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

Personalized AI infrastructure

Can a 10-Person Company Use Superintelligence Today? A Direct Answer

A 10-person company can use today's capable AI models, but not a finished superintelligence. Here is what to use now, what to avoid and the first steps.

Be Human typographic cover with the title: Can a 10-Person Company Use Superintelligence Today? A Direct Answer
Conceptual illustration created with AI · Be Human

A 10-person company can use superintelligence today only in the practical sense that it can use the most capable AI models available now, connected to its own approved records, tools and permissions. It cannot buy a finished superintelligent colleague, and any vendor who promises one deserves close scrutiny. The sensible answer is to build a small, governed layer first, so that stronger models can later improve the work without forcing a rebuild of the company's setup.

Key takeaways

  • Today a small team can get real value from capable models applied to one narrow, repeated role.
  • The value depends on the layer around the model: memory, tools, permissions, approval rules and checks.
  • No one can sell you a finished system that runs the company unsupervised, so be wary of offers that imply one.
  • Start with one role, one set of current records and one success criterion you can observe.
  • Design the layer so a stronger model changes the engine while your records, permissions and tests stay in place.

What a 10-person company can use today

For a team this size, the useful question is not which model is the most powerful but which repeated task is worth handing to an agent. Capable models already do well at drafting, summarizing long documents, classifying incoming requests, answering questions from approved material and preparing work for a person to review. The output is only as good as the material the agent can see, so the most common first pattern is retrieval: the agent looks up approved documents before it answers and can show which document it used.

A good first role has three traits. The work repeats every week, the source material already exists in a reasonably current form, and a mistake is visible and reversible before it reaches a customer or a payment. A quote drafter that a salesperson reviews before sending fits that profile. An agent that issues refunds on its own does not, at least not at the start.

If you are still deciding how much structure a company this size needs, the minimum stack for a ten-person team is a useful reference point before you choose tools.

What a small company should not expect

Three expectations cause most of the trouble. The first is autonomy: a model that reasons well still needs an owner who decides what it may do and when it must stop. The second is accuracy without records. If the warranty policy is outdated or two documents contradict each other, the agent will repeat the contradiction with confidence. The third is safety by default. Agents that read emails, web pages or uploaded files can be steered by instructions hidden in that content. The OWASP Top 10 for LLM Applications treats this class of risk as a design problem, not a wording problem, so the answer lies in limiting what the agent can reach and do.

The NIST AI Risk Management Framework is a practical reference for deciding these controls before launch, especially the questions of who is accountable for each output and how the system is monitored after it goes live.

The layer that gains from better models

The part of the system that lasts is not the model. It is the set of decisions around it: which memory the agent may read and which version of each document counts as current, which tools it can call and with what scope, what it may only draft, what it may send or change, when it must hand off to a person, and which test questions prove it is still working. Treat the model as an engine you can replace. The memory, permissions, approval rules and logs are the chassis and controls, and they belong to the company.

When a stronger model appears, a layer built this way turns the upgrade into a controlled test: run the same evaluation set on the new model, compare the results with the current one, and change approval thresholds only where the evidence supports it. A layer built around one vendor's prompts and undocumented permissions turns the same upgrade into a rebuild. For a fuller picture of how the parts fit together, see how to draw your AI stack on one page.

Illustrative scenario: a parts distributor's supplier questions

Illustrative scenario (not a client case): a ten-person distributor of replacement parts receives the same questions every week about lead times, warranty terms and compatible models. The owner gives one operations agent read access to the current catalog and the approved warranty policy, and asks it to draft replies in the shared inbox. The agent may not send messages, change prices or promise refunds. Any request about refunds or exceptions goes to the operations lead, with the relevant policy paragraph attached.

After the first two weeks, the team reviews the log. Several drafts quoted a warranty period from an old folder that nobody had cleaned up, so the owner archives the outdated policy and records who approved the replacement. Drafts are then accepted with lighter edits, and the operations lead keeps the refund boundary exactly where it was. The example shows the order of work: clean the source, set the boundary, observe the results, then widen the scope only when the logs justify it.

First steps in order

Work through these steps before you buy more tools or ask for a broader rollout:

  • Name one role, one accountable owner and one repeated task that happens often enough to measure.
  • List the records the role needs and mark which of them are current, approved and easy to find.
  • Write stop rules that state what the agent may read, draft, send and change, and what it must never touch.
  • Define one success criterion you can observe in the logs, such as the number of drafts accepted without edits in a week.
  • Build a small test set from real questions whose correct answers a person has already checked.
  • Give the agent the fewest tool permissions that still let it finish the task, and record every tool call.
  • Schedule the first permission review before launch, not after the first incident.

Be Human designs, implements and operates this kind of layer for companies that want AI agents to keep their identity, memory, channels, tools, permissions and operating context under human control. A first engagement defines one useful role, the memory and tools it needs, its approval boundaries and an observable success criterion. Managed operations can then include monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements. The service is remote and available in English and Portuguese.

Measuring results and handling model changes

Evaluation is what turns a promising demo into an operating tool. OpenAI's guide to evals describes the same discipline used across the industry: write test cases with expected outcomes, score the results, and rerun them after every change to the model, the documents or the tools. A small test set that a person maintains is more useful than a large one nobody reads.

Review permissions on a fixed schedule rather than only when something breaks. A quarterly permission review keeps tool access aligned with the role as the business changes, and it creates the evidence you will need when you decide whether to raise an approval threshold. If you want to find the first task worth handing to an agent, take the free Be Human AI diagnostic.

Frequently asked questions

Can a company with ten people use superintelligence today? Yes, in the practical sense that it can use the most capable AI models now through a governed layer of approved records, tools and permissions. It cannot buy a finished superintelligent system that runs operations without supervision, and claims of that kind should be treated with caution. The realistic goal is one well-defined role that saves review time on repeated work and can be checked.

Is it safe to let an AI agent read customer data? It can be, but only with written rules about what the agent may read, why it needs that access and where its outputs may go. Limit access to the data the task requires, log every read and action, and review those permissions on a schedule. Treat content the agent reads, such as incoming emails, as untrusted input that may contain instructions.

Should we wait for better models before we start? No, because the groundwork is the same regardless of model strength: current records, a named owner, stop rules and a test set. Waiting leaves those gaps open, and the gaps are what make any later upgrade expensive. Start with one role and keep the model replaceable.

What should we prepare first? Pick one repeated task, gather the current records it depends on, and write down what the agent may draft, send and never touch. Then define one success criterion and a small set of test questions with checked answers. A person who knows the work should choose the task and review the first results before the scope grows.

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 →