The 90-day agent governance implementation playbook

The CISO of a mid-market insurance company told me his governance story in two sentences: “We had a board-approved AI strategy. We did not have a single document that said which agents were running, who owned them or what data they could access.”

He is not unusual. Grant Thornton’s 2026 AI Impact Survey found that 78% of executives lack strong confidence they could pass an independent AI governance audit within 90 days. Meanwhile, 75% of boards have approved major AI investments. Only 48% have set governance expectations for those investments.

The gap between investment and oversight is where risk accumulates.

This playbook closes that gap in 90 days. Not by building a perfect governance program on day one, but by building a functional one. One that gives you visibility by day 30, control by day 60 and evidence by day 90.

Before you start: three prerequisites#

1. Executive sponsorship. You need a CTO or CISO who will fund the program, attend the kickoff and enforce decisions when teams resist. Governance without authority is a suggestion box.

2. A cross-functional founding team. Security, compliance, legal, engineering and at least one business unit representative. Organizations with cross-functional AI governance teams deploy AI 40% faster and face 60% fewer post-deployment compliance issues than those using siloed approaches.

3. A “won’t do” list. Write down what this program will not attempt in 90 days. It will not build custom tooling from scratch. It will not govern every AI model in the organization (only agents). It will not rewrite the enterprise data classification policy. Scope discipline is survival discipline.

Phase 1: Discovery and inventory (days 1-30)#

The goal of this phase is simple: find every agent, assign every owner and kill everything that should not be running.

Week 1-2: Shadow agent audit#

You do not know how many agents you have. Nobody does. The shadow AI problem is universal: 68% of employees use AI tools without IT approval, and 98% of organizations report unsanctioned AI use.

To find what is running:

• Scan OAuth grants across your identity provider for connections to AI platforms (OpenAI, Anthropic, Google AI, Cohere, Azure AI)
• Analyze API gateway logs for traffic to LLM endpoints
• Review cloud billing for AI service subscriptions that were not provisioned through IT
• Check expense reports for SaaS AI tool purchases on corporate cards
• Interview engineering teams. Ask: “What AI tools do you use daily that IT did not set up?” Frame it as discovery, not enforcement. You will get more honest answers.
• Scan for MCP server connections and IDE extensions with agent capabilities

Most enterprises discover 2-3x more agents than they expected. One financial services firm I worked with found 47 active agents when they expected 15.

Week 2-3: Stakeholder mapping#

For every discovered agent, answer four questions:

• Who built it? The creator might have left the company.
• Who owns it now? If nobody, assign a temporary owner immediately.
• What data does it access? PII, financial data, health records or internal-only documents.
• What happens if it stops? Some agents are critical path, while others are experiments nobody remembers starting.

Build this into a simple registry. Even a structured spreadsheet with mandatory fields is better than nothing. The fields you need at minimum:

• Agent name and ID
• Owner (named individual, not a team alias)
• Purpose (one sentence)
• Data classification (public, internal, confidential, restricted)
• Risk tier (low, medium, high, critical)
• Status (active, inactive, under review, decommissioned)
• Dependencies (upstream and downstream systems)

Week 3-4: Risk assessment baseline and quick wins#

Classify every agent into risk tiers. Use a simple matrix:

	Low data sensitivity	High data sensitivity
Low autonomy	Low risk	Medium risk
High autonomy	Medium risk	High/Critical risk

Then execute the quick wins:

• Revoke credentials for agents that are inactive or whose owners have left
• Disable agents that access data above their justified need
• Rotate shared credentials used by agents and humans
• Publish a one-page acceptable use policy that defines minimum requirements for running an agent in the enterprise

These quick wins create immediate risk reduction and, more importantly, organizational momentum. Governance becomes visible. People start paying attention.

Phase 1 deliverables checklist#

• Complete agent inventory with ownership assignments
• Shadow agent audit report
• Stakeholder map (agent owners, data stewards, executive sponsors)
• Risk classification for every discovered agent
• Quick-win execution report (credentials revoked, agents disabled)
• One-page acceptable use policy (published and acknowledged by engineering)
• Phase 2 scope and resource requirements

Phase 1 RACI#

Activity	CISO	Governance Lead	Engineering	Business Units
Shadow agent audit	A	R	C	C
Stakeholder mapping	I	R	C	R
Risk classification	A	R	C	I
Quick-win execution	A	R	R	I
Acceptable use policy	A	R	C	I

R = Responsible, A = Accountable, C = Consulted, I = Informed

Phase 1 failure modes#

• Discovery theater. Scanning only sanctioned channels and declaring the inventory complete. If you did not check OAuth grants, expense reports and engineering interviews, you missed agents.
• Classification paralysis. Spending three weeks debating whether an agent is “medium” or “high” risk. Use the simple matrix above. Perfection is the enemy of progress.
• No quick wins. If you finish Phase 1 without revoking a single credential or disabling a single agent, you have done an audit, not governance.

Phase 2: Foundation (days 31-60)#

The goal of this phase: build the policy framework, select tooling and prove governance works on a pilot cohort of your 10 highest-risk agents.

Week 5-6: Policy framework#

Write four documents. Not 40. Not a 200-page governance manual. Four:

1. Agent registration policy. What information must be provided before an agent is approved for deployment. Minimum: owner, purpose, data classification, risk tier, dependencies, access requirements.

2. Access control policy. How agents receive credentials, how those credentials are scoped to least-privilege, how often they are rotated and what triggers revocation.

3. Review and monitoring policy. How often agents are reviewed (monthly for high-risk, quarterly for low-risk), what continuous monitoring covers and what triggers an ad-hoc review.

4. Decommissioning policy. The process for retiring an agent, including dependency audit, credential revocation, state preservation and compliance documentation.

Each policy should be under three pages. Write them for engineers, not lawyers. If the people who deploy agents will not read them, the policies do not exist.

Week 6-7: Tooling selection#

You need three capabilities. They can be one platform or three separate tools:

• Agent registry. A centralized inventory with mandatory fields, ownership tracking and dependency mapping. This replaces the spreadsheet from Phase 1.
• Policy enforcement. Automated checks that validate agents against your policies at registration, deployment and runtime. Integration with CI/CD is non-negotiable. If governance is a manual gate, engineering will route around it.
• Monitoring and alerting. Continuous observation of agent behavior, credential usage, data access patterns and drift from approved configurations.

Evaluation criteria, in priority order:

• Time to first value (can you onboard 10 agents in a week?)
• CI/CD integration (does it fit in the existing deployment pipeline?)
• Compliance evidence generation (can it produce audit-ready reports?)
• Scalability (will it handle 500 agents?)
• Total cost of ownership

Week 7-8: Pilot governance for top-10 agents#

Select your 10 highest-risk agents. Apply the full governance process:

• Register each in the new registry with complete metadata
• Review and adjust access controls to match least-privilege
• Configure monitoring and alerting
• Run the first formal review against the review policy
• Document findings and remediation actions

This pilot serves two purposes. First, it validates that the policies and tooling work in practice. Second, it creates a reference implementation that other teams can follow. When you tell the marketing team “here is how the fraud detection team governs their agents,” you have a concrete example, not an abstract policy.

Phase 2 deliverables checklist#

• Four governance policies (registration, access control, review, decommissioning)
• Tooling selection decision with evaluation rationale
• Tooling deployed and configured for pilot cohort
• 10 highest-risk agents registered and governed
• First formal review completed for pilot cohort
• CI/CD integration demonstrated (at least one pipeline)
• Phase 3 onboarding plan and timeline

Phase 2 RACI#

Activity	CISO	Governance Lead	Engineering	Legal
Policy framework	A	R	C	C
Tooling selection	C	R	R	I
Pilot governance	A	R	R	I
CI/CD integration	I	C	R	I

Phase 2 failure modes#

• Policy perfection: spending six weeks on policy revisions instead of getting 80% right and iterating; ship version 1.0 and update it after the pilot reveals what is missing
• Tooling analysis paralysis: evaluating 12 vendors for eight weeks when you should set a two-week evaluation window with clear criteria, make a decision and move on
• Pilot avoidance: selecting 10 low-risk agents for the pilot because they are easy, when you should pick the hard ones (if governance cannot handle your riskiest agents, it cannot handle anything)

Phase 3: Scale (days 61-90)#

The goal of this phase: onboard the full agent estate, automate enforcement, map to compliance frameworks and deliver the first executive report.

Week 9-10: Full estate onboarding#

Take the process you validated with 10 agents and roll it out to the rest. Prioritize by risk tier:

• Critical and high-risk agents first (week 9)
• Medium-risk agents next (week 9-10)
• Low-risk agents last (week 10)

For each tier, the onboarding process is the same:

• Register in the agent registry
• Validate access controls
• Configure monitoring
• Assign review schedule

At scale, this only works if registration is self-service with automated policy checks. If every agent requires manual review from the governance team, you will create the bottleneck that drives shadow adoption.

Week 10-11: Automated enforcement#

Move from manual policy checks to automated enforcement:

• At registration: Automated validation rejects agents that do not meet minimum metadata requirements
• At deployment: CI/CD pipeline checks verify risk classification, access scope and monitoring configuration before an agent reaches production
• At runtime: Continuous monitoring flags drift from approved configurations, unauthorized data access and credential anomalies
• At decommissioning: Automated workflows trigger credential revocation and compliance documentation

The goal is that governance runs without a human in the loop for low and medium-risk agents. Humans review high-risk decisions and exceptions.

Week 11-12: Compliance mapping and executive reporting#

Map your governance controls to the compliance frameworks that matter to your organization:

• EU AI Act: agent registry satisfies Article 12 record-keeping, risk classification maps to the Act’s risk tiers and monitoring covers Article 9 risk management
• SOC 2: agent access controls map to CC6 (logical and physical access controls), monitoring maps to CC7 (system operations) and review cadence maps to CC4 (monitoring activities); see the full mapping in our SOC 2 guide
• ISO 42001: agent lifecycle management maps to the AI management system requirements
• Industry-specific: HIPAA for healthcare agents, PCI DSS for payment-related agents and SEC/FINRA for financial services

Then build the executive dashboard. Your board and C-suite need to see:

• Total agent count (registered vs. estimated unregistered)
• Risk distribution (how many agents at each tier)
• Compliance posture (percentage of agents meeting policy requirements)
• Open findings (unresolved governance issues, by severity and age)
• Trend lines (are things getting better or worse?)

Week 12: Operational runbook#

Document the steady-state operating procedures:

• How new agents are registered and approved
• How reviews are scheduled and conducted
• How incidents are reported and escalated
• How agents are decommissioned
• How the governance program itself is reviewed and improved

This runbook is the difference between a project (which ends) and a program (which persists). Without it, governance degrades the moment the implementation team moves on to the next initiative.

Phase 3 deliverables checklist#

• All discovered agents registered and governed
• Automated enforcement active for registration, deployment and runtime
• Compliance mapping documented (EU AI Act, SOC 2, ISO 42001, industry-specific)
• Executive dashboard operational
• Steady-state operational runbook published
• First monthly governance report delivered to executive sponsor
• Post-implementation review completed with lessons learned

Phase 3 RACI#

Activity	CISO	Governance Lead	Engineering	Compliance	Executive Sponsor
Full estate onboarding	A	R	R	I	I
Automated enforcement	I	R	R	C	I
Compliance mapping	C	R	I	R	I
Executive reporting	I	R	I	C	A
Operational runbook	A	R	C	C	I

Phase 3 failure modes#

• Onboarding backlog: trying to register 200 agents manually in two weeks, when self-service registration with automated validation is the only way this works at scale
• Compliance theater: mapping controls to frameworks without verifying they work; run a simulated audit before you tell the board you are compliant
• No steady-state plan: declaring victory at day 90 and disbanding the team, when governance is not a project (budget for ongoing headcount of 2-3 FTEs per 100 agents in a hub-and-spoke model)

Decision points along the way#

At each phase transition, stop and make explicit decisions:

Day 30 decision: Go/no-go on Phase 2.

• Do we have a complete enough inventory to proceed?
• Are the quick wins generating organizational support?
• Is the executive sponsor still engaged?

Day 60 decision: Go/no-go on Phase 3.

• Did the pilot validate our policies and tooling?
• Can the tooling handle full-estate onboarding?
• Do we have the engineering resources for CI/CD integration?

Day 90 decision: Transition to steady state.

• Are we confident in the compliance mapping?
• Is the executive dashboard telling a coherent story?
• Do we have the headcount and budget for ongoing operations?

If the answer to any critical question is no, extend the phase rather than rushing forward with gaps. A 120-day implementation that works beats a 90-day implementation that collapses at day 91.

Sources#

Source	Date	URL
Grant Thornton, 2026 AI Impact Survey	2026	https://www.grantthornton.com/services/advisory-services/artificial-intelligence/2026-ai-impact-survey
ElevateCX, AI governance RACI matrix	2025	https://elevateconsult.com/insights/designing-the-ai-governance-operating-model-raci/
Gartner, 40% of agentic AI projects canceled by 2027	Jun 2025	https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
Microsoft, Agent Governance Toolkit	Apr 2026	https://opensource.microsoft.com/blog/2026/04/02/introducing-the-agent-governance-toolkit-open-source-runtime-security-for-ai-agents/
Nudge Security, AI agent discovery	Mar 2026	https://securityboulevard.com/2026/03/ai-agent-discovery-inventory-and-govern-shadow-ai-agents-nudge-security/
Gravitee, State of AI Agent Security 2026	2026	https://www.gravitee.io/blog/state-of-ai-agent-security-2026-report-when-adoption-outpaces-control
OneTrust, Responsible AI governance guide	2026	https://www.onetrust.com/blog/responsible-ai-in-2026-a-3-step-guide-for-governance-that-scales/