prompt configurations that produced them. This is the distinctive function Weave provides in the Cost and Resource signal category: version-to-cost causality. When a model update or prompt change causes unexpected cost increases in production, Weave connects the production cost signal back to the specific experiment that introduced the change — which is what makes the monitoring actionable rather than merely informational.

The governance evidence angle is where Weave's experiment tracking history becomes a monitoring asset. Every model version, every prompt change, every evaluation result, and every production trace is stored with the same versioning and attribution logic that W&B's experiment tracking has applied to training runs since 2018. That means when an audit or regulatory review asks which model version was running during a specific time period, what evaluation results supported its deployment, and what prompted the change from a previous version, W&B's records answer those questions with the precision of a version control system rather than the approximation of a documentation process. Merck's CDAO described this directly: W&B "enables us to track every prompt, response, cost, and latency in real time, turning black-box models into auditable, optimizable assets." ARIA, which CoreWeave launched on Weave on June 29, 2026, uses that same data foundation to automate research analysis — building dashboards and visualizations from production experiment data rather than requiring researchers to manually triage results.

W&B Weave's position as a second option at Cost and Resource (primary: Langfuse) reflects a genuine difference in buyer profile rather than a quality ranking. Langfuse is the right choice for teams that want open-source, self-hosted LLM cost tracking without a dependency on W&B's experiment tracking ecosystem. W&B Weave is the right choice for the substantial share of ML engineering teams — 40% of the Fortune 500, 1 million active developers — already using W&B for experiment tracking, for whom Weave is a natural extension of existing infrastructure rather than a new vendor relationship.

✓ What We Like

• Version-to-cost causality: Connects production cost and quality signals back to specific experiment runs, model versions, and prompt changes rather than presenting aggregate metrics without attribution.

• 1 million+ active W&B users: An existing developer base that adopts Weave as an extension of infrastructure they already use, rather than as a new tool requiring a new learning curve.

• CoreWeave's ARIA agent: The June 29, 2026 launch of an agentic research assistant built on Weave demonstrates active investment in the platform one day before this article's publication.

• Fortune 500 customer breadth: OpenAI, Meta, NVIDIA, Toyota, Samsung, and Salesforce as named customers across AI research, manufacturing, and enterprise software.

• Governance-ready audit trail: Experiment tracking history that answers regulatory questions about which model version was running when and what evaluation supported its deployment.

⚠ What to Know

• CoreWeave acquired W&B in May 2025; the long-term product roadmap is now influenced by CoreWeave's GPU cloud strategy, which may evolve differently from what an independent W&B would have prioritized.

• Maximum value comes from using Weave alongside W&B's experiment tracking infrastructure; teams not already on W&B get less incremental benefit from Weave than existing W&B users do.

• Weave's cost per trace ingestion ($0.10/MB) can produce unexpected costs for teams with high-volume LLM production traffic; budget modeling before deployment is important.

• Weave focuses on LLM and agent observability; traditional ML model drift monitoring remains more mature in W&B's Models product than in Weave specifically.

Signal Categories Covered

Cost and Resource
Governance Evidence
LLM Tracing
Experiment Attribution

Regulatory Frameworks

SOC 2
HIPAA
GDPR
ISO/IEC 27001
NIST 800-53

Best For

• Existing W&B users: ML engineering teams already using Weights & Biases for experiment tracking who want LLM cost attribution and tracing in the same platform rather than a separate tool.

• Research-heavy AI teams: Organizations where experiment management and model evaluation are central workflows, and where connecting production LLM behavior to specific research decisions matters for both performance and governance.

• Teams needing change attribution governance: Organizations subject to audit requirements that ask what model version was deployed when and what evaluation evidence justified that deployment.

Pricing: Free personal tier. Pro: $60/month. Enterprise: custom. Weave ingestion billed at $0.10/MB over free tier. Contact W&B directly or request a match through GetAIGovernance.net.

Zenity — Best for Agent Intent Monitoring and Execution-Path Policy Compliance

The Platform That Examines What an Agent Was Trying to Do, Not Just What It Said

Choose Zenity if: you have AI agents running across your enterprise environment — in SaaS platforms, cloud-built applications, and on endpoint devices — and you need to know whether those agents are operating within their authorized scope, with monitoring that tracks execution paths rather than just prompt content.

Founded: 2021

HQ: New York, NY and Tel Aviv, Israel

Employees: 65+

Funding: $55M+ total ($16.5M Series A + $38M Series B)

Recognition: Named in two categories in the Gartner Hype Cycle for Agentic AI (April 2, 2026). Named "Company to Beat" in Gartner's 2026 AI Vendor Race report. Carahsoft federal distribution partnership (June 17, 2026). Named customers include WM (Waste Management), with specific metrics: 575,000 resources, 1,500 connectors, 200+ environments governed through the platform. Fortune 20, Fortune 50, and Fortune 200 customer references on the company website.

Zenity's platform is organized into three pillars — Observe, Govern, and Defend — that collectively cover the Intent and Policy Compliance signal category by examining agents at the execution-path level rather than the prompt level. The Observe function continuously discovers and inventories AI agents across SaaS platforms, cloud frameworks, and endpoint devices, including shadow agents built by employees without security team knowledge. The Govern function applies AI Security Posture Management: evaluating agent configurations, permissions, tool integrations, memory access, and instruction sets against organizational policy before agents run, surfacing misconfigurations and over-permissioned access that create compliance exposure. The Defend function, which includes the Correlation Agent launched in late 2025, provides runtime detection and response: it connects every signal, data point, and interaction from across the agent lifecycle into a single coherent narrative about what an agent was attempting to accomplish.

The Intent and Policy Compliance signal category in GAIG's framework is specifically about whether agents stay on task, whether their tool calls match their authorized scope, and whether their outcomes align with organizational policy — not just whether individual prompts look safe. The Correlation Agent addresses exactly this: it interprets what an agent was trying to do by connecting tool calls, memory access patterns, data usage, and control flow into a story rather than treating each interaction as an isolated event. This catches sophisticated attacks that prompt-level monitoring misses: a gradual sequence of individually innocent-looking tool calls that together constitute unauthorized data exfiltration, or an agent that modifies its own instruction context across multiple sessions rather than through a single obviously malicious prompt.

The WM (Waste Management) deployment is the most specific named case study in Zenity's public documentation: 575,000 resources, 1,500 connectors, and more than 200 environments governed through the platform, with Monica Taylor Boggan, head of WM's business information security office, on record stating that Zenity gave her team "the visibility and control to operate governed autonomous agents at enterprise scale." The Gartner Hype Cycle for Agentic AI recognition in two categories from April 2026 — alongside the "Company to Beat" designation in Gartner's AI Vendor Race — provides the analyst validation that enterprise procurement teams need to justify the selection. The Carahsoft federal distribution partnership from June 17, 2026 (13 days before this article's publication) means Zenity is accessible through SEWP V and other government procurement vehicles, which opens federal market adoption alongside the commercial enterprise base.

✓ What We Like

• Execution-path analysis: Zenity's Correlation Agent monitors what agents were attempting across full execution paths rather than evaluating individual prompts in isolation — the distinction matters for catching sophisticated multi-step intent violations.

• Gartner Hype Cycle for Agentic AI (two categories): Independent analyst recognition from April 2026 in both "Agentic AI Security" and a second agentic AI category.

• Gartner 2026 AI Vendor Race "Company to Beat": A specific competitive positioning designation from an independent analyst that reflects Gartner's view of Zenity's market trajectory.

• WM deployment with specific metrics: 575,000 resources and 1,500 connectors governed in a named Fortune 200 customer is a verifiable and scale-credible reference.

• Carahsoft federal distribution: Government procurement accessibility through established federal acquisition channels, announced June 17, 2026.

• Coverage across SaaS, cloud, and endpoint: Agent monitoring that reaches across environments rather than only cloud-native agents, which matters for enterprises with legacy and hybrid infrastructure.

⚠ What to Know

• Zenity is primarily a security and governance platform for AI agents; its placement in this monitoring guide reflects the Intent and Policy Compliance signal category specifically. Organizations evaluating Zenity for broader AI security should review the AI Security Platforms 2026 guide alongside this one.

• No offensive security or red-teaming capabilities — Zenity monitors and prevents, but doesn't proactively test what an attacker could exploit in agent configurations.

• Developer toolchain coverage is surface-level for endpoint AI coding agents like Cursor and Copilot; MCP tool execution monitoring is not the same depth as MCP server governance at the platform level.

• Custom enterprise pricing only; no published self-serve evaluation tier.

Signal Categories Covered

Intent and Policy Compliance
Agent Discovery and Inventory
AI Security Posture Management
Execution-Path Anomaly Detection
Policy Enforcement

Regulatory Frameworks

OWASP LLM Top 10
MITRE ATLAS
NIST AI RMF
EU AI Act

Best For

• Enterprise security teams with agents across mixed environments: Organizations where agents run in SaaS platforms, cloud frameworks, and endpoint devices simultaneously and need a single visibility layer across all three.

• Organizations with sensitive data and autonomous agents: Companies where agents have access to PHI, PII, or financial data, and where monitoring whether those agents stay within authorized scope is a legal or operational necessity.

• Federal and regulated industry organizations: Government agencies and regulated enterprises that need agent governance accessible through federal acquisition vehicles with analyst validation behind the placement.

Pricing: Not publicly listed. Contact Zenity directly or request a match through GetAIGovernance.net.

Sources

The following sources were used in the research and writing of this guide. Claims are attributed to specific sources. Platform capabilities described without external citations are drawn from vendor documentation listed below.

1. Arize AI, "Arize AX Changelog," release notes through June 2026. https://arize.com/docs/ax/release-notes

2. Arize AI, "Arize Phoenix," product page. https://arize.com/phoenix/

3. Arize AI, "Arize AX," Azure Marketplace listing, June 2026. https://marketplace.microsoft.com/en-us/product/saas/arizeai1657829589668.arize_ai

4. AppSecSanta, "Arize AI Review 2026: AI Observability & LLM Evaluation," June 2026. https://appsecsanta.com/arize-ai

5. GitHub, Arize-ai/phoenix repository. https://github.com/Arize-ai/phoenix

6. Arthur AI, "Arthur Launches Agent Discovery & Governance (ADG) Platform on Google Cloud Marketplace," January 7, 2026. https://www.arthur.ai/blog/arthur-launches-agent-discovery-governance-on-google-cloud-marketplace

7. Arthur AI, "Arthur in 2025: Building Trust and Governance for the Agentic AI Era," December 22, 2025. https://www.arthur.ai/blog/2025-recap

8. AppSecSanta, "Arthur AI 2026: Model Monitoring & AI Governance," May 19, 2026. https://appsecsanta.com/arthur-ai

9. Arthur AI, LinkedIn company page, tokenmaxxing and AI governance content, June 2026. https://www.linkedin.com/company/arthurai

10. Braintrust, "Announcing our Series B," February 17, 2026. https://www.braintrust.dev/blog/announcing-series-b

11. SiliconANGLE, "Braintrust lands $80M funding round to become the observability layer for AI," February 17, 2026. https://siliconangle.com/2026/02/17/braintrust-lands-80m-series-b-funding-round-become-observability-layer-ai/

12. Tracxn, "Braintrust — 2026 Company Profile, Team, Funding & Competitors." https://tracxn.com/d/companies/braintrust

13. Coralogix, "Coralogix Raises $200M to Scale the Observability Backbone for the Age of AI," June 3, 2026. https://coralogix.com/blog/coralogix-raises-200m-to-scale-the-observability-backbone-for-theage-of-ai/

14. Advent International, "Coralogix Raises $200M to Scale the Observability Backbone for the Age of AI," June 3, 2026. https://www.adventinternational.com/news/coralogix-raises-200m-to-scale-the-observability-backbone-for-the-age-of-ai/

15. Coralogix, "Coralogix Acquires Aporia," December 23, 2024. https://coralogix.com/blog/coralogix-acquires-aporia/

16. Tracxn, "Coralogix — 2026 Company Profile, Team, Funding & Competitors." https://tracxn.com/d/companies/coralogix

17. Evidently AI, official website. https://www.evidentlyai.com/

18. PitchBook, "Evidently AI 2026 Company Profile." https://pitchbook.com/profiles/company/469919-98

19. Y Combinator, "Evidently AI company profile." https://www.ycombinator.com/companies/evidently-ai

20. Fiddler AI, "Fiddler Raises $30M Series C to Power the Control Plane for AI Agents," Business Wire, January 27, 2026. https://www.businesswire.com/news/home/20260127042634/en/

21. Fiddler AI, "Fiddler Series C: The Control Plane Moment for AI," blog post, April 8, 2026. https://www.fiddler.ai/blog/series-c

22. Fiddler AI, "Fiddler Raises $30M Series C to Deliver the First Control Plane for AI," press release, January 27, 2026. https://www.fiddler.ai/press-releases/fiddler-raises-30m-series-c

23. ClickHouse, "ClickHouse Raises $400M Series D, Acquires Langfuse," January 16, 2026. https://clickhouse.com/blog/clickhouse-raises-400-million-series-d-acquires-langfuse-launches-postgres

24. ClickHouse, "ClickHouse Welcomes Langfuse: The Future of Open-Source LLM Observability," January 2026. https://clickhouse.com/blog/clickhouse-acquires-langfuse-open-source-llm-observability

25. Langfuse, "Langfuse Joins ClickHouse," company announcement. https://langfuse.com/blog/joining-clickhouse

26. Langfuse, "How did we get here?" company handbook. https://langfuse.com/handbook/chapters/story

27. Levo.ai, product updates page, 2026 Launch Week announcements. https://www.levo.ai/resources/product-updates

28. AppSecSanta, "Levo.ai Review 2026: eBPF API Security Platform," May 19, 2026. https://appsecsanta.com/levo-ai

29. PitchBook, "Levo (Software Development Applications) 2026 Company Profile." https://pitchbook.com/profiles/company/481424-68

30. Gartner Peer Insights, "Levo.ai Reviews & Ratings 2026." https://www.gartner.com/reviews/market/application-security-testing/vendor/levo/product/levo-ai

31. Yahoo Finance / ACCESS Newswire, "Oasis Security Raises $120M Series B to Secure the Rise of Enterprise AI Agents," March 19, 2026. https://finance.yahoo.com/sectors/technology/articles/oasis-security-raises-120m-series-160000141.html

32. SiliconANGLE, "Oasis Security Raises $120M to Secure Nonhuman Identities," March 19, 2026. https://siliconangle.com/2026/03/19/oasis-security-raises-120m-secure-non-human-identities-across-ai-cloud-environments/

33. CB Insights, "SUPERWISE company profile and funding." https://www.cbinsights.com/company/superwiseai

34. CheckThat.ai, "SUPERWISE AI Reviews 2026: What Users Really Think," May 20, 2026. https://checkthat.ai/brands/superwise/reviews

35. SUPERWISE, "Named a Sample Vendor for AI Observability in Banking in Gartner Hype Cycle for Generative AI in Banking, 2025," company announcement. https://www.cbinsights.com/company/superwiseai

36. SUPERWISE, official website and AMP product page. https://superwise.ai/

37. SiliconANGLE, "CoreWeave debuts ARIA agent to automate AI research in Weights & Biases," June 29, 2026. https://siliconangle.com/2026/06/29/coreweave-debuts-aria-agent-automate-ai-research-weights-biases/

38. UsagePricing, "Weights & Biases Pricing," June 2026. https://www.usagepricing.com/blueprint/weights-biases

39. Tracxn, "Weights & Biases — 2026 Company Profile, Team, Funding." https://tracxn.com/d/companies/weights-biases

40. Weights & Biases, W&B Weave documentation. https://docs.wandb.ai/weave

41. Zenity, "Zenity Named in Two Categories in the 2026 Gartner Hype Cycle for Agentic AI," Business Wire, April 15, 2026. https://www.businesswire.com/news/home/20260415309905/en/

42. Akto, "Zenity Security 2026: Features and Comparison," March 31, 2026. https://www.akto.io/blog/zenity-security

43. Channel Insider, "Zenity Adds Agentic Browser Protection, LLM Defense Tools," December 5, 2025. https://www.channelinsider.com/security/tools-and-platforms/zenity-ai-security-features/

44. Zenity, official website and Correlation Agent product page. https://zenity.io/

45. GetAIGovernance.net, "AI Monitoring Signals Explained," updated June 18, 2026. https://getaigovernance.net/blog/ai-monitoring-signals-explained

46. GetAIGovernance.net, "AI Compliance Certifications, Frameworks, and Laws Explained." https://getaigovernance.net/blog/ai-compliance-certifications-frameworks-and-laws-explained

47. Cisco, "Cisco Completes Acquisition of Galileo," May 22, 2026 update to original April 9, 2026 announcement. https://blogs.cisco.com/news/cisco-announces-the-intent-to-acquire-galileo

48. Cisco, "Acquisitions by Year," acquisitions list including Galileo Technologies (April 9, 2026). https://www.cisco.com/site/us/en/about/corporate-development/acquisitions/acquisitions-list-years/index.html