Yes. You do not need to document every business process before testing AI. You need one clearly defined task with recognisable inputs, an acceptable output, a named reviewer and a safe route for unusual cases.
If employees disagree about the correct result, source data conflicts or an error could cause harm before anyone catches it, clarify the task before automating it.
Undocumented does not mean unclear
A process can be consistent without being written down. An experienced employee may follow the same steps, use stable rules and produce a predictable result each time. The process exists, but much of it remains in that person’s head.
This creates operational risk, especially when the work depends on one employee. It does not automatically rule out a limited AI test. You may still be able to observe the task, define its boundaries and check whether similar inputs produce acceptable results.
An unclear process is different. Employees may use different data, apply conflicting rules or consider different outcomes correct. Writing down one person’s approach will not resolve those disagreements.
The practical question is not whether you have documented everything. It is whether you can define and evaluate one specific task. If your business cannot agree on what a good result looks like, it cannot reliably judge the AI’s output.
Check the task with five questions
Before introducing AI into recurring work, answer five questions.
Are the inputs recognisable?
You should know what starts the task and what information it uses. This could be an email, order form, meeting transcript or standard spreadsheet. If employees search several systems and decide case by case which value to trust, the inputs are not stable enough yet.Can you describe an acceptable output?
The result does not need to be identical every time, but its required content and format should be clear. A draft customer reply, for example, might need to answer the question, avoid making new commitments and identify missing information.Can a named person review it?
Someone must have the knowledge and authority to decide whether the result is usable. Assign this responsibility to a person or role, not vaguely to “the team”.Can you catch an error before it causes harm?
A flawed internal summary can usually be corrected during review. An incorrect price sent to a customer may create an immediate commercial problem. Review should happen before the output affects a customer, payment, contract, commitment or business record.Is there a clear route for exceptions?
The system should flag uncertain or unusual cases and send them to a specific person. An exception is manageable when it can be recognised. It becomes dangerous when it passes through as routine work.
A task that fails one question is not necessarily unsuitable for AI. The unclear answer shows what you need to fix before testing it.
Consider both task clarity and the impact of an error
Assess each proposed task on two dimensions: how clearly it is defined and how serious a wrong output could be.
| Task clarity | Impact of an error | Recommended approach |
|---|---|---|
| High | Low | Test on a limited set of real cases with human review. |
| Low | Low | Define the inputs, expected output and owner before testing. |
| High | High | Restrict permissions, retain records and require qualified approval. Keep AI in a supporting role where appropriate. |
| Low | High | Standardise the workflow before automating any consequential step. |
Apply the matrix to one recurring task, not an entire department. Clear rules alone do not make a high-impact task safe. One error can still affect prices, payments, contracts or customer commitments, so the controls should reflect the potential harm.
Start with assistance, not full automation
AI can support one task even when the wider process is not ready to run automatically from beginning to end.
Suitable first uses often prepare work for a person instead of making the final decision. AI might:
- draft a response,
- classify an incoming request,
- extract information from a document,
- compare text with a checklist,
- summarise a conversation.
A person then checks the output before using it. This keeps decisions such as approving a price or changing a delivery commitment with someone who has the necessary context and authority.
Human review reduces risk, but it does not guarantee correctness. Reviewers need enough time and information to challenge the output. If they routinely approve suggestions without checking them, the control exists only on paper.
Standardise the work first when people cannot agree on the result
AI is likely to spread inconsistency when the business has not agreed on how the work should be done. Warning signs include:
- Different employees consider different outcomes correct for the same case.
- Source systems contain conflicting values, with no agreed authoritative source.
- Nobody owns the final decision or the quality of the result.
- Exceptions are more common than the standard route.
- Employees rely on context that is missing from the stated inputs.
- An error cannot be detected before it reaches a customer or changes a commitment.
- Success is measured only by speed or volume, without checking correctness.
Standardise the relevant workflow first when the task passes between teams, changes data in several systems or affects prices, payments, contracts or customer commitments. The same applies when decisions must be explained later, sensitive information is involved or mistakes are difficult to reverse.
You do not need to map every activity in the company. Define the workflow far enough to identify who decides, which information takes priority, what evidence must be retained and where unusual cases go.
Test agreement with representative past cases
Give several representative past cases to the people who would review the AI’s output. Ask them to assess each case independently using the same proposed criteria.
If they regularly disagree, do not treat that as an AI problem. Settle the rule, identify the correct data source or assign decision authority first.
If they reach similar conclusions and can explain why, you have a stronger basis for a limited test. Use real cases, keep human review in place and record where the proposed rules fail or exceptions appear.
Frequently asked questions
Do we need a complete process map before testing AI?
No. You need a clear boundary around the task being tested. Define its trigger, inputs, acceptable output, reviewer and exception route.
Which AI tasks are usually easier to review?
Drafting, classification, information extraction, comparison and summarisation are often easier to inspect before use. Their suitability still depends on the source data, the reviewer’s expertise and the effect of a mistake.
Can human approval make a high-risk AI task safe?
Human approval is one control, not a guarantee. The reviewer needs the necessary context, time and authority. You should also retain enough information to show what the AI produced, what the reviewer checked and what was approved.
What should happen when AI is uncertain?
The case should be visibly flagged and routed to a named person or role. The system should not hide uncertainty by forcing every case through the standard route.