Walk into any pitch this year and someone will ask for 'AI on the website.' What they almost always mean is a chatbot — a friendly bubble in the bottom-right that answers FAQs and, on a good day, books a call. We build those. We also think they are the least interesting thing AI can do for a brand, and the reason so many of these projects quietly underwhelm six weeks after launch.
The framing is the problem. A chatbot is a feature: a box you bolt on, scoped by how many questions it can deflect. A sales engineer is a role: the person who joins the call when the conversation gets technical, who knows the product cold, who can read whether you're a tyre-kicker or a signed contract waiting to happen — and who knows exactly when to bring in a human and get out of the way.
When you brief AI as a role instead of a feature, almost every decision downstream changes: what it needs to know, what it's allowed to decide, how you measure it, and what 'good' looks like. This essay is the pattern we use to make that real — four layers, in the order we build them.
A chatbot deflects questions. A sales engineer moves a deal forward.
The distinction the whole essay turns onIt has to remember the brief
The first layer is memory, and it's the one most teams skip. A generic assistant starts every conversation from zero — it has no idea who you are, what page you arrived on, what campaign sent you, or what you asked the last time you visited. A sales engineer who reintroduced themselves every five minutes would be fired by lunchtime.
Memory means the concierge is grounded in two things: the brand's own context — positioning, services, the real difference between the Sprint and the Programme — and the visitor's session — referrer, pages viewed, the half-finished form, the question they asked before they got distracted. None of that is exotic. It's a retrieval layer over your own content and a thin profile of the session. But it changes the first sentence the AI gets to say from 'How can I help?' to something that already knows why you're here.
In practice this is unglamorous. On a live site, the concierge reads the section you're sitting in when you open it — offering to scope work on the Services page, offering to summarise the essay on an Insights page. Same model, different memory.
It has to know what you actually sell
The second layer is the one clients are most nervous about and the one that pays for the whole project. A help-desk bot can get away with knowing your FAQ. A sales engineer has to know the inventory: what you offer, what it costs, what's in scope, what isn't, and where the genuine trade-offs live.
This is uncomfortable because it forces a decision most brands defer: how much of your commercial reality are you willing to put in front of a machine that talks to strangers? Our answer is a tiered one — and it's the heart of the pattern.
We tier it. Public knowledge — positioning, services, process, the shape of an engagement — the AI shares freely, because it's already on the site. Qualified knowledge — indicative ranges, typical timelines, what moves the price — it shares only once a visitor has shown genuine intent. Human-only knowledge — specific quotes, contract terms, anything that commits the studio — it never touches; it names the boundary and hands over.
Notice the AI isn't deciding the commercials. It's deciding which tier a given visitor has earned — and that's a judgement call, which is the next layer.
It has to qualify, and it has to say no
A good sales engineer spends as much energy disqualifying as qualifying. The wrong-fit lead who consumes three weeks of scoping is more expensive than the deal you politely declined on day one.
So the third layer is judgement: the concierge reads the conversation for the signals that actually predict fit — budget posture, timeline, decision authority, the specificity of the problem — and routes accordingly.
Crucially, 'routes accordingly' sometimes means saying no. A concierge that enthusiastically promises everything to everyone isn't helpful; it's a liability that books calls your team has to un-book.
What it never does is invent a number, promise a date, or commit the studio to anything a partner would have to honour. Within those limits, every turn it re-reads fit and picks one of three moves: go deeper, hand to a human, or wind down gracefully.
It has to know when to step back
The final layer is the one that separates a tool from a teammate: knowing the moment to stop talking and bring in a person.
The best sales engineers are defined as much by their handoffs as their pitches — the clean briefing note, the warm introduction, the context the human doesn't have to re-gather.
So the concierge's job at the end isn't to close. It's to hand over a deal that's already warm: a short, structured summary of who the visitor is, what they need, which tier of knowledge they reached, and why it thinks they're worth a call.
Here's the test. If your AI's proudest metric is 'questions answered,' it's still a chatbot. If it's 'qualified conversations handed to a human, with context,' you've built the sales engineer.
Build the role, not the widget
None of these four layers requires a frontier model or a research budget. They require a decision: to treat AI as a colleague with a job description rather than a feature with a backlog.
Memory so it knows the context. Inventory so it knows the product. Judgement so it qualifies honestly. Handoff so it knows when it's done.
Do that, and the bubble in the corner stops being the thing you apologise for in month two. It becomes the most patient, best-briefed member of the team — the one that's always on, never tired, and smart enough to know when to get a human.
That's not a chatbot. That's a sales engineer.