Centralise the machinery and devolve the decisions. One company-wide policy, one register of agents, one enforcement layer that every agent's actions pass through, and one record of verdicts. Within that, each agent belongs to the owner of the business process it works in, who sets its limits and answers for its outcomes, just as they do for their team.
Why does company-wide governance break down?
Usually because agents arrive from several directions at once. Finance buys one inside its accounting software. Customer service builds one. A team in operations connects a model to a spreadsheet. Each is governed, if at all, by whoever set it up, under different rules, with records in different places.
Within a year, nobody can answer the board's simplest question: what can our agents do, and who said so?
What should be central, and what should be local?
| Element | Central | Local (process owner) |
|---|---|---|
| Policy | What agents may never do; approval of new agents; standards for testing and records | How the policy applies to their process |
| Register | One list of every agent, owner, capabilities and depth | Keeping their entries accurate |
| Enforcement | One layer checking every action against bounds | Setting their agent's limits |
| Platform floor | Actions withheld for every agent | None |
| Testing standard | What "certified" means | Supplying their own cases |
| Record | One append-only store of verdicts | Reviewing their agents' holds and refusals |
| Oversight | Risk, internal audit, board reporting | Answering for outcomes |
The pattern mirrors how companies already govern people. Central HR sets policies; managers run their teams. Our answer on who owns an AI agent's decisions in a company explains why the process owner, not IT, carries the outcome.
What are the practical steps?
- Find every agent. Survey teams, check software licences and look for service identities calling model providers. Register what you find.
- Assign an owner to each. If nobody will own it, it stops.
- Write the company policy. Short: what no agent may do, how agents are approved, tested, recorded and retired.
- Route every agent through one enforcement layer. Agents should not hold direct credentials to business systems.
- Set a common testing standard. Each capability earns depth on the owner's own cases.
- Report centrally. Risk and internal audit see all agents through the same record.
- Report to the board on authority, not technology.
Can each vendor's built-in controls do this?
No. Vendor and model guardrails are useful, but each covers only its own product, with its own rules and records. A company with five vendors has five control designs and no single view. Our answer on whether a model's own guardrails can govern an AI agent explains the gap. The principle that holds across all of them is to govern the action, not the model.
A worked example
Take a mid-sized insurer with agents in claims, underwriting support, finance and HR. A risk review finds nine agents, three of which nobody had registered, including one in HR that drafted responses to grievances using a personal model account.
The insurer registers all nine, assigns owners, and stops the HR agent until it can be rebuilt under the policy. It routes the remaining eight through one enforcement layer, so claims and finance agents meet the same kind of checks and file to the same record. The chief risk officer reports quarterly to the board: nine agents, their owners, what each may do and at what depth, refusals and holds by type, and any incidents.
Which controls make the answer hold?
- No agent outside the register. Unregistered agents are found and stopped.
- One enforcement layer. No direct access around it.
- Owners with real authority, who can tighten limits instantly.
- A framework above the controls. The NIST AI RMF Govern function for AI agents gives a structure for roles, policies and accountability.
- Board oversight of what agents may do. See boards should govern agent authority, not agent technology.
The topic guide, agent governance is a bound, not a model, sets out the principles behind each of these.
Where Pryme fits
Pryme is the layer between models and the business. Every agent, whichever model it runs on, meets the same five bounds at the moment of action, and every verdict is filed in one append-only record. Owners set their own limits; the platform floor is the same for all.
Questions people ask
Who should own the company-wide policy?
Usually the chief risk officer or an equivalent senior manager, with the board approving it. Technology supports it but should not own it.
Do we need a central AI team?
A small one helps: to run the register, the testing standard and the enforcement layer. It should not own agents in the business.
How do we handle agents inside software we buy?
Register them, assign an owner, and route their actions through your enforcement layer where possible. Where not, limit what the software's agent can reach.