EssayAI Adoption and Business Decisions
Enterprise AI Gets Stuck on What “Everyone Already Knows”
Policies and historical data preserve outcomes but often omit why an exception was made. Enterprise AI needs a living business judgment brief containing rules, cases, disagreements, and escalation boundaries.
Original Chinese version
On this page
Not long after the finance team’s AI system went live, sales came back with another complaint. The system had flagged a client-entertainment expense as an exception.
The reimbursement request included the purpose, invoice, amount, date, and approval. Everything was complete. The only problem was that the amount exceeded the normal limit. Based on the existing policy, AI had done nothing wrong. An over-limit expense should trigger a warning.
Sales had an explanation. The client had just placed a large order. The team had temporarily raised the level of hospitality, and a business leader had approved it. Why should the expense still be rejected?
Finance felt equally frustrated. Before the system went live, they had handed over the policy documents and historical data. They had provided everything the implementation team requested. Wasn’t the system supposed to learn from that material?
The original expectation was that AI would do more than flag unusual transactions. It would also look at how similar cases had been handled and offer an initial judgment on common exceptions.
Then the team reopened the historical expense records used to configure and test the system. The same type of over-limit client entertainment had sometimes been approved and sometimes rejected. The data preserved the outcome, but not why an exception had been made or what reasoning had supported the decision.
That missing layer was precisely what AI had failed to inherit.
If AI Has the Policies and the Data, Why Is It Still Missing the Point?
In a real business, many judgments depend on the circumstances at the time. A policy can state a general principle. A specific case still depends on who the client is, what stage the business is in, why an exception is needed, and who is willing to own the decision.
These judgments do not form a complete set of “hidden rules” waiting for a vendor to discover. Some are merely habits left over from the past. Some are interpreted differently across departments. Others are no longer appropriate today. AI can see what the company did before, but it cannot know which outcomes reflect genuine agreement, which reflect unresolved conflict, and which are simply old habits.
This also helps explain why AI can appear so capable when writing code. Coding rules are not universally simple, but feedback is often faster and clearer. The code compiles or it does not. A test passes or fails. A change can or cannot be rolled back. Many business decisions leave only an outcome. They do not preserve why that outcome made sense at the time.
The organization is not merely missing a few more rules. It first has to separate judgments that history has mixed together.
Policies and Data Are Not Enough for Enterprise AI
Giving AI thicker policy manuals will not solve the problem. However carefully you write them, the documents will eventually contain words such as “reasonable” and “special circumstances.” Are you prepared to let AI define what counts as reasonable on its own?
The more useful approach is to return to real events. Put finance, business, and technology in the same room and reopen past expense claims one by one. Why was this one approved? Why was that one rejected? Ask the people who actually made the decisions to explain what they considered at the time.
As the cases accumulate, some judgments will prove stable enough to become automatic checks. Others will resist hard rules but can still be organized into decision principles, positive and negative examples, and risk signals that help AI analyze the case and ask better questions.
Disagreements will surface as well. Sometimes the policy never addressed the issue. Sometimes two departments have been interpreting the same situation differently. That means the company either failed to make an existing consensus explicit or never had a shared standard in the first place. The relevant owners must decide which interpretation should become official.
Other judgments should not become standardized even after the context is clear. In those cases, define when the work must return to a person and who has the authority to decide. Before that point, AI can organize the information and identify the risk.
After working through the cases, the team can turn what used to happen through experience into a business judgment brief that can be discussed, tested, and revised. It contains more than rules. It includes decision principles, reference cases, unresolved disagreements, and the boundaries at which AI must escalate to a person.
At that point, the organization can also see which deployment problems will not disappear simply because the model becomes more capable. A large language model can find patterns in historical data. It cannot decide on the company’s behalf which patterns remain valid standards and which are outdated habits that should be discarded. The business must lead that decision, with the affected functions confirming the result together.
Someone Has to Own the Judgment End to End
Once this work begins, someone has to own it all the way through. One meeting is rarely enough. A new situation appears today; another exception appears tomorrow. The brief has to be extended and revised. Someone must lead the work and ensure that the decision standard always has an owner.
That person is usually someone with both the authority to interpret the process and the ability to move the organization. They do not need to configure the model or tune its parameters. The IT team or vendor can handle the technical implementation. Their responsibility is narrower and more important: bring the relevant functions together and make the business judgment explicit enough to be used.
With that ownership in place, the business is no longer handing over a policy document and a pile of data. It is handing over a living account of how the organization makes decisions—one that can continue to be tested and updated.
When AI enters a company, the hardest part is often not teaching it to read an expense claim. It is getting the organization to articulate the judgments it has always made through experience. The things people assume “everyone already knows,” but have never actually laid out, are the layer most likely to be skipped during implementation.
Management Advice
Do not expect IT or a vendor to make sense of the business on your behalf. Technical implementation can be outsourced. Deciding which judgments should become the official standard cannot. Someone on your side who understands the business has to make that call. Make the judgment explicit, and AI finally has something it can carry forward.
Do not stop after handing over a policy document and a process diagram. Bring real cases, disagreements, and exceptions so the technology team or vendor can see how the business actually decides. You are not handing over files. You are handing over a living system of judgment.
Next time, we will use management thinking to crack open another AI question.