Sadad Siddi · 18 January 2026 · 5 min read
AI is useful when it removes repetitive judgement or summarises information a person would otherwise hunt for. It is expensive theatre when it is bolted onto a product that still cannot do the basic job reliably.
A simple filter
- Is the task frequent, bounded, and currently done by a person copying or classifying information?
- Can a wrong answer be caught quickly by a human, or would it cause silent harm?
- Do we have the source documents, logs, or structured data needed to ground the model?
- Will the feature still be valuable if the model is slow, uncertain, or unavailable?
Prefer tools around the work
The highest-return applications we see are unglamorous: drafting replies from existing knowledge, extracting fields from documents, routing tickets, and flagging anomalies. They sit beside a durable system of record. The model does not become the system of record.
If a workflow fails those tests, we automate with rules, better UX, or a clearer process. Intelligence should earn its place in the product the same way every other feature does — by making the business measurably easier to run.
- ai
- automation
- product strategy