Business perspective

When AI Meets the Unwritten Rules of Your Business

Ask an experienced employee how a process works, and you will usually get two answers.

First, they explain the procedure. Then come the qualifications.

“Unless the customer was promised something different.”

“If the delivery was late, we check with operations.”

“For that kind of request, speak to Nadia before doing anything.”

Those qualifications can carry a surprising amount of business knowledge. They reflect what people have learned about customers, commitments and situations that the written procedure does not fully cover.

When a company introduces AI into that process, those small qualifications become important decisions.

The system may be able to read the policy, find the customer record and prepare a convincing response. But what should it do when the policy and the circumstances point in different directions?

Someone has to decide when the assistant can make an exception and when it needs approval.

Information is not authority

Imagine a customer asking for a refund.

The policy allows refunds within 30 days. The request arrives on day 35, so the straightforward answer appears to be no.

An experienced employee looks further. The delivery was delayed. The customer contacted support twice before receiving the order. Someone on the account team may have promised flexibility.

The employee knows to pause and check.

Now imagine an AI assistant handling the same request. It has the refund policy and the order date. It gives a clear, polite refusal.

The response is consistent with the information it was given. It may still be the wrong outcome for the business.

Giving it more information helps, but creates another decision. If it can see the delivery history and the account manager’s note, is it allowed to make an exception? Does that note constitute approval? What if the promise is vague or contradicts a later instruction?

The company has to decide how those facts should affect the outcome.

This is one of the less obvious things AI adoption can reveal: how much of a process depends on people interpreting information that has never been brought together clearly.

Experienced employees fill the gaps. They know where to look, whom to ask and which situations need a manager. An AI system needs a deliberate way to recognise those gaps and bring them to the right person.

There is a difference between having enough information to form an answer and having the authority to decide.

In the refund example, the assistant might identify the late delivery, summarise the customer’s previous messages and flag the account manager’s promise. That could save an employee considerable time.

Issuing the refund is a separate responsibility.

A company may let the assistant handle standard cases while referring disputed commitments or larger amounts to a manager. Another company may choose different boundaries. What matters is that someone has made the decision, and the system’s access and capabilities reflect it.

Otherwise, the technology can end up settling a management question by default.

For example, connecting the assistant to the refund system may give it the ability to issue refunds before anyone has agreed which cases it should handle independently. A broad instruction such as “resolve customer issues” can leave similar room for interpretation.

The same issue appears when an assistant prepares a quotation or processes a purchase request. Finding the price or the right laptop is useful work. Approving a discount or deciding that a purchase is justified requires additional judgment.

Start with the exceptions you already have

Making room for judgment does not mean documenting every possible exception. Trying to anticipate everything can make a process slow and brittle. Businesses need room to respond to circumstances.

A practical starting point is to examine the cases that already require a conversation.

Take a small set of recent requests that employees escalated, corrected or handled differently from the standard procedure. Ask the people involved what made each case unusual.

Often, the answer will reveal something specific: missing information, conflicting records, an earlier promise, an approval limit or a decision that belongs to another team.

Those examples tell you more about the work than a demonstration built entirely around straightforward requests.

They also deserve scrutiny. An exception might reflect sound judgment, an outdated rule or a habit that nobody has questioned. Looking at actual cases helps distinguish judgment worth preserving from practices the business should reconsider.

That gives leaders a useful opportunity before automation expands: agree which decisions should be consistent, where discretion is valuable and who should exercise it.

The assistant can then support those choices rather than reproduce every workaround it encounters.

Make the handover useful

“Ask a human” is incomplete unless the system knows whom to ask, what to show them and what it may do while waiting.

A manager reviewing the refund should be able to see the relevant dates, the delivery problem, the earlier promise and the reason the assistant paused.

Without that context, the manager must investigate the case again or approve a recommendation without enough information.

The quality of the handover affects how much time the automation actually saves. A system that processes routine requests quickly but sends confusing cases to the wrong team can simply move work elsewhere.

A better test of readiness

Working through these questions can also expose commitments that are hard to find, rules that employees interpret differently and decisions with no clear owner. Resolving those gaps makes the process easier for everyone, including the next employee who joins the team.

For a leader, this offers a useful way to judge readiness. A successful demonstration shows that the AI can handle the examples presented to it. Looking at exceptions shows how the company expects it to behave when the work becomes less straightforward.

Both matter once the system is connected to customer records, operational tools or actions with real consequences.

Before giving AI more responsibility, sit with someone who does the work today and ask:

“When do you stop following the usual process, and how do you know what to do next?”

Their answer may reveal the part of the business your AI most needs to understand.