Skip to content
be-human.ai
← Back to the journal

Personalized AI infrastructure

Superintelligence Governance: A Committee Charter for Small Companies

A short committee charter for small companies: who approves agent scope changes, who reviews access, and who can stop an agent, with a first-month plan.

Be Human typographic cover with the title: Superintelligence Governance: A Committee Charter for Small Companies
Conceptual illustration created with AI · Be Human

Superintelligence governance, in practice, is the set of decisions a company makes before its AI agents can do more: who approves a change to what an agent may do, who reviews the access it holds, and who can stop it when something goes wrong. A small company does not need a board or a compliance department to do this well. A committee of three or four named people, a one-page charter, a fixed meeting rhythm and clear stop rules covers most of the ground.

Key takeaways

  • Give the committee three jobs: approve scope changes, review access, and hold the authority to stop an agent.
  • Keep the group to three or four people, with one named chair who makes the final call.
  • Write down what an agent may do without asking, so routine work moves and unusual work is visible.
  • Review access on a fixed schedule and keep the evidence outside the agent.
  • Test the stop procedure while nothing is wrong, so it works when something is.

Why a small company needs a committee at all

An AI agent rarely gets dangerous through one large decision. It gets broader through many small ones: a read permission added for a report, a draft tool connected to the inbox, a shared folder opened so the agent can answer a question faster. Each change looks reasonable on its own. Taken together, they can leave an agent holding access that nobody remembers approving.

Without a named group, these decisions default to whoever built the agent or whoever asked for the change most recently. That is the real governance gap in small companies, and it grows as the agent becomes more capable. A committee does not slow the work down if it only handles the decisions that matter. Its purpose is to keep judgment with people who can be held accountable.

The NIST AI Risk Management Framework organizes this work around governing, mapping, measuring and managing risk. A small charter can borrow the governing part without adopting the full framework. You need to know who decides, on what basis, and where the record lives.

Who sits on the committee and how decisions are made

Three or four people is enough. The roles matter more than the titles:

  • An executive sponsor who owns the budget and the final decision when members disagree.
  • An operations lead who knows the process the agent touches and can judge whether a change fits it.
  • A technical owner, internal or external, who understands the tools, permissions and integrations.
  • A person from finance, legal or data protection when the agent handles money, customer personal data or contracts.

Decide in advance how the committee works. Name one chair. Agree that a decision needs the chair plus one other member, and that the sponsor can overrule a decision only in writing, with a reason. Keep minutes short: what was requested, what was decided, who approved it, and what changes in the permission list as a result.

The charter: what it must contain

A charter for a small company fits on one or two pages. Use this checklist to draft it:

  • Purpose and scope: which agents are covered and what each one is for, in one sentence.
  • Members, chair and quorum, with names and backups for each role.
  • Decisions reserved for the committee, such as new tools, new data sources, any action that sends something outside the company, and any spending commitment.
  • Decisions delegated to the operator, such as adjusting wording, changing a schedule within limits, or pausing a single task.
  • Access review cadence, the owner of each review and the evidence it produces.
  • Stop authority: who can pause an agent, who can suspend it, and who must approve a restart.
  • Records: where decisions, permission lists and incident notes are stored, and who can read them.
  • Review date for the charter itself, so the document does not go stale without anyone noticing.

Keep the charter in a place the agent cannot edit. If the agent can rewrite its own rules, the charter stops being a control.

Approving scope changes

Scope changes are the most common decision and the easiest to mishandle, because most of them look small. A simple way to handle them is to classify each request by what the agent could do after the change:

  • Reads information that is already available to the team: the operator can approve, and the change is logged.
  • Drafts output that a person reviews before it leaves the company: the operations lead approves.
  • Writes to an internal system or changes a record: the committee approves in writing.
  • Sends a message, payment or commitment outside the company, or touches customer personal data: the committee approves, with the data protection or finance member involved.

The Be Human approach to change requests asks for the same information each time: the business reason, the exact permission being added, the data it reaches, the test that will show it works, and the way it can be reversed. For a fuller version of this flow, see how to request and approve changes to a managed AI agent.

Illustrative scenario: a small training company has an agent that drafts replies to course enquiries for a human to send. The operations lead proposes letting the agent send simple confirmations directly. Under the classification above, this moves the agent from drafting to sending outside the company, so the committee decides. It asks for a test set of sample enquiries, a rule that the agent never confirms refunds or prices, and a way to switch sending off within minutes. The change is approved for confirmations only, and the decision is written into the charter's scope section.

Reviewing access and stopping an agent

Access reviews are where governance becomes concrete. Once a quarter, the technical owner exports every tool, data source and permission the agent holds, and the committee compares that list with the approved scope. Anything not in the approved scope is removed or justified in writing. Anything with an expiry, such as access granted to a contractor, is checked against its end date. The OWASP authorization guidance recommends deny by default, and this is the practical form of that idea: each tool starts closed and opens only through a recorded decision. For the steps and the evidence to keep, see the quarterly AI agent permission review.

Stopping an agent needs its own procedure, and it should not depend on the agent's judgment. The OWASP Top 10 for LLM Applications lists excessive agency among the main risks, which is one reason a stop rule must sit outside the agent. Define three levels:

  • Pause a single tool or action type while the team investigates.
  • Pause all outbound actions, so the agent can still read and draft.
  • Suspend the agent entirely, revoking its credentials until the committee approves a restart.

Any committee member and the operator can invoke the first level without waiting for approval. Only the committee can lift a suspension. Rehearse the procedure once after the charter is written and again after major changes, and note how long each step took.

How a managed approach carries the work

Be Human designs personalized AI infrastructure in which an agent keeps its identity, memory, channels, tools, permissions and operating context under human control. A first engagement defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion, which gives the committee a written scope to govern from. In managed operations, Be Human can handle monitoring, recurring evaluations, permission review, integration maintenance and incident response, and it brings the committee the evidence needed for each decision rather than asking members to reconstruct what happened. The committee still decides. The managed service keeps the record and the mechanics current.

Starting the committee this month

A practical sequence for a small company that has agents running, or about to start:

  • Write a list of every agent in use or planned, with the tools and data each one touches.
  • Name the four roles and the chair, and confirm that each person accepts the role.
  • Draft the charter using the checklist above, and keep it to two pages.
  • Classify each current permission against the four scope levels and flag anything that needs a decision.
  • Schedule the first access review and the first stop rehearsal on the calendar before the charter is signed.
  • Agree that any change proposed outside the committee is paused until it is classified.

Governance grows with the agent. Start with the decisions you already face, and the charter will show you where it needs more structure. If you are unsure which task should go to an agent first, the free Be Human AI diagnostic can help you find it.

Frequently asked questions

Does a small company need a formal AI board to govern its agents? No. A committee of three or four people with a short written charter is enough for most small companies. What matters is that the decisions are named, the chair is clear, and the records are kept somewhere the agent cannot change them.

Who should be allowed to stop an agent? Any committee member and the operator should be able to pause an agent or a single tool at once, without waiting for approval. Restarting a suspended agent should require the committee, so that a pause does not quietly become a change of scope.

How often should access be reviewed? Quarterly is a reasonable default for a small company, with an extra review whenever a tool, contractor or data source is added or removed. Each review should compare the current permission list against the approved scope and record what was removed or justified.

What if an agent's scope has already grown without approval? Pause the expanded permissions, write down what the agent has been doing with them, and bring the case to the committee as a scope change. Decide whether to approve, narrow or remove each permission, then update the charter so the gap does not reappear.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →