Ship simple intake before adding AI

A founder we spoke to last quarter had spent eight weeks adding AI to her dental practice's website.

The model quality was good. The latency was fine. The transcripts were clean. None of that mattered, because the underlying contact form was still asking for "your message" in a single text area, and the new AI was paraphrasing whatever the patient typed before sending it on. The data going in was unstructured. The data coming out was unstructured-with-extra-steps.

This post is the longer version of why that failure was inevitable, and what the patient version looks like.

The mistake AI projects keep making

The mistake is treating AI as the solution and intake as the implementation detail.

Most of the AI-on-website projects we watch start the same way: a founder is convinced their visitors would convert better if there were "a chatbot" or "a voice agent" or "a smart form" — something AI-shaped. They spend three weeks evaluating vendors, two weeks integrating, one week tuning prompts, and they ship.

The result is usually a slick AI experience layered over a fundamentally vague intake question. The visitor still doesn't know what to say. The AI is still guessing what the business cares about. The brief on the other end is still a paragraph that needs interpretation. The AI has improved nothing about the underlying problem; it has only added cost.

The actual problem at the top of the funnel is rarely "we don't have AI." It's "we don't know what we want to know." Until that clarifies, AI is an expensive way to obscure ambiguity.

The intake-first sequence

Here's the order we recommend, in plain language:

Define what you wish you knew. What 4–6 things would make a partner, broker, or practice owner say "this is a real lead, send it to me now"?
Ship the simplest form that captures those 4–6 things. No AI. Maybe branching, maybe not. Cheap to build, cheap to change.
Run it for a quarter. Watch what comes through. Refine the schema based on what's missing. Every refinement makes you a better product owner of your own data.
Now add AI — to surface insights, to handle the conversational variants, to make the intake friendlier. Not before.

We've watched the impatient version of this and the patient version. The patient version compounds. The impatient version produces a demo that nobody can keep alive without the original developer.

What "the simple version" actually looks like

A workable intake form for a service business asks about seven things. Three required:

A name (so you can address the visitor)
An email (so you can respond)
A topic (because routing matters more than people expect)
A budget hint (optional)
A timeline hint (optional)
A short message (because some context is irreplaceable)
A honeypot field (to slow the spam)

That's it. No live chat. No AI clarification. No 12-field qualification deluge. And — this part matters — the data that comes through is already structured. You can filter, route, summarize, and prioritize it without a model in the loop.

When you add AI later, it writes into the same schema. The schema is the deliverable. The AI is the input device. The brief is the product.

Why this is hard for studios

A studio's incentive is to sell the AI. AI is novel; AI commands a premium; AI is what gets the press. Selling "we'll fix your intake form first" is a harder pitch. It sounds basic. It sounds like the kind of thing the client could do themselves.

But the client couldn't do it themselves, because the question "what do we wish we knew about every inquiry" requires sitting in their seat for a week and watching the partners triage. It requires asking "why didn't this lead convert" enough times that patterns emerge. That's the real work. The form is the artifact at the end.

We've made our peace with leading with the unsexy thing. The clients who get it are the ones we want.

Three questions to ask before adding AI

If you're considering an AI intake project, before signing the contract, answer these:

Do you have a structured definition of "good lead" vs. "bad lead"? If not, you're not ready. AI will not invent this for you.
Can you describe a single inquiry from this week using exactly the schema you'd want? If you can't produce one example yourself, the AI cannot produce examples reliably.
What does the brief look like when it arrives? If you can't draw the email or the JSON or the Slack message, you're describing a feature, not a deliverable. Define the deliverable; build backwards.

If those three questions are easy, you're ready for AI. If they're hard, you have a smaller, cheaper project to do first. That smaller project will earn the right to do the bigger one.

Closing

The contact form is the part that has to work first. AI is what makes a working form better, faster, smarter. It is not what makes a vague form clear.

If you're at the start of an intake project, do the simple thing first. Ship it. Watch it. Refine it. Then call us about the rest.

Want to see what we'd write for your intake? Tell us about it.