Personalized AI infrastructure
Superintelligence Company Roles: Who Staffs the Agent Layer
Who owns, reviews, operates and approves a company's AI agents, and what each role must decide before an agent gets more permissions.

Superintelligence company roles are the people who own, review, operate and approve the AI agents a business depends on. A company needs four distinct responsibilities around each agent: an owner who sets its purpose, a reviewer who checks its work, an operator who keeps it running, and an approver who decides when its permissions grow. Naming these roles before an agent gains access to more systems is what keeps its memory, tools and channels under human control.
Key takeaways
- Every agent needs four named responsibilities: owner, reviewer, operator and approver. In a small company one person can hold two of them, but the duties should still be written down separately.
- The owner defines the job and its success criterion. The approver decides permission changes, and the person who asks for a change should not be the only one who approves it.
- Reviewers check outputs against a standard written before launch, not against impressions collected after something goes wrong.
- Permissions should expand only after the roles have run for a defined period and produced evidence that the current boundaries work.
- Each role belongs in a one-page runbook for the agent, so the responsibilities survive staff changes and vendor changes.
Why the agent layer needs named roles
An AI agent is not a single tool that someone switches on and forgets. It holds memory of your customers or operations, calls external systems, sends messages through channels, and works inside permissions that someone granted. Each of those elements can change over time. A new integration gets added because it was convenient, a memory source gets widened because a team asked for better answers, and a channel gets enabled because a pilot went well. When nobody is named for these decisions, each change looks reasonable on its own, and the agent slowly ends up with more reach than anyone intended.
The NIST AI Risk Management Framework treats governance, accountability and ongoing monitoring as core parts of managing AI risk, not as paperwork added at the end. Access control guidance from OWASP makes a similar point from the security side: permissions should be checked against explicit rules, and the rules should be maintained by people who are accountable for them. Roles are the organizational form of those rules. They tell you who is allowed to decide, who is expected to notice a problem, and who must act when something goes wrong.
The four roles and what each one decides
The owner is the business leader accountable for what the agent is for. The owner writes the purpose in one or two sentences, names the success criterion that can be observed, and decides which customer records, documents and systems the agent may use at all. The owner does not need to read logs every week, but the owner must be able to say whether the agent is doing its job and whether it should keep doing it.
The reviewer checks the quality of the agent's work against the standard the owner defined. In practice this means sampling outputs on a schedule, recording cases where the agent was wrong, unclear or out of scope, and deciding which of those cases require a change. The reviewer is often a subject expert in the business process, such as a support lead or an operations manager, rather than a technical person. The reviewer also watches for unusual inputs, because agents that read external text can be steered by instructions hidden in that text, a risk that the OWASP Top 10 for LLM Applications describes in detail.
The operator keeps the agent running. The operator monitors availability, reads the event trail, maintains integrations when a connected system changes, and handles incidents. The operator can pause the agent when a clear stop rule is triggered, but the operator cannot unilaterally widen what the agent is allowed to do.
The approver decides when permissions, memory sources, tools or channels change. The approver should be someone with authority over the business area affected, and for anything involving money, personal data or outbound messages to customers, the approver should be a person who can be held accountable for the consequences. The approver is the role most often missing in early deployments, and its absence is the usual reason agents accumulate access quietly.
The owner decides what the agent is for, the approver decides what it may touch, and neither of them has to do the operator's daily work.
A decision guide for who approves what
Not every change needs the same level of scrutiny. A practical approach is to sort decisions into tiers and assign an approver to each tier before the agent goes live. The exact tiers depend on your business, but the following decision guide is a sound starting point.
- Low impact: changing wording of responses, adding a reference document that is already public inside the company. The operator may make the change, and the reviewer checks it in the next sample.
- Medium impact: connecting a new internal data source to the agent's memory, adding a read-only tool. The owner approves after the reviewer confirms the source is accurate and appropriate.
- High impact: enabling write actions in a business system, allowing outbound messages to customers, changing spending or refund limits. The approver signs off in writing, and the change is tested in a limited setting first.
- Restricted: anything involving sensitive personal data, regulated decisions or changes to the agent's own authority rules. These require the approver plus a second review by a person outside the operating team.
Be deliberate about thresholds. Raising an approval threshold because a tier feels slow is a decision in itself, and it deserves the same evidence as any permission expansion. The related guidance on approval thresholds for superintelligent agents explains why such thresholds should move slowly.
Step-by-step: staffing the agent layer before expansion
Use this sequence before the first agent receives any permission beyond read-only access.
- Step 1: Write the agent's purpose and one observable success criterion. If you cannot describe how you would know it works, the role is not ready.
- Step 2: Name the owner, reviewer, operator and approver by person, not by department. Write the names and a backup for each role.
- Step 3: List every memory source, tool and channel the agent currently has, and mark which ones are read-only and which can change something.
- Step 4: Assign the decision tiers from the guide above and record who approves each tier.
- Step 5: Define stop rules: the conditions under which the operator pauses the agent and notifies the owner. Keep them specific, such as a repeated refusal to follow a boundary or an integration returning errors.
- Step 6: Run the agent with the current permissions for a defined period, and have the reviewer log results against the success criterion.
- Step 7: Only after the evidence supports it, let the approver consider expanding one permission at a time, starting with the change of lowest impact.
Keep the output of these steps in a single runbook page for each agent. A runbook that lives in someone's inbox will not survive a staff change, and a role that exists only in memory will be forgotten as soon as the person who set it up moves on.
Illustrative scenario: a regional services company deploys an agent to answer customer questions about scheduled visits. The owner is the operations director, the reviewer is a senior dispatcher, the operator is an IT contractor on call, and the approver is the general manager. Six weeks in, a dispatcher notices that the agent quotes a visit window that the scheduling system no longer uses. The reviewer logs the case, the operator checks the integration and finds a field was renamed, and the owner decides the agent should be paused for that topic only. The general manager approves the fix before the agent resumes answering questions about visit windows. Nothing in that sequence required a new tool or a larger budget. It required that each person knew which decision was theirs.
Separating the operator from the approver
The most common governance gap in small companies is that the person who wants a change also approves it. The operator asks for a new integration because it would reduce manual work, and the same operator marks it approved because there is no one else available. This is not a failure of intent, but it removes the check that approval is meant to provide. Separate the two duties even if the company is small. If the only available approver is also the operator, write that exception down, limit it to low-impact changes, and schedule a review of the exception.
The change request process should record who asked, who reviewed, who approved and what evidence was attached. The guide on how to request and approve changes to a managed AI agent describes a workable format. Keeping that record is also what allows a later audit to answer the question of why an agent can do something today that it could not do last year.
How Be Human's managed approach staffs the layer
Be Human designs, implements and operates personalized AI infrastructure in which each 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 that role needs, the approval boundaries around it, and an observable success criterion. Managed operations then provide the operator function through monitoring, recurring evaluations, integration maintenance, permission review and incident response, while the client keeps the owner and approver decisions. Changes move through a documented request process, and the reviewer checks agent outputs on a schedule that the client agrees to. The aim is that the company always knows who decides, who checks and who can pause the work.
Keeping the roles current
Roles drift. The owner changes departments, the reviewer goes on leave, and an integration is replaced by a system nobody has connected to the agent before. Schedule a short review each quarter in which each role holder confirms they still hold the responsibility, that the stop rules still match the business, and that every permission the agent holds still has an owner who can justify it. The quarterly permission review checklist gives a format for that session. When the review shows that an agent has the right roles and a clear record, you have evidence to make the next decision well. If you are not sure which of your processes is the right first task to hand to an agent, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Do I need four different people for one small AI agent? You do not necessarily need four people, but you do need four distinct responsibilities with a named person for each. A company of ten people often assigns the owner and approver to the same leader, the reviewer to a team member who knows the process, and the operator to an IT contractor or managed provider. The critical rule is that the person requesting a change should not be the only person approving it.
When should the approver be someone outside the operating team? Any change involving sensitive personal data, regulated decisions, outbound messages to customers or the agent's own authority rules should receive a second review from someone outside the team that operates it. This protects the company from a blind spot that forms when the same people build, run and approve the same system. For low-impact changes, an internal approver is usually enough.
How do the roles change when an agent is expanded to more channels or locations? Each new channel or location creates new decisions that need an owner and an approver, even if the agent's core purpose stays the same. The owner of the original agent may need a local counterpart who understands the specific customers or rules. Write those additions into the runbook before the expansion goes live, so the permissions are not granted first and assigned later.
What should happen if a role holder leaves the company? Name a backup for every role when the agent is set up, and update the runbook within days of any departure. Until the replacement is confirmed, the operator should keep the agent within its current permissions and avoid approving changes. Treat a departure as a reason to run an early review rather than waiting for the next quarterly session.
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.


