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

Personalized AI infrastructure

Superintelligence Company Org Chart: Who Owns Agents, Memory, Tools

A practical org chart for a company running AI agents: who owns agents, memory, tools and approvals, and how to name one accountable person for each.

Be Human typographic cover with the title: Superintelligence Company Org Chart: Who Owns Agents, Memory, Tools
Conceptual illustration created with AI · Be Human

A superintelligence company org chart names one accountable person for each part of the company's AI operation: the agents, the memory they draw on, the tools they can call, and the approvals that decide when an agent must stop and ask. These owners do not need to be engineers. They need the authority to say what an agent may do, the evidence to see whether it is doing it well, and a standing review slot to change either.

Key takeaways

  • Four roles carry most of the weight: agent owner, memory steward, tool owner and approval owner, with an executive sponsor above them.
  • Each role needs decision rights, not only technical skill, so the owner can change scope, pause the agent and sign off on exceptions.
  • Memory is the role most often left unowned, yet it decides what an agent believes about your customers, orders and policies.
  • Every tool permission should trace back to a named approver and a review date.
  • Ownership is only real when the owner can read the action log, the evaluation results and the list of recent changes.

Why capability growth changes who needs to own what

When an assistant only drafts text for a person to read, a mistake costs the time it takes to notice and correct it. When an agent can read a customer record, update a ticket, send a message or start a workflow, the same mistake can reach a customer before anyone looks. Stronger models widen what agents can do well, and they widen what they can do wrongly in the same stretch of work. The organizational question therefore moves from whether the agent works to who decides what it may touch, who is responsible for what it remembers, and who gets the call when it does something nobody expected.

The NIST AI Risk Management Framework organizes AI risk work around governing, mapping, measuring and managing. You do not need to adopt the framework in full to use its logic. Each of those activities needs a person behind it, and a governance activity without an owner tends to become a document that nobody opens. For a company that is building its first layer of agents, the useful starting point is the five layers a company should own, because the org chart should match that layered structure rather than a list of vendor products.

The five roles a company needs to name

Five roles are enough for most companies that run more than one agent. In a company with a single agent, one person may hold two or three of them, but the roles should still be written separately so that the person knows which decision they are making at any moment.

  • Executive sponsor: sets the business purpose, accepts the budget envelope and settles disputes between teams about what the agent should cover.
  • Agent owner: decides the agent's job, its success measure, its tone and its hours, and responds when the agent behaves badly.
  • Memory steward: decides what the agent may remember, where each fact comes from, how long it is kept and who can correct it.
  • Tool owner: maintains each integration, holds the credentials and confirms that each tool exposes only the actions the agent needs.
  • Approval owner: defines which actions need a human confirmation, sets the thresholds for money, personal data and outgoing messages, and reviews exceptions.

The executive sponsor is usually the person whose budget or customer relationship is at stake. The approval owner is often in finance, operations, legal or the function that carries the risk if something goes wrong. Choosing these people is the first real decision, and it is harder than it looks, because it asks managers to accept accountability they may have delegated to a developer or a vendor. The role of the sponsor is described in more detail in what an executive sponsor for an AI project must do.

How to assign each role to a named person

A role without a name is a suggestion. Write one role card per owner, one page each, and keep the language plain. The card should answer what the person decides alone, what they must escalate and what they can never decide without a second signature. Use the following sequence to make the assignment.

  • Write the agent's job in one sentence and name the function that sponsors it.
  • Name one person for each of the five roles on the card, using a personal name and not a team name.
  • Give each owner three decisions they can make alone, one they must escalate, and one they can never make.
  • Agree on the evidence each owner reads every month, such as the log of actions taken, the failed requests and the approvals granted or refused.
  • Set a review date for every role and list the events that trigger an earlier review, such as a new tool, a new data source or a complaint from a customer.
  • Name a backup for each role who already reads the same evidence, so a holiday or a resignation does not leave an agent without an owner.

A backup matters more than most leaders expect. Ownership that lives only in one person's head fails at the moment it is most needed. A company that wants to formalize the review body above these roles can use the steering committee charter as a starting structure.

A role card that cannot be read in two minutes is probably trying to do too much.

Who owns memory and tools, in practice

Memory means the information an agent can reach when it answers a question or carries out a task. Most business agents reach company knowledge through retrieval, where relevant documents are fetched at the moment of need. The AWS guidance comparing retrieval and fine-tuning explains why many teams prefer retrieval for knowledge that changes often: the source documents stay under the company's control and can be corrected without retraining a model. That design choice gives the memory steward concrete work. They own the list of sources, the rules for updating them, the rules for removing outdated material and the answer to the question of who is allowed to add something new.

Tools are the actions an agent can take. The Model Context Protocol gives agents a common way to connect to outside tools and data, but a protocol does not decide who may use which capability. The tool owner makes that decision by scoping each credential to the smallest set of actions the job requires. The OWASP authorization guidance states the principle plainly: deny by default and grant only what a specific function needs. Applied to an agent, that produces a short list of permitted actions per tool, reviewed like any other access list.

Illustrative scenario: a regional distributor gives its order-tracking agent read access to the warehouse system and a narrow write permission that only adds internal notes to an order. The tool owner keeps write access away from shipping and invoicing. Later, the warehouse team asks the agent to change delivery dates. The request goes to the approval owner, who decides that the change needs a human confirmation step, not a broader credential.

Who approves an action, and how often is ownership reviewed?

Approvals are the part of the chart that changes most as agents improve. Early on, a person confirms most external actions. Later, some low-risk actions can run without confirmation, but only after the approval owner has seen enough evidence to justify the change. Raise limits slowly and in one direction at a time. The approval owner should not be the same person as the agent owner, because the person who wants the agent to succeed is rarely the best person to decide how much freedom it should have. Guidance on setting these limits is covered in approval thresholds for stronger agents.

Ownership is only as good as the record behind it. The OpenTelemetry observability primer describes traces, metrics and logs as the means to understand what a system did and why. For an agent, owners need the same information in business terms: which tools it called, which memory it used, which actions required approval, and which were refused. Without that record, the owner is guessing. Review the chart each quarter against the record, and use a repeatable permission review to check that each credential still matches the job. A step-by-step method is set out in how to run a permission review for your AI agents.

How managed operations keep the chart current

In a managed engagement, Be Human helps the company draft the first version of the chart, writes the role cards with the named owners, and configures the agent's memory sources, tool scopes and approval thresholds to match what those owners have decided. Ongoing operation covers monitoring, recurring evaluations, integration maintenance, permission review and incident response. Changes are made through a recorded request and an approval step rather than by informal edits, so the owners can see what changed and why. The company keeps the decision rights. Its named people approve scope and thresholds, and Be Human carries out the technical work and reports the results.

If you want to identify the first task worth handing to an agent, and the owner who should sign off on it, the free Be Human AI diagnostic is a practical place to begin.

Frequently asked questions

Does a company need a dedicated AI team to own its agents? Not necessarily. A small company can assign the five roles to existing managers, provided each person has a written role card and a named backup. The chart matters more than the headcount, because the questions it answers do not disappear when the team is small.

Can the agent owner also be the approval owner? It is better to separate the two. The person who wants an agent to perform well tends to set thresholds more generously than an independent reviewer would, so an approval owner from finance, operations or risk keeps the limits honest.

How often should the org chart be reviewed? Review it at least once a quarter and whenever a tool, a data source or the purpose of an agent changes. Use the same evidence each time, such as the action log and the list of approvals, so that each review compares facts rather than impressions.

What happens if an owner leaves the company? Their role card should already name a backup who reads the same evidence. Transfer open decisions in writing, rotate any credentials the person held, and update the chart in the same week, so that the agent never runs without a named owner.

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 →