What an AI agent can actually do on your site (in 2026)

Most "AI feature" pitches we see in 2026 collapse into four real patterns. Knowing the four — and what each costs to ship — saves a quarter of debate.

Pattern one: high-confidence search over your own content. RAG over a structured knowledge base. Costs and quality are now production-acceptable. Ships in two weeks.

Pattern two: drafting on behalf of the user. Form filling, email composition, briefing notes. Works best when the user reviews and edits — never when the output is read by someone else without a check.

Pattern three: summarisation of incoming content. Customer messages, sales notes, support tickets. Cheap, useful, low-risk because the original is still a click away.

Pattern four: decision-support. Visualising a recommendation, surfacing precedent, comparing options. Higher-stakes; needs evaluation harnesses; ships in months.

The patterns that are NOT ready for unsupervised production: end-to-end customer service without a human in the loop, automated medical or legal advice, anything where a model can directly cause irreversible action. They will be ready — they aren't yet.

If your AI roadmap doesn't read like "we'll start with pattern one and three in Q1, evaluate two in Q2, decide on four in Q3," you don't have an AI roadmap. You have a vendor pitch.