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

Personalized AI infrastructure

Three Human Roles Behind Every Company AI Agent: Owner, Steward, Approver

Name the three human roles around your AI agents: an agent owner, a memory steward and an approver, so every memory change and permission has an owner.

Be Human typographic cover with the title: Three Human Roles Behind Every Company AI Agent: Owner, Steward, Approver
Conceptual illustration created with AI · Be Human

A company that runs AI agents needs three named human roles around them. The Agent Owner decides what the agent is for, the Memory Steward controls what the agent remembers and how that memory changes, and the Approver signs off on actions above a stated risk level. When these roles are written down, every memory edit and every new permission traces back to a person who agreed to it, which is what leaders need when something goes wrong.

Key takeaways

  • Every agent needs an owner who is accountable for its purpose and results, not only for whether it is running.
  • Memory changes are decisions. Assign one steward who approves what is added, corrected or deleted.
  • Approvers should work from written thresholds, so a yes is a defined act and not a favor.
  • Roles can be held by existing staff, but each needs a written scope, a named backup and a review date.
  • A role without a named person is a risk. Check all three roles against the agent's actual activity at least once a quarter.

Why every agent needs three named roles

An AI agent is not a static document or a simple tool. It reads records, keeps notes from earlier conversations, calls other systems and sometimes acts on behalf of the business. Each of those capabilities changes what the company is exposed to. If nobody is named for the agent's purpose, its memory and its permissions, change happens by drift: a team lead asks for one more integration, an employee corrects a customer note inside the agent's memory, a developer widens a connection to make a test pass. Each change may be reasonable on its own. Together they produce an agent that nobody can fully explain.

Three roles fix this without creating a committee. In a company of ten people, the roles may sit with two or three staff members. In a larger business, they map to departments. The point is not the org chart. The point is that each memory change and each permission has a name attached and a stated reason. A broader governance structure can sit above these roles, and the committee charter guide explains how that layer can stay light enough for a small company to run.

What the Agent Owner decides

The Agent Owner is the person accountable for the agent's job. This person writes the one-sentence purpose, names the users, defines the observable success criterion and decides when the agent should be expanded, paused or retired. The owner is usually a leader close to the work: an operations manager for a support agent, a finance lead for a reconciliation agent, or a clinic administrator for appointment reminders.

The owner does not need to understand model internals. The owner needs to answer five questions in writing: what the agent is for, what it must never do, who depends on its output, how success will be measured, and what happens when it fails. If the owner cannot answer these, the agent is not ready for production. The proof of concept, pilot or production stage check is a useful place to force the question before anything goes live.

What the Memory Steward controls

Memory is the part of an agent that most often goes wrong quietly. An agent that remembers a customer's preferred contact channel is useful. An agent that remembers an outdated refund amount will repeat that error in every conversation that touches it. The Memory Steward owns the quality and the boundaries of what the agent keeps.

Concretely, the steward approves what enters long-term memory, decides how corrections are made, sets retention rules, and separates confirmed facts from opinions or unverified notes. The steward also decides what must not be remembered at all, such as personal data the business has no reason to keep or confidential know-how that belongs in a controlled store. Stewards should work from a written list of memory sources, with one rule for each source stating whether the agent may write to it. When answers start to drift, the steward is the first person to check whether memory caused the change. Memory changes should follow the same discipline as other changes, which is why the change request process for managed agents covers memory as well as connections and prompts.

What the Approver signs off

The Approver decides whether an action the agent proposes may go ahead. This is where the business sets its risk boundary. A common pattern is to let the agent draft, classify and recommend freely, require approval before anything reaches an outside party, and require a second approval before any payment, deletion or change to customer records. The exact thresholds depend on the business. They should start low and be raised only after evidence builds up, as described in the guide on approval thresholds for superintelligent agents.

Permissions belong to the same role. Every tool an agent can call is a permission, and each should be granted for a specific purpose with a named reviewer. Standard authorization practice applies directly. The [OWASP authorization guidance](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) describes the principle of granting only the access a task needs and checking access on each request. The Approver reviews the permission list at a fixed interval, following the routine in the quarterly permission review guide.

How to assign the three roles in five steps

Use this sequence in one working session with the leaders who will hold the roles.

  • List every agent in use or planned, with one line describing what it does and who uses it.
  • For each agent, record the Agent Owner's name and the five written answers about purpose, limits, dependents, success and failure.
  • Name a Memory Steward and list each memory source the agent reads or writes, with a yes or no on whether the agent may change it.
  • Name an Approver and write the three action levels: draft freely, approve before any external contact, approve twice before money, deletion or record changes.
  • Name a backup for each role, set the date of the next review, and store the one-page record where the team can find it.

Keep the record short. A single page per agent is enough for most small companies. Operating instructions and stop rules can live in a separate runbook, such as the one-page runbook template.

Illustrative scenario: a regional services company runs a sales assistant agent. The sales director is the Agent Owner, and the purpose is answering product questions and preparing draft quotes. The operations coordinator is the Memory Steward, approving what the agent may store from customer calls and removing outdated pricing notes. The finance lead is the Approver, who must approve any quote above an agreed value before it reaches a customer. When a sales representative notices an old discount in the agent's memory, the steward traces the note to its source, removes it and records the reason. Nobody has to rebuild the event from chat logs weeks later.

How managed operations keep the roles working

Naming roles is the easy part. They decay when people change jobs or when agents change behavior. Be Human designs, implements and operates personalized AI infrastructure in which an agent's identity, memory, channels, tools, permissions and operating context stay under human control. A first engagement defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion, and the three human roles are written into that scope from the start. In managed operations, monitoring and recurring evaluations show whether memory and permissions still match the agreed purpose, and changes route through the named owner, steward and approver. Records of agent actions, of the kind described in the [OpenTelemetry observability primer](https://opentelemetry.io/docs/concepts/observability-primer/), give stewards and approvers evidence instead of recollection.

The first step is to find the one task worth handing to an agent and the roles that task would need, and the free Be Human AI diagnostic can help you identify it.

Frequently asked questions

Can one person hold all three roles? It is possible in a very small company, but it weakens the control. The person who benefits from a change should not be the only one who approves it, so separate the Approver from the Owner wherever you can and document any exception.

Does the Memory Steward need technical skills? The steward needs judgment about business facts, not engineering skill. The steward must know which records are accurate, which are outdated and which should never be stored, and engineers can carry out the changes the steward approves.

How often should the three roles be reviewed? Review them at each permission review and whenever a person in one of the roles changes jobs or leaves. A quarterly cadence works for most businesses because it catches drift before it becomes a habit.

Does this structure satisfy legal requirements for AI systems? Named roles are an operating practice, not a legal conclusion. Frameworks such as the [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) describe governance functions that these roles help support, and rules such as the EU Artificial Intelligence Act may apply depending on where you operate and what the system does, so confirm your obligations with qualified counsel.

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 →