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

Personalized AI infrastructure

Company Superintelligence for a Ten-Person Team: The Minimum Stack

A ten-person company needs five things before an AI agent is useful: scoped memory, narrow tools, explicit permissions, review habits and a success test.

Be Human typographic cover with the title: Company Superintelligence for a Ten-Person Team: The Minimum Stack
Conceptual illustration created with AI · Be Human

A ten-person company can run company superintelligence now with a deliberately small stack: one scoped memory, a short list of tools, written permissions, a recurring review habit and one measurable success test. Nothing in that list requires a large platform or an internal research team. It requires a clear decision about what the agent may know, what it may touch, and which moments need a person to approve.

Key takeaways

  • The minimum stack has five parts: scoped memory, a short tool list, written permissions, a review habit and one observable success test.
  • Start with one role and one memory scope. Widening knowledge before the role is clear adds risk without adding usefulness.
  • Give each tool a narrow contract that states what it reads, what it writes and which actions need a person to approve.
  • Review what the agent did on a fixed schedule, not only after something goes wrong.
  • Extend capability in steps, each backed by evidence from the last review, so the stack grows as models improve without being rebuilt.

What the minimum stack actually contains

The minimum stack has five parts. Memory is the set of documents, records and notes the agent may draw on, stored so you can inspect it and remove items. Tools are the systems the agent can call, such as a shared inbox, a ticket queue or a read-only view of order records. Permissions are a written list of what each agent identity may read, write, send or delete. The review habit is a recurring check of actions, errors and escalations. The success criterion is one measure that shows whether the agent is doing its job.

Notice what is absent: a custom model, a data lake, a standing AI team. Those can come later if a specific task needs them. For a ten-person company, the first version should fit on one page that an operations lead can explain aloud in a few minutes. If the drawing needs many boxes, the scope is already too wide. A simple architecture view helps here, and Draw Your AI Stack: A One-Page Architecture Map for Leaders shows how to produce one.

Start with one role and one bounded memory

Most small teams begin with a broad question: what should our AI know? A more useful question is which single job an agent should own, and what it needs to know to do that job. A role such as drafting first replies to shipping questions, sorting supplier invoices into a review queue, or preparing a weekly stock summary is narrow enough to test and concrete enough to matter.

Memory should match the role and nothing more. Include the policies, product sheets and approved reply templates the role needs. Leave out personal data, pricing exceptions and unfinished negotiations unless the owner of that information has decided they belong in scope. Label every item with its source and the date of its last review, so that when the agent relies on a policy you can check whether the policy is still current.

Illustrative scenario: a ten-person distributor where the office manager answers the same kind of shipping question many times each week. The first role is to draft replies to shipping status questions using the carrier policy, the delivery calendar and the reply templates. The memory excludes the full contract archive. When a draft is wrong, the team can trace the error to one document, which is the point of keeping the scope small.

Give the agent a short list of tools with clear contracts

A tool is any action the agent can take outside its own text output. Each tool needs a plain description, defined inputs, a predictable output and a stated limit. Vague tools produce vague behavior, and a tool that returns every field of a record invites the agent to use fields it does not need. Describe tools in terms of business actions, such as looking up order status, rather than in terms of the software behind them.

Use this as a working check before you connect anything: - Name each tool after the business action it supports, not the software it calls. - Start with read access. Grant write access only after the read path has been tested against real cases. - Return only the fields the task needs, and nothing more. - Log every call with a timestamp, the role that made it and the result. - Refuse by default when an input falls outside the documented range, and send that case to a person.

Any tool that reads outside text, such as customer emails or web pages, opens a path for instructions hidden inside that text. The OWASP guidance on prompt injection recommends treating retrieved content as data to be evaluated rather than as instructions to be followed, and keeping the agent's own permissions fixed regardless of what the text asks for.

Write permissions and approval boundaries

Permissions answer two questions: what this identity may touch, and which actions need a person to say yes. Give the agent its own identity rather than letting it borrow a staff login, so its actions can be separated from people's actions in the log. Grant the narrowest access that lets the role work, and record the grant in plain language beside the technical scope, so a manager can read it without opening an administration console.

The OWASP authorization guidance favors deny-by-default checks that the system enforces on every request. That means the boundary lives in the platform that connects the agent to your tools, not only in the instructions given to the model. Instructions shape behavior, but they are not a control. For a small team, three bands are usually enough: actions the agent may take freely, such as reading and drafting; actions it may take and report, such as sending a routine reply from an approved template; and actions that always require a person, such as refunds, price changes, contract language and deletions. The approval thresholds guide explains why these thresholds should start strict and move slowly.

Build review habits before you add capability

A review habit is a scheduled look at what the agent did, what it refused, and where a person corrected it. Without one, you learn about problems from customers, which is the most expensive way to learn. The NIST AI Risk Management Framework treats governance and measurement as ongoing functions rather than one-time approvals, and the same idea applies at small scale. The observability concepts described by OpenTelemetry translate directly: the action log is your record of what happened, staff corrections are your most useful signal, and the success criterion is the measure you watch over time.

Keep three records. The first is the action log, covering every tool call. The second is a correction log, where staff note each time they changed or rejected a draft and why. The third is an escalation review, listing cases the agent should have sent to a person but did not. Review weekly during the first month, then move to a monthly cadence once corrections taper off. Each quarter, run a formal permission check as well, following the steps in Quarterly AI Agent Permission Review: Steps and Evidence.

A four-step rollout for the first month

Go-live should follow a sequence in which each step produces evidence for the next. Skipping ahead is the most common way small teams end up with an agent that has broad access and no record of its decisions.

  • Write the role, its memory scope and its success criterion on one page, and have the person who owns that process sign it.
  • Connect read-only tools and run the agent in draft mode, comparing its drafts with what staff actually sent for the same cases.
  • Enable the narrowest write permission only after drafts meet the success criterion, and keep anything irreversible behind an approval step.
  • Schedule the first permission review and the first evaluation before go-live, so the review habit exists before the first incident.

When models improve, the stack does not need to be rebuilt. You revisit the same five parts: widen memory only where a documented gap appears, add tools only with a written contract, and raise approval thresholds only when the review record supports it. What the First Months With a Managed AI Agent Look Like describes how this cadence plays out in practice.

Be Human designs, implements and operates this kind of minimum stack for companies that want the agent's identity, memory, channels, tools and permissions kept under their control. A first engagement defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion. Managed operations can then cover monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, delivered remotely in English or Portuguese. The goal is to keep the stack small, visible and easy for your team to change.

If you are not yet sure which task deserves an agent first, the free Be Human AI diagnostic can help you identify it.

Frequently asked questions

Is a ten-person company too small to run company superintelligence? No, because the minimum stack is defined by scope, not by company size. A small team often benefits most, since one clear role and a single owner for permissions are easier to govern than a sprawling setup. The main requirement is discipline about what the agent may know and what it may do.

How many tools should the first agent use? Use only the tools that the first role cannot work without. A single read tool and one tightly scoped write tool is a common starting point for a small team. Add others only when the review record shows a specific gap that they would close.

Who should own the permission list? The owner of the business process should decide what the agent may do, and the person responsible for systems should enforce it in the platform. Splitting those roles prevents a situation where nobody can explain why an agent holds a given access. Review the list on a fixed schedule and whenever the role changes.

When should we extend the stack? Extend it when the current role meets its success criterion consistently and the review records point to a specific gap. Widening memory, adding tools or lowering approval thresholds should each follow a written request and a check of the evidence. Moving faster than the evidence usually creates the incidents that make teams distrust the agent.

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 →