Practical AI / GUIDE + WORKSHEET
AI, rules, or a human: choose the right workflow
Decide where deterministic rules, AI assistance, and human judgment belong in a business workflow using a practical decision matrix.
THE STARTING POINT
Use ordinary rules when the decision is explicit and the inputs are structured. Consider AI assistance when interpreting variable language or documents adds value. Keep human judgment where context, uncertainty, or consequences require it. A useful workflow can combine all three without giving AI permission to make every final decision.
Separate interpretation from action
An email may need interpretation to suggest its topic, while assigning it to a named queue follows an ordinary rule. Sending a binding response may require approval. Split those steps so each uses an appropriate method. This makes the system easier to test and avoids treating a plausible model answer as authorization to act.
Compare with the simplest baseline
Try a clear form, a lookup, or a small set of rules before assuming a model is necessary. Measure whether AI improves the cases that matter enough to justify evaluation, review, and maintenance. Include an unknown or needs-review route rather than forcing a prediction for every input. The decision should reflect the work, not enthusiasm for a particular technology.
Design the review boundary
Identify actions with external consequences and decide what evidence a reviewer needs. A reviewer should see the original input, proposed interpretation, and resulting action. Establish how rejected suggestions improve the process and when the workflow should fall back to manual handling. Test unusual inputs and meaningful mistakes, not just demonstrations that look convincing.
WORKED EXAMPLE / ILLUSTRATIVE
A fictional customer inbox
AI suggests a topic for a free-form message. A rule selects the corresponding review queue. A person decides whether a proposed customer reply is appropriate. The system assists interpretation without independently committing the business to a response.
| Task | Candidate method | Reason |
|---|---|---|
| Check whether required ID exists | Rule | Explicit condition |
| Suggest topic of free-form message | AI with evaluation | Variable language |
| Approve a disputed refund response | Human under business policy | Context and consequences |
MAKE IT USEFUL
Rules, AI, or human decision
Evaluate each step separately.
Your notes stay in this page and are not sent to Smithers. Download or copy them before leaving; refreshing clears them.
Before you put it to work
- Separate interpretation from action.
- Compare with a simple baseline.
- Allow an unknown result.
- Test consequential mistakes.
Model confidence language is not a verified probability of correctness. Validate the actual task and design review around the consequences.
IF THIS LOOKS FAMILIAR
Your tools. Your particular mess.
The worksheet is yours to use. If the difficult part is making it work with the systems you already have, that's the kind of thing we help with.