Most AI features on business websites are a chat bubble in the corner that nobody clicks. The company paid for it, the agency shipped it, and it answers questions the FAQ already answered — worse, and more slowly.
Meanwhile the things models are genuinely good at go unbuilt, because they're less demo-able. Here's the split, from features I've actually shipped rather than features that photograph well in a deck.
What works
Search that understands a question, not keywords
The highest-value, least-glamorous AI feature there is. Your site search currently fails anyone who doesn't already know your internal vocabulary — someone types "does it work outdoors" and gets nothing, because your page says "IP65 rated." Semantic search fixes exactly that gap, and unlike a chatbot it has a visible, measurable before-and-after: search sessions that end in a result instead of a bounce.
Support that answers from your own documentation
Useful when — and only when — it answers from your content and links to the source. That's retrieval, not a chatbot with your logo on it, and the distinction is the entire ballgame. I've written up when it's the right call in RAG for business websites.
Drafting the first version of something a human then edits
Product descriptions from spec sheets, alt text for a large catalogue, a first-pass summary of a long document. The pattern that works is always the same: the model produces a draft, a person approves it, and nothing reaches a customer unread. Anywhere a human stays in the loop, this is reliable and immediately economic.
Extraction and routing
Pulling structured fields out of unstructured input — a CV, an invoice, a long enquiry — and sending it to the right place. Invisible to the customer, and often the highest actual ROI on this list, because it removes a task someone is doing by hand every day.
What doesn't work
A general chatbot on a small site. If your site has twelve pages, a visitor can read the relevant one faster than they can negotiate with a chat widget. You've added a slower path to information that was already there.
Anything that must never be wrong, without a retrieval layer. Refund policy, pricing terms, legal or medical specifics, delivery commitments. A model asked to answer from memory will produce something confident and occasionally invented. That's not a bug you prompt your way out of — it needs the answer grounded in a source, or it needs to not be an AI feature.
AI as a headline feature on a page with a broken conversion path. If the buy button hides on mobile and the page takes six seconds to load, an assistant does not help. Fix the leak first.
The question worth asking first
Not "where could we use AI?" but: what task does someone here repeat, that has a right answer, that we currently do by hand?
That question routes you to extraction, routing, drafting and search — the things that pay for themselves. The other question routes you to a chat bubble.
What shipping it properly involves
The model is the easy part. Production is the rest:
- Evaluation on cases you care about — a fixed set of real inputs and expected outputs, so you can tell whether a change improved things or just changed them.
- A fallback. Models are slow sometimes and unavailable occasionally. The feature must degrade to something useful rather than a spinner.
- Cost caps and logging. Per-call cost is small; unbounded per-call cost across a traffic spike is not.
- An honest "I don't know" path. The single most valuable behaviour you can build in, and the one most often skipped.
An AI feature without those four is a prototype that happens to be in production.
I'm Divyansh Sood. I build AI features into websites and products — and I run my own AI product, WebSeek.ai, on three model providers simultaneously, so the trade-offs here are ones I've paid for rather than read about. Describe the task you'd want automated and email me — if the honest answer is that you don't need a model for it, you'll get that instead.