Opinion
Companies run on rules nobody ever wrote down
Revenue, churn and customer health often mean different things across the same company. As AI enters decision-making, those unwritten definitions can no longer remain informal.
Ask the CFO, the head of sales and the customer success leader to define an active customer. In many companies, you will get three different answers.
Finance may count only paying accounts. Sales may include customers who signed but have not yet gone live. Customer success may exclude accounts with no recent usage. Each definition can be reasonable within its own workflow. The trouble begins when the same company uses all three without a shared rule.
Companies accumulate thousands of definitions like these. What counts as booked revenue? When is a deal truly closed? Which type of churn matters? What makes a forecast conservative? Which exception overrides the standard process?
That ambiguity is not limited to data. In my experience, many people struggle to define exactly what they want from a system, and even fewer can describe the desired outcome precisely enough for someone else to reproduce it. AI makes that lack of clarity visible. Once that logic is automated, however, an unclear rule can stop being a local disagreement and start shaping decisions across the company.
Formal systems capture transactions, fields and approvals. People add the meaning behind the data. For example, a manager may reject a forecast and explain why, while an analyst may determine which of several revenue fields is authoritative. So when a new employee joins the company, she learns that a customer can be contractually active and operationally dormant at the same time.
This accumulated judgment is often called tribal knowledge. Together, the definitions, exceptions and priorities form the company’s invisible operating system. Today, employees mostly learn it through meetings, corrections, spreadsheet comments and repeated exposure to edge cases. The workflow may be documented, while the rules that decide the outcome remain in people’s memory.
I often compare a general-purpose AI system to a brilliant PhD graduate arriving on their first day. Intelligence helps, but usefulness arrives after onboarding. The organization still has to teach the system how it defines its goals, priorities and terms, which data sources matter, what exceptions exist and who is allowed to see what.
Connecting AI to Salesforce, NetSuite, a data warehouse or a collection of spreadsheets gives it access to records. That access is only the starting point. The next task is reasoning: deciding how fields relate, which definitions apply and whether the result makes business sense. A system may find several fields labeled “revenue.” People who know the business must explain which field applies to existing customers, which one belongs in a board report and which one should never be used for forecasting.
That is where the hidden layer becomes consequential. Companies now expect AI to analyze, recommend, monitor and act across core workflows. An experienced employee can sense that an answer looks wrong, ask a clarifying question and search for missing context. An automated system acts on the logic available to it. The result may be a consistent answer that is consistently wrong.
Self-service tools raise the stakes further. A finance analyst can build a forecasting script in minutes. A marketing team can automate a workflow. An operations team can create its own dashboard. The speed is valuable, but local definitions can quickly harden into local software. One team calculates revenue one way, another uses a different formula for customer acquisition cost, and several dashboards present incompatible versions of the business.
This is why business teams are becoming central to AI adoption. Technical teams build and maintain the infrastructure, data pipelines, security and access controls. Finance, sales, customer success and operations understand the outcomes, exceptions and trade-offs that define useful results. The strongest implementations will bring those two forms of knowledge together, with clear ownership for the business context the system uses.
Companies should treat business logic as a managed asset. The scope should remain focused: identify the definitions and exceptions that materially change decisions, assign owners, capture real examples and keep the rules current as the business evolves. For example, a correction from a knowledgeable user should become organizational knowledge and remain available beyond a single meeting or spreadsheet comment.
The risk of leaving tribal knowledge undocumented becomes most obvious when an experienced employee leaves: the systems may retain every record while losing the reasoning that made those records useful. Making business logic transferable reduces that dependency, helps new hires become productive faster and gives teams a shared basis for explaining decisions.
AI will make this discipline more urgent because systems cannot reliably operate on knowledge that the organization has never made transferable. The next phase will be judged by whether companies can use AI repeatedly inside real decisions with consistent definitions, context and accountability. The companies that benefit most will be those that understand themselves well enough to teach others, human or machine, how they work and what outcomes they expect.
The first step is simple to state and difficult to execute: write down the rules that already run the company.
Dr. Daniel Zahavi is co-founder and CEO of Arito. He previously co-founded LEVL Technologies, which was acquired by Comcast in 2022.
