The pattern that works
Start where AI assists rather than decides — drafting, summarising, suggesting. Users stay in control, errors are visible and correctable, and you learn how your customers actually use it before betting anything on autonomy.
Adding AI to a working product is a different problem from building an AI product. The constraint is not capability — it is not destabilising something customers already depend on.
Where to start
| Feature type | Risk | Value | Good first move? |
|---|---|---|---|
| Summarise existing content | Low | Moderate | Yes |
| Draft text the user edits | Low | High | Yes |
| Search over user's own data | Low | High | Yes |
| Suggest next actions | Low | Moderate | Yes |
| Auto-categorise records | Moderate | High | With review |
| Take actions automatically | High | High | Later |
| Make decisions for the user | Very high | Varies | Rarely |
"Draft, don't send" is the highest-value low-risk pattern. The AI composes; the user reviews and sends. You capture most of the time saving, errors are caught before they matter, and users feel in control — which is what determines whether they keep using it.
The infrastructure decisions that are hard to reverse
- Abstract model calls behind your own interface from the first feature. Provider switching becomes configuration rather than refactoring.
- Cost attribution per customer. Without it you cannot tell which accounts are expensive, or price the feature rationally.
- Decide the data policy before building. What customer data leaves your environment, under which vendor terms, and what you tell customers.
- Build the evaluation set early. It is what lets you change models or prompts safely later.
- Instrument quality signals — acceptance rate, edit rate, regeneration rate.
The data policy decision is the one that causes retrofits. If your terms of service and DPAs do not cover sending customer content to a third-party model provider, you may need to update them, notify customers, or offer an opt-out. Finding this out after launch is considerably worse than before.
Pricing and cost exposure
Unlike most features, AI has a marginal cost per use. That changes the economics:
- Model the cost per active user at realistic usage before launch.
- Identify your heaviest plausible user — a small number of power users can dominate spend.
- Set per-plan limits, communicated clearly rather than enforced silently.
- Route simple requests to cheaper models — most calls are not the hard ones.
- Cache aggressively where content is stable.
A feature that delights users and loses money on your heaviest accounts is not a success. Model the worst-case user, not the average one.
Designing for trust
Adoption depends on whether users believe the output:
- Show sources for anything factual, linked to the underlying record.
- Make output editable rather than final.
- Indicate uncertainty honestly rather than presenting everything with equal confidence.
- Allow regeneration with feedback, so a poor result is not a dead end.
- Never hide that it is AI-generated. Users find out, and discovering it feels like deception.
- Provide an off switch for customers who prefer not to use it.
Rollout that protects the existing product
- Internal use first — your own team on your own data.
- Opt-in beta with a small group of engaged customers who will give feedback.
- Measure acceptance and edit rates — a high edit rate means the output is not good enough yet.
- Widen gradually, watching cost per customer alongside usage.
- Default on only when quality and economics are both proven.
Signals worth watching
| Signal | What it tells you |
|---|---|
| Acceptance rate | Whether output is usable as-is |
| Edit rate and size | How far off it is |
| Regeneration rate | First attempt failing |
| Feature abandonment | Tried once, not returned to |
| Cost per active user | Whether the economics hold |
| Support tickets mentioning it | Confusion or distrust |
What not to do
- Do not add AI because competitors did. A feature nobody asked for still costs money to run.
- Do not make it unavoidable. Users who do not want it should be able to work as before.
- Do not let it write to customer data unreviewed in the first version.
- Do not launch without cost attribution. You will not be able to price it or diagnose a bill.
Adding AI to a product customers already rely on? Tell us the workflow you want to assist. See our AI agent service, cost control, and observability.