Most AI features in business software exist because someone decided the product needed AI. A chatbot here, a summary there, maybe some auto-generated content. They demo well. They rarely survive contact with actual daily use.
The AI features we build start from the opposite direction: what repetitive, high-volume task is consuming human attention that could be better spent elsewhere?
For a counselling practice, that meant clinical note enrichment — extracting structured data from free-form progress notes so practitioners spend less time on paperwork and more time with clients. For a legal platform, it meant document intelligence — parsing co-parenting agreements to surface clauses and obligations that matter to the user's specific situation.
Both features are invisible. There's no chatbot, no 'Ask AI' button. The intelligence is woven into the workflow so naturally that users forget it's there. That's the goal.
We use Claude for most of our AI work — Sonnet for high-volume processing where cost and speed matter, Opus for complex reasoning where accuracy is critical. The model selection is part of the engineering, not a marketing decision.
If your AI feature needs a tutorial to explain why it's useful, it's solving the wrong problem.