The Agent Governance Playbook
What enterprises actually ask for when they go looking for control over their AI agents: six requirements, a maturity ladder from alerting to enforcement, and the twelve questions that show where your agent governance stands today. Drawn from working sessions with AI platform, finance, and technology leaders at global enterprises.
- The six requirements, in the buyers’ own words
- Why seat-based plans break gateway-only governance, and what fits instead
- The privacy conditions that decide whether a deal is even possible
- A 12-question self-assessment of your agent governance today
11 pages · PDF · Free
The six things every enterprise asked for
Different enterprises, different entry points: an AI platform team, a technology CTO fifteen vendors into an evaluation, a fintech CTO staring at the AI bill, and an architecture group standardizing dozens of vendors. All of them arrived at the same spec without being led there. This is that spec.
Observability
"Why did the agent answer that?", answerable at the user, model, and tool level by a non-engineer.
The agent inventory
Owner, creator, model, access, and change history for every agent, exportable for audit.
Centralization
One pane over every agent, every cloud, every tool: built in-house and bought from vendors alike.
FinOps
Cost by department, application, agent, and person, including seat-based plans no gateway can see.
Alerts that land
Cost thresholds, quality drops, PII and prompt-injection attempts, delivered to email, Slack, and Teams.
Built for the explosion
Two agents today, thirty next year: the layer enterprises deliberately buy before the sprawl.
Can you list every AI agent running in your company right now?
Not the approved ones. The running ones. Most governance programs start with policy documents, but you cannot govern what you cannot list: which tools, which agents, who owns each one, what it costs, and whether anyone actually uses it.
| Tool | Seats | Licence / mo | Idle / mo | Usage · 30d | Top department |
|---|---|---|---|---|---|
| Claude Code | 44/60 | $6,000 | $1,600 | $4,310 | Engineering |
| Claude Cowork | 22/30 | $900 | $240 | ~$41 | Customer support |
| Copilot Studio | 6/15 | $3,000 | $1,800 | ~$493 | Customer support |
| GitHub Copilot | 87/120 | $2,280 | $627 | ~$98 | Engineering |
| ChatGPT Enterprise | 52/80 | $2,000 | $700 | $163 | Data & AI |
| Cursor | 31/40 | $800 | $180 | $625 | Engineering |
| Databricks Genie | - | - | - | ~$1,051 | Data & AI |
| Custom agents | - | - | - | $5,112 | Engineering |
Illustrative figures. The inventory reconciles what the company bought against what people actually use, who owns it, and the initiative behind it; every figure says where it came from and how fresh it is.
“It has been very difficult for us to have observability of what people are using.”
“A director will ask me: why did this agent answer that? I need to see it quickly, at the level of the user, the model, and the tool.”
“I saw the latest AI bills, but I have no clue where all this is used, where all of this is actually going.”
Quotes from recorded enterprise working sessions, translated where needed and anonymized at the participants’ preference.
Where does your agent governance stand today?
The twelve questions in this playbook are a self-assessment for CTOs and AI leaders: can you list every agent, attribute every dollar, and trace every answer? Get the PDF, find the gaps, and decide what to uplift first.