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

Personalized AI infrastructure

Superintelligence for Small Companies: A Plain Start Under 20 People

A plain guide to superintelligence for companies under 20 people: which layer to build first, how to scope one agent role, and where approvals belong.

Be Human typographic cover with the title: Superintelligence for Small Companies: A Plain Start Under 20 People
Conceptual illustration created with AI · Be Human

Superintelligence for a company under 20 people means preparing the parts of the business that a more capable AI agent will depend on: clean records, a defined memory, explicit permissions and a named person who approves changes. You do not need a research budget or a data science team to begin. The first step is one useful role, the records it relies on, and a boundary it cannot cross without a human decision.

Key takeaways

  • The durable asset is the company layer: records, memory, permissions and approvals. It outlasts any single model or tool.
  • Start with one role and one observable success criterion, not a company-wide platform.
  • Write down what an agent may read, draft, write and approve before it touches a customer, an invoice or a schedule.
  • Keep a named human owner for approvals, and raise approval thresholds only when evidence supports it.
  • Small teams benefit most from being explicit. A clear one-page description beats a large, vague program.

What the superintelligence lens means at your size

The superintelligence lens asks one practical question: if AI agents become much more capable than they are today, what does your company need in place so they help without causing harm? It does not ask you to predict when that will happen or to agree with any particular forecast. You can hold the question calmly, because the preparation is largely the same whatever the timeline turns out to be.

For a company with fewer than 20 people, the scarce resource is attention, not headcount. Nobody has time to build an internal AI department, and you should not try. What a small company can do is decide early which knowledge, records and decisions belong to the company itself, and make sure any agent working for it reads from and writes to those places under rules that a person set and can change.

Why the company layer matters more than the model

Models change quickly, and a company that ties its operating knowledge to one vendor's chat history will have to start over whenever it switches. The parts that last are the company layer. It includes identity (what the agent is for and how it represents you), memory (what it keeps between conversations and where that memory lives), tools (which systems it may call), permissions (what it may read, draft, write or approve) and the source records that ground its answers.

The retrieval pattern shows why this matters. When an agent checks approved documents before answering, its output is tied to records you can inspect, instead of to whatever the model happens to recall. The same design makes errors easier to trace, because you can see which record was used.

Security follows the same logic. OWASP's guidance on prompt injection warns that text an agent reads from emails, web pages or support tickets can contain instructions that the agent may follow by mistake. Treat retrieved content as data, never as commands. OWASP's authorization guidance points in the same direction: check access on every request and grant the least privilege the job needs. The NIST AI Risk Management Framework organizes this work into govern, map, measure and manage, which gives a small company a shared vocabulary without requiring a large program.

Start with one role, not a platform

The first move is deliberately narrow. Choose one role, such as answering routine status questions from clients or preparing the first draft of supplier follow-ups. Use this guide to test a candidate:

  • The work repeats at least weekly and has a clear end state.
  • The records it needs already exist somewhere a person would look, even if they are messy.
  • A mistake is visible and reversible within a day.
  • One named person owns the outcome and will review what the agent does.
  • Success can be stated in one sentence that someone could check, such as a reply drafted the same business day with the order status confirmed in the system of record.

If a candidate fails two or more of these tests, it is a weak first role. Choose another one, or fix the records first. Fixing records is not wasted effort, since every later agent will need them too. For a fuller view of the minimum parts, see Company Superintelligence for a Ten-Person Team: The Minimum Stack.

A step-by-step plan for the company layer

Build the layer in this order. Each step produces a short document that a new team member could read in ten minutes.

  • Step 1: Write the role in one paragraph. Name the people it serves, the decisions it may make alone and the decisions it must hand back.
  • Step 2: Name one source of truth for each fact the agent needs. Price lists, customer records and policies should each have one owner and one location. Duplicate copies are where agents learn wrong things.
  • Step 3: Define memory. Decide what the agent may remember between conversations, who may edit those entries and when they expire.
  • Step 4: List tools and permissions. For each system, mark the action as read, draft, write or approve. Begin with read and draft only, and grant write access after the first evaluation passes. The AI agent permission review walks through this check step by step.
  • Step 5: Set approval boundaries. Any action that commits money, changes a customer commitment or sends an external message needs a named approver until evidence supports a higher threshold.
  • Step 6: Build a small evaluation set. Collect a dozen real past cases where the correct answer is known, and run the agent against them before every change. The OpenAI evals guide describes this habit of testing against defined cases rather than judging by impression.
  • Step 7: Keep logs. Record each tool call, each draft and each approval so that a person can reconstruct a bad outcome later.
  • Step 8: Schedule a review. Each quarter, confirm who has access, whether the success criterion still matters and whether the source records are still current.
A layer you can explain on one page is a layer your team can maintain.

Illustrative scenario: a six-person consulting office

This scenario is illustrative and does not describe a specific client. A six-person consulting office wants an agent to answer clients' questions about project status. The owner first writes a one-paragraph role and names the project tracker as the source of truth. The agent is allowed to draft replies, but not to send them. After two weeks of drafts, a project lead compares each draft with the answer she would have given. One draft states a deadline as firm because an old note called it tentative. The team changes the memory rule so unconfirmed dates are ignored, adds that case to the evaluation set and reruns the cases. Only then does the owner allow routine status updates to be sent automatically, while any message that mentions budget or scope still goes to a project lead for approval.

The lesson is not that the agent was careless. The company had not yet decided which facts count as confirmed, and the layer made that gap visible before it reached a client.

How Be Human approaches this in managed operations

Be Human designs and operates this layer for companies that prefer not to build it in-house. A first engagement defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion. Managed operations then cover monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, with each change requested and approved rather than made quietly. The work is remote and available in English and Portuguese. For a clear picture of how the first months unfold, see What the First Months With a Managed AI Agent Look Like.

If you want to find the first task worth handing to an agent, take the free Be Human AI diagnostic.

Frequently asked questions

Is superintelligence relevant to a company with fewer than 20 people? It is relevant in a practical sense, because the preparation that matters is the company layer: records, memory, permissions and approvals. A small team does not need a large program to benefit from being explicit about these things. The same work also makes everyday AI tools safer and more predictable today.

Do I need a developer or a data team to start? Not for the first step. You need one person who understands the work well enough to name the source of truth and the approver, and who will review the agent's output for a few weeks. Technical integration can be designed and operated by an outside team if you prefer not to build it internally.

What should I never give an agent on the first day? Avoid write access to financial systems, customer commitments and any channel where a mistake reaches outside the company without review. Start with read and draft permissions, then widen access only after the evaluation set passes and the logs show the agent behaving as described. Raise approval thresholds slowly and in writing.

How do I know whether the agent is working? Define the success criterion before launch, in one sentence that a person can check. Run a small set of past cases against the agent before each change, review the logs for actions taken, and record every correction a human makes. If corrections fall in a known area, fix the memory or the source record for that area rather than adding more instructions.

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 →