When not to build with AI
A practical test for when rules, search or ordinary software beat machine learning on cost, latency, auditability and trust.
Start with the decision, not the technology
The useful question is not 'where can we use AI?' but 'which decision or task is slow, costly or inconsistent today, and why?'. Once the cause is clear, the right tool usually becomes obvious — and it is often not a model.
Machine learning earns its place when the pattern is too complex or changeable to write by hand and enough representative examples exist to learn it. Outside that zone it adds cost and uncertainty without adding value.
Compare the options honestly
Each approach has a natural fit. The table below is a starting point for a feasibility conversation, not a rule book.
| Situation | Usually better choice | Why |
|---|---|---|
| Logic fits in a policy document | Rules or workflow software | Predictable, cheap to run and easy to audit. |
| Users need to find known documents | Search with good metadata | Simpler to evaluate; add generation only if summaries are genuinely needed. |
| Few or unrepresentative examples | Collect data first | A model cannot learn a pattern it has not seen. |
| Complex, variable patterns with labelled history | Machine learning | This is where models outperform hand-written logic. |
Warning signs a project should pause
None of these mean AI is impossible — they mean the groundwork is missing.
- Nobody can describe what a correct output looks like.
- Every error would need a full, case-by-case explanation.
- No owner is available to monitor quality after launch.
- The expected saving is smaller than the cost of evaluation and maintenance.
The hybrid path
Many reliable systems combine both: rules handle the clear cases and hard constraints, and a model handles the ambiguous remainder with a human review step. This keeps the auditable parts auditable while still capturing value where patterns are fuzzy.
Working on something similar?
Discuss your workflow, available data and operating constraints with Talee’s engineers to assess a suitable next step.
Talk to an engineer