Skip to content
AI governance AI governance platform CISO

AI Governance Platform: An Enterprise Buyer's Guide

A policy PDF cannot stop a prompt or an agent action. What to require from an AI governance platform: enforcement, masking, audit evidence, spend, agents.

Mathieu Perochon
Mathieu Perochon Founder, GuardWeaver
Updated November 19, 2026 … min read
Central control layer linking users, agents, models, and an audit ledger over MCP and A2A

Most enterprises wrote an AI usage policy before they had any way to enforce it. Employees paste contracts into chat tools, agents call internal APIs, and the policy PDF sees none of it. IBM’s analysis of its 2026 Cost of a Data Breach Report puts a number on the drift: the share of studied organizations reporting security incidents involving unapproved AI use rose from 20% to 43% (IBM, September 2026).

This guide is for CIOs and CISOs evaluating software to govern how employees and AI agents use AI. It covers the capabilities to require, a comparison of the four common approaches, and a requirements checklist you can take into vendor demos.

What is an AI governance platform, and what is it not?

It is software that sees how people and agents use AI, applies your policy at the moment of use, and keeps evidence of every decision taken.

The term covers two different jobs, and the difference decides which vendors belong on your shortlist.

Model risk management serves data science teams that build or fine-tune models. It inventories models, documents training data, runs bias and robustness tests, and gates releases. The NIST AI Risk Management Framework, released on January 26, 2023, describes much of this work under its MAP and MEASURE functions.

Usage governance serves the CIO and CISO. It controls what employees and agents do with AI every day: which data goes into prompts, which documents an assistant may retrieve, which actions an agent may take in a business tool, and how much it all costs.

Gartner’s market sizing also treats these as distinct needs. In August 2026 it divided the market for securing AI into four segments: AI application security, AI usage control, AI governance platforms, and AI gateways. This guide focuses on usage governance. If your main risk is a credit-scoring model built in-house, you need model risk tooling first.

Why is a written AI policy not enough?

A policy is read once, then applied thousands of times by people and agents nobody watches. Without enforcement and a log, compliance is unprovable.

The data on policies without enforcement is consistent. In IBM’s 2025 report, based on Ponemon Institute research into 600 breached organizations, 63% had no AI governance policy or were still developing one. Among those with a policy, only 34% performed regular audits for unsanctioned AI (IBM Cost of a Data Breach Report 2025). In the 2026 edition, 92% of organizations that reported an AI-related breach lacked proper AI access controls (IBM X-Force, July 2026).

Analysts describe the same shift. In its strategic predictions for 2027 and beyond (September 2026), Gartner writes that organizations are moving “from policy-based AI governance to operational governance that embeds controls directly into AI systems and workflows.” In a February 2026 Q&A, it reported that organizations that deployed AI governance platforms were 3.4 times more likely to achieve high effectiveness in AI governance, based on a survey of 360 organizations in the second quarter of 2025. That is a correlation, not proof of cause, but it points the same way.

The frameworks your auditors use expect the same. NIST AI RMF GOVERN 1 asks that policies be “in place, transparent, and implemented effectively.” ISO/IEC 42001, the AI management system standard published in December 2023, follows a Plan-Do-Check-Act cycle: the policy is the Plan, and the evidence for Check has to come from somewhere.

On the regulatory side, the EU AI Act’s AI literacy duty (Article 4) has applied since February 2, 2025, and the Article 50 transparency duties since August 2, 2026. Obligations for Annex III high-risk systems now start on December 2, 2027, after the Digital Omnibus on AI postponed them. The logging duties of Articles 12 and 26(6) apply to high-risk systems only, so do not assume every internal assistant falls under them.

Which capabilities should an AI governance platform have?

Seven: usage visibility, enforcement before the action, data masking, an audit log, spend caps, external agent coverage, deployment choice.

Visibility on usage

You cannot write good rules for usage you have not seen. Ask for a dashboard of usage per team, agent, and model, with the questions asked and the spending behind them. A useful test: can the platform run rules in an observation mode that flags without blocking, so you calibrate before you enforce?

Policy enforcement before the action

The decisive question is where the rule applies. A rule checked after the answer has reached the user, or after the agent has written to the CRM, is a report. Ask the vendor to list every enforcement point: when the request arrives, when documents are retrieved, when an agent calls a tool, when it delegates to another agent, when the answer goes out. Then ask what a rule can decide besides block or allow.

Sensitive data masking

Masking replaces personal or secret data before it reaches the model provider. Ask which data classes the detectors recognize, and where masking applies: the prompt only, or also the conversation history, the retrieved context, tool arguments, and agent handoffs. A platform that masks the prompt but forwards the full history leaks the same data one turn later.

Audit log as evidence

An audit log supports an audit only if nobody can rewrite it. Ask whether it is append-only, whether entries are chained so tampering is detectable, whether it can be exported with an integrity proof, and how long it is retained on your plan. Ask what it records: decisions, refusals, masking events, and changes to the rules themselves. The log is evidence among others. It does not make you compliant by itself.

Spend controls

Agentic workflows can multiply model calls without anyone noticing. In the same September 2026 predictions, Gartner expects 60% of Global 500 companies to embed AI FinOps control at inference by 2028. Ask for caps per user, per agent, and per model, applied in real time, plus support for your own provider keys so you see the provider’s real bill.

Coverage of external agents over MCP and A2A

The Model Context Protocol connects an agent to tools and data, while A2A connects agents to each other (A2A and MCP compared). An agent built by another team or a partner can reach your systems through either. Ask whether those external agents cross the same enforcement points and land in the same log as your employees. If they bypass the platform, your governance stops at your own chat window.

Deployment model and data location

A governance layer sees every prompt and every answer, which makes it one of the most sensitive systems you run. Ask where it is hosted, whether it can run on your own servers, and whether it works fully offline with a local model if your sector requires it.

How do the main approaches to AI governance compare?

Policies set intent, network controls block destinations, gateways route model calls, governance platforms add agent rules and evidence.

The four approaches are not exclusive. Most enterprises combine at least two. The table describes what each one typically covers. Individual products vary, so test each row in a demo.

Requirement Policy document only Network or CASB controls AI gateway AI governance platform
Where control applies In the employee’s head At the network edge, by destination On model API calls routed through it At each step of an AI exchange, including agent actions
Sees prompt and answer content No Partly, depending on inspection Yes, for routed calls Yes
Masks sensitive data before the model No Rarely Often Expected
Controls agent tool calls and handoffs No No Rarely Expected, for traffic it handles
Evidence for auditors Signed acknowledgments Traffic logs Request logs Decision log tied to rules
Spend control per user, agent, model No No Often per key or route Expected
Covers unsanctioned consumer tools On paper only Yes, by blocking or allowing domains No Only through a detection feature, if offered

Two lessons come out of the table. Network controls remain the only layer that reaches a consumer AI tool opened in a browser, so a governance platform does not retire them. And a gateway that only sees model API calls cannot judge an agent’s action in your ERP, which is where agentic risk sits. Gartner’s April 2026 guidance on agent sprawl lists monitoring agent behavior and enforcing policy among six steps, and notes that only 13% of organizations think they have the right AI agent governance in place.

What should your AI governance requirements checklist include?

Twelve testable questions, each answered in a live demo on your own use case rather than in a slide deck or a written questionnaire.

  1. Can I see usage, questions, and spending per team, agent, and model?
  2. Can rules run in observation mode before blocking?
  3. At which points of an exchange can a rule apply, including retrieval, tool calls, and agent delegation?
  4. What can a rule decide besides allow and block?
  5. Which sensitive data classes are detected, and where is masking applied?
  6. Is the audit log append-only, chained, and exportable with an integrity proof?
  7. Does the log record changes to the rules, not only user requests?
  8. How long is the log retained, and can that match my retention policy?
  9. Can I cap spending per user, agent, and model, with my own provider keys?
  10. Do external agents connecting over MCP or A2A cross the same rules?
  11. Where is the platform hosted, and can it run on my servers or offline?
  12. What does the platform not cover, in the vendor’s own words?

The last question is the most revealing. A vendor that cannot name its limits has not mapped where its enforcement points stop.

Where does GuardWeaver fit in an AI governance stack?

GuardWeaver is a usage governance platform: its rules apply in real time to employees and to external agents that go through it.

According to its governance page, a GuardWeaver rule can take four decisions: allow, deny, withhold, or flag. Rules can act at eight points, including when a question arrives, when documents are retrieved, when an agent calls a tool, when it hands off to another agent, and when the answer goes out. With Premium and Enterprise, automatic masking recognizes ten data classes, such as email addresses, IBANs, payment card numbers, and API keys, and replaces them before they reach the model. That masking covers the query, the history, the context sent to the model, tool arguments, delegation, and the response.

The audit log is append-only. With Premium and Enterprise, each entry is chained to the previous one and the log can be exported; with Enterprise, it is also sealed in segments and timestamped by a third-party authority. Spending caps per user, agent, and model come with Premium and Enterprise, and every plan uses your own AI provider keys with no markup. With Premium and Enterprise, external agents that route through your MCP or A2A servers cross the same enforcement points as your teams, as the orchestration page explains. GuardWeaver runs as SaaS hosted in Europe or, with Enterprise, on your own servers, up to an air-gapped install with a local model.

What it does not do matters as much. GuardWeaver governs the AI traffic that passes through it. It is not a model risk management tool for data science teams. Shadow AI detection is listed as “coming soon” on the pricing page and is not available today, so consumer tools opened outside the platform still need network controls and a clear policy. For the data protection side of a deployment, our GDPR checklist for knowledge base chatbots covers the questions to ask, and our guide to AI agent orchestration platforms goes deeper into MCP and A2A.

To see how rules, masking, and the audit log work together, start with the governance page. The pricing page shows which governance features each plan includes: you can create a free account or try Premium with no credit card.

Frequently asked questions

What is an AI governance platform?

Software that shows how employees and AI agents use AI, applies the organization's usage policy at the moment of use, and keeps a record of each decision. Gartner lists AI governance platforms as one of four segments of the market for securing AI, next to AI application security, AI usage control, and AI gateways.

How is AI usage governance different from model risk management?

Model risk management validates the models a data science team builds: training data, testing, bias, and approval before release. Usage governance controls what people and agents do with AI every day: which data they send, which actions agents take, and what it costs.

Does the EU AI Act require an AI governance platform?

No article requires a specific tool. The AI Act requires AI literacy measures (Article 4) and transparency duties (Article 50), and its logging duties under Articles 12 and 26(6) apply to high-risk systems. A platform helps produce evidence for these duties. It does not replace them.

What should an AI audit log record?

Each evaluated request, the decision taken and the rule that took it, every refusal, masking, or withheld answer, and every change to the rules themselves. It should be append-only and exportable so an auditor can verify it.

Can an AI governance platform control external AI agents?

Only the agents whose traffic goes through it. When external agents call your tools over MCP or delegate tasks over A2A through the platform, the same rules and the same log apply to them as to your employees.

All guides on this topic →

Try GuardWeaver on your own documents

15 days of Premium included, no credit card. A demo is still available if you prefer.

Try it free

Book a demo