
AI and Knowledge Succession in Family Businesses: Recording Rules
How to record rules, partner agreements and know-how in a family business so AI agents and successors do not depend on one person. Practical steps included.
Want to apply this in your company?
Book a call with Juan →In a family business, superintelligence in practice means AI agents and models that run parts of the operation under explicit rules, not a system that replaces the owner. For the operation not to depend on a single person, business rules, approval criteria, partner agreements and the know-how currently held in someone's head need to be written down, with an owner, a review date and clear limits. That record is what makes succession possible without interrupting customer support, purchasing and day-to-day decisions.
Key takeaways
- Dependence on one person usually shows up in exceptions: discounts outside the price list, negotiated deadlines, customers who will only talk to one partner.
- A rule that exists only in someone's memory cannot be delegated to an AI agent or to a new employee.
- The record should say who decides, how far each agent can act on its own, and when the conversation passes to a person.
- Every significant decision needs a short record: what was decided, by whom, with what information, and what changed afterward.
- The record is only useful if it is reviewed with the same discipline the company applies to its cash flow.
What depending on one person means in a family business
In most family businesses, operational knowledge builds up in a few people. One partner knows which supplier falls behind in the rainy season, another knows the deadline exception agreed with an old customer, and whoever runs the back office remembers why a certain product has a different return policy. None of this is wrong in itself. The problem appears when these rules exist only as someone's habit and were never written down.
When a company starts using AI in customer support, quotes or after-sales, that implicit information becomes a practical problem. The agent answers with the general rule, the customer asks for the exception they heard about before, and nobody knows whether the exception still applies. The person who knew is busy, travelling or leaving the partnership. The risk is not technological. It is the lack of a place where knowledge stays available to whoever has to decide.
It helps to separate two things that are often treated as one. The first is operational knowledge, such as deadlines, price tables and technical criteria. The second is the agreement between partners, such as who approves discounts above a certain level, who signs contracts and how disagreements are resolved. Both need to be recorded, but with different access levels. For the access and permissions side, the OWASP Authorization Cheat Sheet is a sound technical starting point.
Which rules and knowledge to record first
It does not make sense to try to document the whole company at once. The most useful order starts with what generates repeated decisions or financial risk. Typically, the initial list includes:
- Commercial exceptions: who can grant them, under what conditions and up to what limit.
- Deadlines promised to customers and suppliers, with the reason for each special deadline.
- Approval criteria for purchasing, spending and payments, including amounts that require two signatures.
- Return, warranty and repair policy, with the situations where the general rule does not apply.
- Customer communication rules, such as sending hours, tone and topics that must never be handled by automated message.
- Partner agreements on roles, substitutions and what each one can decide alone.
Each item should answer four questions: what the rule is, who decides, when the rule does not apply and where the evidence is. If the rule depends on personal customer data, the company also needs to check how that data is handled. Brazil's data protection authority, ANPD, is the official source for understanding LGPD obligations on this point. The topic of automated decisions has its own scope, which is worth reading before delegating any approval to a system, as we explain in our article on Article 20 of the LGPD.
How to set up the decision record, step by step
The record does not need sophisticated software to start. A spreadsheet or a shared document is enough, as long as it has an owner and a history. The sequence below works as an implementation checklist:
- Bring together the partners and the person who handles customers day to day, and list the situations that generate the most questions or exceptions in recent months.
- For each situation, write the general rule, the known exception and the name of who approves it.
- Mark which decisions an agent can make, which require a person and which should never be automated.
- Assign an owner to each record, responsible for keeping the text up to date.
- Create a simple history of relevant decisions: date, what was decided, who decided, the information used and the change in the rule, if any.
- Agree on a periodic review, for example whenever a supplier changes, prices change or someone leaves the operation.
- Test with real questions: an absent partner should be able to answer a common exception from the record, without calling anyone.
This last test is the most revealing. If the record cannot answer a real situation, it is not ready for succession. The same applies to whoever builds the AI side: the agent should only consult rules that are written and reviewed, not what someone remembers hearing.
A rule that exists only in one person's memory is not a company rule.
Illustrative scenario: the deadline exception nobody recorded
Illustrative scenario: a distributor of construction materials has a partner who has long negotiated different delivery deadlines for three large customers. Customer messages on WhatsApp are handled by an AI agent that checks a standard deadline table. When a customer asks for the same terms the other three receive, the agent gives the general deadline and passes the conversation to the partner. The partner is travelling, the conversation stalls and the customer goes to a competitor.
After a review, the company records the four exceptions, each with its reason, validity period and the person who can extend it. It also defines that the agent may state the standard deadline and confirm that a staff member will reply at an agreed time, but may not promise deadlines outside the table. None of these changes depends on new technology. It depends on writing down what was already being done and making the agent's limits explicit.
Where AI helps and where it should not decide
AI works well for consulting a record, organizing information and drafting a response. It is less suitable for deciding alone when the rule requires judgment about a relationship, a dispute between partners or an exception with no precedent. In these cases, the agent should log the request, gather the context and pass the decision to a person with authority.
There is also a technical caution. Rule texts stored in documents can be used as context, but that requires version control and access control. The NIST AI Risk Management Framework is useful for structuring this conversation inside the company, because it addresses governance, risk mapping and accountability, not only the model.
If the company does not yet have a policy for how employees use AI tools, the decision record should cover that too: which data can be pasted into external tools and who authorizes new integrations. That subject deserves its own treatment, and our company policy for employee ChatGPT use shows the kind of limit that tends to be useful.
How we apply the map, limit, pilot and transfer method
In our work with companies, the sequence usually starts by mapping the current process and the knowledge that supports each step. Then we set limits for the agent's data and actions, and pilot a real workflow with a decision record and a success metric agreed in advance. Only after that do we transfer the work to the team, with written rules, defined owners and a clear decision on whether to expand, adjust or stop.
This method does not remove the dependence on people, but it organizes how knowledge is transferred. At the end of the pilot, the company has not only a working agent but a set of rules that anyone taking over the operation can read. For companies that want to understand what is realistic to expect from this kind of system, the superintelligence glossary for executives helps separate technical terms from practical capability.
Frequently asked questions
How long does it take to record the rules of a family business? It depends on the size of the operation and how available the partners are, but the most efficient approach is to start with a short list of recurring situations and expand from there. The first version should be simple, with an owner, the general rule and the known exceptions. What matters most is that the first version gets used and reviewed, not that it covers everything at once.
Can an AI agent make decisions on behalf of a partner? In general, it should not make decisions that involve judgment, exceptions or disputes between partners. It can consult written rules, state standard conditions, log requests and pass the conversation to a person with authority. The more sensitive the decision, the clearer the boundary needs to be between what the agent does and what belongs to a human.
Do you need special software to maintain a decision record? Not necessarily. A spreadsheet or a shared document with change history can be enough at the start, as long as it has an owner and a periodic review. More structured tools make sense when the number of rules grows or when agents need to query those rules automatically under access control.
Does a rules and decisions record have LGPD implications? It can, especially when the record contains personal data about customers or when automated decisions affect people. The company should define which data goes into the record, who can access it and how long it is kept. For the legal side, it is worth consulting ANPD's official guidance and, if needed, a qualified professional.