Personalized AI infrastructure
Superintelligence for Small Business: Four Steps to Take This Month
A practical plan for small business owners: write the process, keep a memory file you own, map agent permissions and set a review rhythm.

For a small business owner, superintelligence is less a date on a calendar than a steady rise in what software can do without close supervision. The most useful response this month is not a new platform but four concrete habits: a written process list, a memory file the company owns, a permission map that states what an agent may do, and a review rhythm that checks all three. Each habit matters more as models improve, because a stronger agent carries out a vague process faster and reaches further into your systems.
Key takeaways
- Write down one recurring process step by step before you give any agent a task.
- Keep a memory file that your company owns, versions and reviews, independent of any single model vendor.
- Map permissions by action (read, draft, act), not by the name of a tool.
- Review permissions, outputs and incidents on a fixed rhythm, with an extra review after every change.
- Raise an agent's autonomy only after the review log shows the agent has earned the change.
What superintelligence changes for a small company
Most owners hear the word and ask whether they must prepare for one dramatic event. The more useful framing is operational. As models handle longer chains of work with less supervision, the limiting factor moves away from the model and toward the company itself: whether you can describe the work, state what the business knows, and say who approves what. A company that cannot answer those three questions will not gain much from a more capable model, and it may create risk faster than it creates value.
The encouraging part is that these answers are mostly documentation and decisions, which an owner can produce in the normal rhythm of running the business. You do not need to understand model internals to write a clear process, decide who signs off on a quote, or decide which records an assistant may change. Planning that avoids the common mistakes is covered in Superintelligence Myths That Derail Company AI Plans, and the four steps below are the practical starting point.
Step one: Write the process list
Pick one recurring process that touches customers, suppliers or money, and write it as a short sequence of plain steps. Do not start with the most ambitious workflow. Start with one that happens often enough to show results and rarely enough that a mistake can be caught before it spreads. For each step, record:
- The trigger that starts the step and the input it needs.
- The decision made at that step, and the person who makes it today.
- The output produced and where it is stored.
- The point where the process stops and waits for a human.
- The usual failure, and what the person does when it happens.
Consider this illustrative scenario: a small catering company receives quote requests by email. Written down, its process shows that the manager signs every quote regardless of size, which identifies the first place where an agent could safely draft instead of send. The same list reveals that the staffing estimate depends on a spreadsheet that only one employee maintains, a dependency nobody had stated. Writing the process exposes the gaps before any tool touches it, which is the real value of this step.
Step two: Keep a memory file you own
A memory file is the set of facts, rules and context an agent is allowed to rely on. It should be a plain document stored where your company controls access, with a named owner, a version number and a last-reviewed date. Keep in it the things that change slowly and matter often: what the company sells and to whom, the tone it uses with customers, the terms it never agrees to without approval, the people who make decisions, and links to the systems of record.
Keep out the things that change quickly or belong elsewhere. Live stock levels, current prices and individual customer histories should be read from their systems at the moment of use, not copied into a file that goes stale. A useful rule is this: if a fact would cause trouble when it is wrong, it needs an owner and a review date. Also avoid letting the memory live only inside a vendor's chat history, because that is hard to inspect, export or correct. The habits behind this file are described in Documentation Standards for a Superintelligence-Ready Company.
Step three: Draw a permission map
A permission map lists what each agent can do, in which systems, and what must happen before an action takes effect. Map by action rather than by tool name. Many problems start when an agent receives access to a whole system when it needed one capability inside that system. For every action, assign one of three levels:
- Read: the agent may look at the information and summarize it.
- Draft: the agent may prepare a message, record or change, which a named person reviews before it is sent or saved.
- Act: the agent may complete the step without review, and only for actions you have explicitly approved.
Next, write the approval boundaries in plain language. Any action that commits money, signs an agreement, sends a first message to a customer, deletes or exports records, or changes another agent's permissions should name a human approver. Set each threshold in words the team can apply, such as "any commitment the owner would need to sign," and lower the thresholds at first. Standard authorization guidance favors granting only the minimum access a task needs, and that principle is the right default for an agent.
Treat any text an agent reads from email, web pages or uploaded documents as untrusted input. A message can contain instructions aimed at the agent, so the permission map should never allow a high-impact action on the strength of text the agent read alone. A person or a deterministic rule should confirm it.
Step four: Set a review rhythm
Capabilities change without warning, and so do the tools and data an agent touches. A review rhythm turns that uncertainty into a schedule that someone owns. A workable pattern for a small company looks like this:
- Weekly: read the exceptions log, meaning every case where the agent stopped, failed or was overridden by a person, and record the cause.
- Monthly: sample a handful of outputs against the success criterion written for the role, and confirm that the memory file still matches how the business actually runs.
- Quarterly: review the permission map line by line, remove access nobody uses, and decide whether any approval threshold should move.
- After any change: rerun the same sample when a model, tool, data source or role description changes, before the change stays in production.
Evidence matters more than impressions. Keep a short log with the date, what was checked, what was found and who decided. Published guidance on evaluating agents, including OpenAI's evals guide, favors repeatable test cases over judging single outputs by eye. A useful test set is small, written by the people who know the work, and reused every time.
Here is an illustrative example. A quote-drafting agent passes its monthly sample, but the quarterly permission review finds that it can still send replies from a tool added later, when those replies were meant to be drafts. The owner removes that permission the same day and adds a line to the exceptions log. Nothing dramatic happened, and that is exactly what the rhythm is designed to produce.
How a managed setup handles these four steps
Be Human designs, implements and operates personalized AI infrastructure for companies, which means these four steps become the core work of a first engagement rather than homework for the owner. The engagement defines one useful role, the memory and tools that role 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, delivered remotely in English or Portuguese. The owner keeps decision rights over the memory file and the approval thresholds, and changes move through a documented request step described in How to Request and Approve Changes to a Managed AI Agent.
If you are not yet sure which process should go first, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Do I need to wait for superintelligence before acting? No. The four steps are useful with the models available today, and they make any later model easier to adopt safely. A clear process, a maintained memory file, a precise permission map and a review rhythm are the same work whether the agent is modest or highly capable.
Do I need a lawyer or a formal governance structure to start? Not for a first step. A written process, a short list of decision owners and a log of reviews are enough for a small company to begin. Legal review becomes more important when agents touch personal data, regulated decisions or contracts, and that is the point to involve an advisor before expanding scope.
Where should the memory file live? It should live in a place your company controls, in a format you can export, read and correct without help from a vendor. A shared document system with access control is usually enough at the start. What matters most is that one named person owns it and that it carries a review date.
How often should I change an agent's permissions? Change them when review evidence supports the change, not on a schedule driven by enthusiasm. Remove unused access during the quarterly review, and raise approval thresholds only after the exceptions log and monthly samples show stable behavior. Any new tool or data source should trigger a review before it goes live.
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.


