Every AI intake demo sounds the same. A friendly voice answers, repeats the caller’s words back, and books something. It looks like the problem is solved. Then you point it at a real client’s phone and watch it come apart on the third call.
The gap is not the model. The gap is everything the model was never told. A general assistant knows how to talk. It does not know your client’s book of work. Intake is not a conversation problem. It is a knowledge problem wearing a conversation’s clothes.
What “AI-powered” actually ships
“AI-powered” describes the engine, not the fit. A general assistant answers any question with confidence, which is exactly the danger. Ask whether the plumber covers a suburb forty minutes out and it guesses. Ask it to book an emergency callout at 2am and it offers a slot the client would never staff. Confident and wrong is worse than slow and right, because the caller believes it.
The demo hides this because the demo controls the questions. Real callers do not. They ask about the one service the client stopped offering last year, the pricing the client keeps deliberately vague, the after-hours rule that lives only in the owner’s head.
The configuration layer
A vertical-shaped agent is a general model plus a specific book of knowledge about one client’s business. That second half is the configuration layer, and it is where intake is won or lost. It holds the things a good receptionist learns in their first month on the job.
Services: what the client actually sells, and the jobs they refuse. Service area: drawn to the postcode, not the metro. Hours: the real ones, including the difference between open and answering. Escalation rules:which situations are a booking, which are a human callback, which are a genuine emergency that rings someone’s mobile now. Booking rules: what can be scheduled without a person, what needs a quote first, what deposit to mention.
None of that is in the model. All of it is in the configuration. Change the client and every line changes with them, which is why a setup that fits one business is worthless bolted onto the next.
The same call, two agents
Here is the difference in the only place it matters: on the phone. Both transcripts below are illustrative, written to show the pattern, not transcribed from a real call.
The first agent is not broken. It is empty. It booked a routine slot for an emergency, in a suburb it never checked, because nobody told it any different. The second one knew the area, knew a leak is not a scheduling question, and knew the booking rule for urgent work. Same model, opposite outcome.
What happens when it does not know
A configured agent is also honest about its edges. The point of escalation rules is that the agent knows which questions are not its to answer. A price it was never given, a job outside the service area, a situation the client wants a human on: those hand off to a callback or ring a real person, rather than getting improvised. An agent that never says “let me get someone to confirm that” is not smarter. It is just more likely to be confidently wrong.
“A general model can hold a conversation. Only a configured one knows your client’s business well enough to be trusted with their phone.”
Why this is the agency’s advantage
The configuration layer is work, and that is good news for you. It is the part a generic vendor cannot do at scale, because it takes someone who understands the vertical and the individual client. That someone is the agency. You already know what home-services callers ask, which jobs pay, which ones waste a truck. The platform gives you a voice that is included in the plan and an agent that answers, books, and follows up. The knowledge that makes it right is yours to load in.
Home services is live today, with veterinary and legal intake next. Each starts from a vertical pack and gets shaped to the individual client. That shaping is not overhead. It is the thing that turns a chatbot into an intake your client would trust with a real emergency.