By John Wheeler | Wheeler Intelligent Systems
The procedure says to release the order. Your most experienced employee says to wait.
Nothing looks wrong on the screen. The customer account is current, the inventory is available, and the delivery date is within the normal schedule. But the employee remembers that this customer’s receiving crew is unavailable on certain afternoons. Sending the order now could mean a failed delivery and a second trip.
That detail might be in an old email. It might be in a note nobody checks. Or it might exist only in the employee’s memory.
Now imagine giving an AI assistant the written procedure and asking it to help process orders. It could follow every step correctly and still miss the fact that changes the decision.
This fictional example points to a practical challenge for businesses putting AI to work. Your best employees know more than the process documents reveal. Before you ask AI to help with their work, spend some time learning what they notice, what they question, and when they choose a different path.
The written process is a starting point
A procedure describes how work should move under the conditions its author considered. Real work brings missing information, conflicting requests, unusual customers, and changes that have not made it into the instructions.
An experienced employee bridges those gaps. A buyer knows which supplier dates need confirmation. A service manager knows which short customer descriptions usually require more questions. An accounting employee knows that a familiar invoice mismatch may have a reasonable explanation, while another deserves attention right away.
Their value includes knowing when to slow down and investigate. A checklist may say what to do next. Experience helps someone decide whether doing it next makes sense.
AI can help organize this knowledge and prepare useful drafts. But it needs access to relevant, current information through an approved workflow. If a fact has never been recorded or supplied, the assistant has no sound basis for treating it as a company rule.
Start with one task and a real example
“Tell me everything you know about your job” is a difficult request. It is too broad, and it puts pressure on someone to remember details outside the setting where they usually use them.
A better opening is: “Let’s walk through a recent case that took more thought than usual.”
Choose one repeated task, such as reviewing a quote, scheduling a service visit, checking an invoice, or releasing an order. Ask the employee to explain a routine case first. Then compare it with an unusual one.
Use records the business is permitted to use, and limit access to the people who need them. For an outside demonstration or experiment, use fictional or properly de-identified examples.
The aim is to understand the decisions inside the work. These five questions can help.
A. What do you check that the procedure leaves out?
This question often brings out the quiet checks people make without thinking of them as separate steps.
In our fictional delivery example, the employee might check the customer’s receiving hours, whether an appointment is required, and whether someone has confirmed access to the unloading area. Those checks explain why a technically ready order may still need to wait.
Ask a follow-up: “What would a new person probably miss here?” That can make the answer more specific.
Write down both the check and its purpose. “Check receiving hours” is useful. “Confirm receiving hours before dispatch because this location sometimes cannot accept afternoon deliveries” gives a future reader more context.
Then establish whether that information is current. A remembered restriction could have changed. Experience is valuable evidence, but it should still be checked before becoming a standing instruction.
B. What makes you stop and ask someone?
Good employees know where their authority ends. They also know when the information is too weak to support a decision.
A quote preparer might stop when the requested work is unclear. A buyer might ask for approval when a substitute material would affect the customer’s requirements. A scheduler might pause when a job requires skills that are not available on the assigned crew.
Capture the trigger, the person who can resolve it, and the information that person needs. Otherwise, a rule such as “ask the manager” can turn into another round of messages because the manager receives an incomplete question.
This is especially useful when designing an AI assistant. You can tell it to identify the missing requirement and prepare a clear handoff. The responsible person can then make the decision with the supporting facts in view.
C. Which exceptions happen often enough to plan for?
Some exceptions are rare. Others happen every week and have simply never been added to the process.
Ask the employee which situations repeatedly create extra work. Perhaps customers send incomplete specifications. Perhaps supplier confirmations arrive after the planning deadline. Perhaps an invoice uses a different unit of measure from the purchase order.
For each recurring case, record how it is recognized, how it is normally resolved, and what would make the usual response unsuitable.
Keep the scope manageable. You do not need a document covering every possible event before learning anything useful. Start with the few cases that matter most, then add new ones as the team encounters them.
D. What looks correct but usually deserves another look?
This question helps uncover warning signs that are easy to overlook.
A total may be mathematically correct while using the wrong quantity. A delivery date may fit the calendar while ignoring the customer’s access rules. A quote may include the requested work while leaving out preparation or cleanup.
Ask for examples and the reason behind the concern. “That customer is difficult” is too vague to be a useful instruction. “This location requires written approval for changes to the delivery window” describes something a person can check.
Be careful to separate a fact from a hunch. If an employee has a concern but cannot yet explain its basis, record it as a question to investigate. Do not quietly turn it into a rule that affects every future case.
E. How do you know the work is actually finished?
Different people can mean different things when they say a task is complete.
For a quote, does finished mean the draft exists, the numbers have been checked, or the customer has received an approved version? For an order, does it mean the goods have shipped or the customer has accepted delivery?
Agree on a practical definition for the task you are documenting. Include the checks required, the record that shows completion, and the person who owns the next step.
This prevents an AI assistant from reporting success merely because it produced an output. It also helps employees see whether a handoff is ready for use.
Turn the answers into something people can maintain
After the conversation, prepare a short working guide. AI can help turn approved notes into a draft, but the employee and process owner should review it for meaning and accuracy.
| Part of the guide | What to capture |
|---|---|
| Task and purpose | What the work accomplishes |
| Required information | Facts needed before proceeding |
| Routine steps | The normal path through the task |
| Important checks | Details that change the decision |
| Common exceptions | How to recognize and handle them |
| Human decisions | Who must approve or resolve a question |
| Completion | What shows the work is ready for use |
| Ownership and review | Who keeps the guide current |
Keep the first version short enough to use during the work. Put detailed examples alongside it rather than burying the main instructions in a long report.
Also distinguish an approved practice from a personal workaround. An employee may have found a way to cope with a broken process that management should repair. Documenting it should open that discussion, not automatically make the workaround permanent.
Ask AI to organize the knowledge without filling the gaps
Here is a practical prompt readers can adapt after gathering notes they are authorized to use:
Help me turn these notes into a working guide for one business task. Use only the facts provided. Separate routine steps, important checks, common exceptions, and decisions requiring a person. Explain what counts as finished. Mark missing or conflicting information as an open question. Do not invent company rules, approvals, customer requirements, or details that are absent. Include a short list of questions for the employee and process owner to review.
The prompt gives the assistant a useful drafting role. It does not establish that the resulting guide is correct. Have the people who understand the work check the draft before anyone relies on it.
Test the guide with someone learning the task
One practical test is to give a newer employee a fictional case and the draft guide. Ask them to explain what they would do, what they would check, and where they would seek help.
Use several cases: a routine one, a common exception, a missing-information case, and a case with conflicting records. Let the experienced employee review the responses.
Where does the learner get stuck? Which wording leads to the wrong choice? What important check remains unstated? Those findings tell you where to improve the guide.
If you also test an AI assistant, use the same cases and look for the same gaps. A smooth explanation is not enough. Check whether it recognizes uncertainty, refers to the right evidence, and respects the approval boundaries.
Make sharing knowledge worth the employee’s time
Give the employee time to contribute and credit for improving the process. Explain how the guide will help with training, coverage, and fewer repeated questions.
Avoid framing the exercise as extracting everything someone knows so the business no longer needs them. Their judgment is part of what makes the process work, and reviewing the guide is itself useful work.
At Wheeler Intelligent Systems, I see this as a practical part of putting AI to work in established businesses. Learn how the work succeeds, record the important decisions, and test whether the assistant helps people use that knowledge well.
You can begin with one task and one experienced employee. The first useful result may be a better process guide, even before AI becomes part of daily operations.
What does your most experienced employee check that nobody has ever put into the procedure?