Use rules when the inputs and the next action are well defined. Add AI where the work involves interpreting varied text, finding a likely category, or drafting language. Many business workflows need both. An enquiry can use AI to propose the service category, fixed checks to identify missing information, and a person to approve the next action.
Choose the approach for each step of the job. That makes it easier to test the result and understand where errors come from. The example below follows a hypothetical service business deciding how to route customer emails.
Start with the decision the workflow needs to make
Write down the input, the expected output, and the facts needed to get from one to the other. Consider a form where customers choose a service from a list and supply a postcode. If an approved table maps each service and postcode to a team, a lookup can route the request. The team can explain the result by checking those two fields against the table.
An email saying “the cooling has stopped again at our north office” creates a different problem. Someone needs to interpret the description, identify the site, and check whether the request belongs to an existing job. AI can propose those interpretations. Keep the original message beside them so the intake owner can resolve ambiguity.
Anthropic’s engineering guidance recommends starting with the simplest solution that works and increasing complexity when needed. For this service business, a useful test is whether the fixed form already captures enough information to route most requests correctly.
Match the approach to the task
| Task | Starting approach | What to verify |
|---|---|---|
| Assign a form request using an approved service and postcode table | Rules and a lookup | Missing values and unmatched combinations reach an owner |
| Categorize an enquiry written in varied language | AI proposes a category, then routing rules apply | Ambiguous or unsupported categories go to manual review |
| Calculate a quote subtotal from approved quantities and rates | Conventional calculations | The quantity and rate source are correct and current |
| Draft commentary from checked reporting figures | AI prepares a draft for review | Each explanation is supported by the supplied figures or notes |
These choices can coexist in one system. A model may help identify which service a customer means, while the approved rate table determines the price. The quote preparation walkthrough shows how that boundary supports pricing review.
Keep permissions and business checks explicit
In the hypothetical intake workflow, AI proposes “maintenance” as the category and a likely customer record. The next steps check whether the category is supported, whether the customer match is unique, and whether the site address is present. If the address is missing, the request stays with the intake owner for clarification.
Record the proposed match as a proposal until the required checks pass. A valid field structure alone cannot establish that the information is correct. OpenAI’s structured output documentation explains that structured outputs can still contain mistakes.
Give the automation only the actions it needs for that stage. Preparing an intake draft and confirming an appointment are separate decisions. Decide who can confirm availability, change prices, or send a customer message, and enforce that decision in the workflow. A sentence in an incoming email should never change those permissions.
Test rules and AI against the same examples
Use permitted historical or synthetic requests with staff-prepared reference outcomes. Include ordinary requests, missing fields, conflicting details, and wording that falls outside the supported services. Run the simplest workable rules through the set first, then compare the proposed AI-assisted version.
For each version, record correct routes, incorrect routes, requests sent for review, and the time people spend fixing the output. Also check whether the reviewer has enough source information to resolve a case. A reduction in manual sorting only helps if correction and review work remain manageable.
If the rules handle the straightforward requests well, keep that route. Test AI on the unresolved text rather than sending every request through a model by default. Agree on what result would justify adding it, including acceptable error handling and ongoing support. Use the automation ROI guide to include review and correction time in that comparison.
Choose a small first release
For this example, the first release could route complete form submissions through the approved lookup and prepare email enquiries for staff review. The team can learn which email categories are dependable before giving the workflow more responsibility.
Document what happens when the lookup has no match, the model is unavailable, or two records disagree. Give those cases a named owner and keep a manual way to complete the job. Review the examples again when the service list or routing table changes.
The First AI Project Kit includes a playbook and editable worksheet for recording inputs, reviewers, and exception paths. Use it to mark which steps are fixed decisions and which require interpretation before asking someone to build the workflow.