Where AI actually helps service businesses
Most service businesses do not need a custom model. They need the first five minutes after a lead arrives to stop being manual.
That is easy to agree with and hard to act on, because it points away from almost everything the category sells. So here is the argument in full: where the delay actually lives in a service business, how narrow the useful surface turns out to be, the two things that make an automated layer safe to leave running unattended, and the version of it answering on this page right now.
The delay is not a capacity problem
In a service business the bottleneck is rarely how much work the team can do. It is attention, and specifically whose attention, at what moment. A brief arrives at 9:40 on a Tuesday. The one person who can answer it properly is in a session until noon. By the time they open it, they are reconstructing context from a form with three fields, deciding which service it even belongs to, and writing from scratch an answer they have written a dozen times in slightly different words.
None of that is the expensive part of the job. All of it happens before the expensive part can start.
Meanwhile the person who sent the brief has done what everyone does: written to three studios in one sitting. The first substantive reply does not merely arrive earlier — it frames the problem. It decides what the project is about, which options are on the table, and what "good" would look like. Whoever frames the problem tends to get the work. That is the real cost of the wait, and it does not appear on any dashboard.
So the useful question is not what a model can do. It is which parts of those first minutes are genuinely mechanical. The answer is narrower than the sales deck and it is the whole opportunity.
The useful surface is four moves wide
The first move is intake. A contact form guesses at the question; a conversation captures it. The difference is not the interface, it is what ends up in the record — the brief in the sender's own words, plus the two or three commercial facts a first reply needs and a form is too rude to ask cold.
The second is routing: attaching a suggested service and a confidence level before a human reads anything. This is the unglamorous one, and it is where most of the time actually goes. Not answering the question — working out what kind of question it is.
The third is the first response, and the rule there is accurate before fast. A reply that is quick and wrong costs more than silence, because it has to be walked back by the same person you were trying to protect.
The fourth is the handoff: the conversation reaches a human with the context already attached. This is the move most deployments skip, and skipping it inverts the whole point. A bot built to deflect protects the calendar by spending the lead. For a studio that trade is backwards, because the conversation is the product.
Two things make it safe to leave running
The first is grounding. The agent answers from a versioned store of your own answers, not from what a model infers about your business at temperature. This is the difference between a system that represents you and one that improvises in your name — and versioned matters, because when it says something wrong you need to know which answer produced it and fix that.
The second is a ceiling. Every model call passes a quota gate. An agent with a language model behind it and no ceiling is an unbounded invoice, and the failure mode is not gradual: it is one loop, one bad night, one bot crawling your endpoint. A ceiling is a few lines of rule and the reason you can stop watching it.
Neither is clever. Both are the difference between a demo and something that runs on a business.
The version answering on this page
The concierge on this site is not an illustration of the service. It is the service, running on the studio's own leads, and it is documented as a case with its parts named: four moves from brief to human handoff, one Durable Object per conversation so each visitor has isolated state, a second Durable Object doing nothing but enforcing quota, a lead store across three versioned migrations, and a nightly digest that reads the same store the admin dashboard reads. The Worker behind it is typechecked before every deploy.
That last detail is the point of publishing any of it. The contact form and the site newsletter post to the same capture endpoint, so there is one store and one truth, which is what makes response time and lead quality measurable rather than felt. A system you cannot inspect is a claim. This one you can poke from the homepage.
What this rules out
Most of the category. Not because the tools are bad, but because they answer a question service businesses do not have. A custom model is a research budget. A dashboard of conversation analytics is a report nobody opens on a Tuesday morning. An agent that closes deals on its own is a liability wearing a product's clothes.
Start where the delay is, not where the demo is. If your bottleneck is the offer, or the site, or the traffic, an intake layer will make a confusing business respond to confusion faster. Fix that first — and any studio worth hiring will tell you so at the diagnosis, not after the invoice.