Decisions & strategy
When not to use AI in your business.
The strongest AI decision can be a smaller change, a standard tool or no project at all. Use the simplest approach that meets the outcome and can be operated responsibly.
The practical answer
Avoid an AI project when deterministic rules already solve the task, reliable evaluation is impossible, failures cannot be contained or the team cannot own the operation. If the problem is unclear, diagnose it before automating it.
- Simplify
Remove unnecessary work or collect a better input.
- Use rules
Apply a form, lookup, template or existing feature.
- Test AI
Use it only where interpretation adds useful value.
1. The task has a clear rule
If every accepted input has an exact answer defined by a stable rule, start with that rule. Required-field checks, date arithmetic, routing by an explicit service selection and looking up an existing customer ID do not need a language model.
For example, a form that asks for the project type and deadline may remove more qualification work than asking AI to infer them from an unstructured message. You can still use AI for a draft summary while keeping deterministic checks around the result.
2. You cannot tell whether the output is good
A convincing answer is not an acceptance criterion. If the team cannot describe correct output, assemble representative cases or identify unacceptable errors, an impressive demo does not resolve the uncertainty. Start by defining the review criteria with the process owner.
Open-ended exploration may still be useful in a sandbox. Treat it as exploration with a person in control, not as evidence that an unattended production workflow is ready. Keep experimental output separate from authoritative records.
3. The failure cost exceeds the review capacity
Consider the action after the answer. A wrong draft that an informed person checks is different from an incorrect message sent to a customer, an unauthorised record update or a decision with serious consequences. Choose the level of automation from the consequence of being wrong.
“Human in the loop” is not enough if the reviewer has no time, context or authority to intervene. If meaningful review and recovery cannot be arranged, reduce the action scope, use deterministic controls or do not automate that decision.
4. The economics only work in the demo
Low task volume, frequent exceptions or constant prompt repair can make automation less attractive than a simpler process. Count setup, licences, reviewer time, monitoring and maintenance. Compare that cost with a measured baseline, not with the best successful example.
A process that changes every week may first need an owner and stable rules. Automating a moving target can turn each operational change into a technical change. It may be worth waiting until the basic service is clearer.
5. Write down the conditions for stopping
Before a pilot, agree what would make you stop, simplify or return to manual work. Those conditions help a team make a decision even when it has invested time in a prototype. Use observed evidence and business consequences rather than a generic maturity score.
- The team cannot obtain authorised, representative examples.
- Review effort removes the expected operational benefit.
- Unacceptable failures remain after the agreed evaluation cycle.
- No person owns monitoring, exceptions and recovery after handover.
- An existing tool or a simpler rule meets the outcome with less operating complexity.
| Condition | Evidence to record | Next step |
|---|---|---|
| A stable rule solves the task | Inputs and rule exceptions | Configure the rule first |
| No reliable way to evaluate output | Missing expected actions or reviewer | Define acceptance before building |
| Review costs exceed useful capacity | Corrections and reviewer effort | Simplify or stop the pilot |
Make a stop decision reviewable
Write the observation, the decision owner and the simpler alternative beside each stopping condition. Pause live automation when a condition is met; investigate with controlled examples before deciding whether to resume.
AI stop-decision worksheet (CSV)
Download the table shown above. Replace illustrative records with authorised examples from your own workflow; this is a working template, not an assessment of your business.
AI stop-decision worksheet (CSV)References and further reading
Entrepreneur and AI consultant, I help service businesses improve their operations. I turn software choices into practical plans. My team at Klicky builds what we decide.
Khalid Boudrari