I've spent the last couple of years designing conversational and AI-assisted products, mostly in the B2B SaaS space. The single most useful principle I've arrived at fits in one sentence: the model interprets; the code executes.
Let me unpack why that split matters for UX.
Don't fake understanding with lists
Early on, every team's first instinct is to handle language with keyword lists, regex, and hand-maintained alias tables. "If the message contains 'price' or 'cost' or 'how much', route to pricing." It works in the demo and collapses in week two, when a real user writes "what would this run me monthly?" and hits the fallback.
Language understanding is exactly what LLMs are good at. Hand the interpretation to the model. But — and this is the design decision — never let the model execute anything directly. The model produces a structured intent; deterministic code validates the schema and performs the action. If validation fails, the user still gets a coherent text response; only the action is withheld. The conversation never dead-ends, and the system never does something the model hallucinated.
That failure mode — text degrades gracefully, actions fail closed — is a UX property. Users forgive a bot that answers imperfectly. They don't forgive one that books the wrong appointment.
Understand needs before you automate them
The temptation is to bolt AI onto whatever flow already exists. The better question is: where in this journey is the user translating their intent into our system's vocabulary? That translation work is what AI should absorb.
In a marketplace context, that meant replacing a filter panel nobody used correctly with a plain-language search that mapped to the same underlying filters. Users didn't want "AI features." They wanted to stop learning our taxonomy. Same backend, radically different comprehension load.
Design the trust budget
Every AI interaction spends or earns trust. A few rules I now treat as defaults:
- Show your interpretation. When the model maps "cancel my Friday slot" to a specific booking, echo it back before acting. One line of confirmation costs a tap and prevents the catastrophic error.
- Cap the blast radius. Reversible actions can run autonomously. Irreversible ones — payments, deletions, messages to third parties — always get an explicit confirm.
- Never say "unlimited," never overpromise. In fintech especially, copy that oversells the AI creates support tickets, not delight.
Instrument the misses
The most valuable research artifact in an AI product is the log of interpretations that failed validation. Each one is a user need stated in the user's own words that your system couldn't serve. I review these weekly the way I used to review usability test recordings. It's cheaper than a research sprint and brutally honest.
Integrating AI into UX isn't about adding a chat box. It's about deciding, deliberately, which layer interprets and which layer acts — and making sure the seam between them is invisible to the user and inspectable by you.