AI in marketing
Where AI belongs is a workflow question.
Start with the work a system needs to perform, then decide whether a language model, a fixed rule or a person should handle it.
Begin with a bounded task.
“Add AI to marketing” does not establish an input, a useful output or an acceptance condition. “Summarise these approved documents and link each claim to its source” is a task that can be evaluated.
Describe the frequency of the work, the information it needs and the consequence of an incorrect result. Those conditions help determine the suitable approach.
Choose the mechanism deliberately.
| Task | Starting point |
|---|---|
| Route a lead by a fixed territory rule | Deterministic logic |
| Draft a summary from supplied material | Assisted generation with source checks |
| Decide a sensitive exception | A responsible human reviewer |
| Retrieve a known record | Search or database retrieval |
Define the evidence of a useful output.
A fluent answer is not enough. Check required fields, source support and prohibited conclusions. For an operational workflow, also check whether the next system can use the output and whether the person receiving it understands its limits.
Use representative difficult inputs during testing, including missing or contradictory information. A demonstration on a clean example cannot establish dependable handling of real work.
Keep a fallback.
Decide what happens when the model cannot produce an acceptable result or a connected service is unavailable. The answer may be a retry, a review queue or a direct handoff to a person. It should not be silent invention.
Measure useful work, not generated volume.
Review accepted outputs, corrections, exceptions and operating cost. Time saved is meaningful only when the task is still completed to the required standard. Expand the workflow after the evidence supports it.