Where AI adds real value to a digital product
A product-first framework for deciding when AI should shape the experience—and when conventional software is better.
Begin with the job, not the model
AI is most useful when it changes the cost, speed or quality of a meaningful job. Starting with a model capability often produces an impressive interaction with no durable reason to exist.
Describe what the user is trying to accomplish, what makes the task difficult today and which part benefits from interpretation rather than deterministic rules. If ordinary software can deliver the result more reliably, use it.
Look for compression and judgement
Strong AI product opportunities often compress an expensive sequence: reading many inputs, finding patterns, producing a first draft or turning an unstructured request into a structured action.
The value is not the generated output alone. It is the time saved before a person can review, decide or continue. That makes the review experience, evidence and editability as important as the model response.
- Summarising a large, changing body of information
- Transforming unstructured material into a useful format
- Ranking options where the criteria can be made explicit
- Creating a first version that a skilled person can improve
Design the failure state first
Every model will be uncertain or wrong sometimes. The product needs a clear response: show sources, express confidence, ask for clarification, require approval or fall back to a deterministic path.
A useful AI interface helps the user understand what the system did and what still needs judgement. Hiding uncertainty may make a demo feel smoother, but it makes a live product harder to trust.
In Wheerli, our visual fashion search product, finding a similar garment is only part of the job. The pipeline also verifies product availability before showing a result, and a monthly cost ceiling bounds the search system’s operating costs. Wheerli is preparing for launch: catalogue coverage is still being validated. Those boundaries are part of the product design, rather than evidence of a proven commercial result.
Evaluate the outcome continuously
Quality needs a concrete definition before launch. Build a small evaluation set from representative tasks, score the parts that matter and keep examples of unacceptable output. Measure the user outcome alongside model performance.
The best AI products combine flexible intelligence with hard product boundaries. Models handle ambiguity; software enforces permissions, records decisions and keeps important states dependable.