The short answer.
Defined rules are enough for calculations, required fields and unambiguous matching. AI can help when free text or documents with different layouts need interpretation. Its results can be wrong. A limited task, appropriate checks and human approval where needed therefore belong in the workflow.
How do you choose between rules, AI and a person?
Consider each step separately. A process is not entirely “AI” or “without AI”. A free-form message can first be interpreted, while the subsequent customer lookup and price calculation follow defined rules.
- 01Can this decision be described reliably with rules?YES →RULEUse a fixed rule.NO ↓
- 02Does unstructured information need interpretation?NO →HUMANClarify the process or decision first.YES → AI CANDIDATE ↓
- 03Can the result be validated independently?YES →AIUse AI and check the result against rules.NO →HUMANHuman review before any follow-up action.
Sensitive decisions or significant consequences of error require approval by a responsible person, regardless of the technical approach.
Which tasks do not need AI?
- Look up a customer ID in existing data.
- Check required details, date limits or permitted status changes.
- Calculate prices from an authoritative price list.
- Send a reminder for a pending approval after a defined period.
For these tasks, it is possible to specify which input should produce which result. These rules can be tested directly. A language model would introduce extra uncertainty without performing a necessary function.
Where can AI support processing?
Enquiries often describe similar needs in very different words. Documents contain the same details in different places. Here, AI can structure information or draft text from confirmed details.
The task remains narrow: choose known categories, capture supported details or prepare a draft. Missing information must not be filled in with plausible guesses. An unknown document type must be allowed to remain unknown.
| Task | Suitable approach | Checks |
|---|---|---|
| Calculate a price from quantity and a price list | Defined rule | Check the formula and price list used. |
| Capture requirements from a free-form email | Possible AI step | Compare details with the message; refer uncertainty for review. |
| Give binding approval to a proposal | Responsible person | Check the content and the specific version. |
How are AI results checked?
Correct formatting is not enough. A field can contain a valid number and still give the wrong amount. Checks therefore need to examine the content too: does the customer ID exist? Do totals add up? Is the detail supported by the original?
Not every statement can be checked automatically. Whether a free-form enquiry was understood correctly may require expert judgement. In that case, the responsible person needs to see the original information and open questions.
When is human approval still needed?
Approval is useful when consequences require conscious responsibility, details cannot be checked sufficiently or an exception falls outside the agreed rules. This might concern sending a proposal or changing important master data.
- The person sees the decisive details and uncertainties.
- They can approve, reject or request clarification.
- Approval applies to a specific version.
- A later change requires another review where appropriate.
How is an AI step limited?
Texts and files received as input are source material, not commands to the system. An instruction inside a document must not change the recipient, permissions or the next task.
I therefore separate reading the content, checking the details and performing the subsequent action. Timeouts, missing values and unknown content have their own handling path. Changes to the model or task instructions need to be checked again against the agreed cases.
The NIST framework provides an overarching approach to managing AI risks. The decision tree on this page is my own practical model.
What does this look like for an enquiry?
An email arrives. Rules check the input and look up the contact. AI can structure free-form requirements and draft a response. Prices and terms come from approved data. A person reviews the draft before it is sent.
If the enquiry instead arrives through a structured form with unambiguous product identifiers, the same workflow may need no AI. The actual task determines the technology.
Could your workflow benefit from automation?
Describe the task and the applications involved. I assess which requirements and open questions matter for implementation.
Review my process