What I do
The whole list, with nothing dressed up.
The first group is the foundation and everything below it plugs into it. Each line is written twice — the plain version in dark type, and the same claim stated properly underneath for anyone who wants to know what is actually going on. You never have to take all of it; most people start with the phone, because that’s the one bleeding.
What your business knows
The interpretable context layer. The part everyone skips, and the reason the rest of it is ever right.
Runs on ICM — Interpretable Context Methodology, a method authored by Jake Van Clief. I apply it; the implementation is mine.
Everything your business knows, written down in one place — your prices, your hours, your rules.
A version-controlled, human-readable repository — plain text, no proprietary format, no lock-in. You can read it, audit it, and take it with you. Nothing holds your operating knowledge hostage.
What’s different for each customer kept apart, so the answer matches the person who’s asking.
One entity, one branch. Because the separation is structural rather than statistical, the system cannot blend three customers’ terms into a plausible average — which is the most common way these tools embarrass the business running them.
Change it once and every part that talks to your customers says the new thing.
One source of truth, many surfaces. The phone, the site and the booking all read the same record, so there is no drift between what your website claims and what your phone says at nine at night.
Nothing goes in that you didn’t say. If it isn’t in there, nothing gets made up.
Provenance is a design constraint, not a policy. Because every fact has a known origin, “I don’t know” becomes a detectable state — which is what makes a safe hand-off possible at all. Competing tools configure themselves by scraping your website; that is precisely why they answer confidently and wrongly.
Your phone
The surface that costs you money every night you are closed — and the one with the most ways to go wrong.
Something answers when you can’t. It tells them your hours and prices and takes the job down.
Speech-to-text, a language model held at low temperature, and speech back — tuned for interruption handling, because real callers talk over a system that pauses too long.
It cannot invent a price, a timeline, or a warranty.
Refused in the system prompt and again in a post-generation filter. Two independent layers, because one instruction is a preference, and a confidently fabricated quote is the likeliest way one of these systems damages the business it answers for.
Emergencies get a real person. It never guesses on those.
A defined refusal set — gas smell, water through a ceiling, no cooling with an infant or an elderly person — bypasses the agent entirely and routes to a human or to emergency services. Non-negotiable, and it ships on day one.
If you can’t pick up, they get a text back in five seconds — so they don’t just call the next guy.
Missed-call-to-text driven off the same record, so the follow-up already knows who called and what about.
You see every call it took, in your own words, in one place.
And every hand-off is logged as a named gap. Fill it once and it stops being a gap — the system gets more accurate because the record does, not because something retrained itself out of sight.
It never records a call unless that was decided on purpose.
Recording is off by default: California is a two-party consent state, and that is a legal constraint that belongs in the build rather than in a terms page written after launch.
Getting found
So the people already looking for you can land on you — including the ones asking an AI instead of Google.
A website, so when someone looks you up they find you and not nothing.
Server-rendered and fast on a phone on a bad connection, which is where the decision actually gets made.
Your Google box — the one with your hours, phone and photos. I keep it right and post to it.
Google Business Profile management, plus reconciling the directory citations that quietly feed it stale data.
Getting your customers to leave you a Google review, automatically, instead of you having to ask.
Triggered off completed jobs in the same record, so the ask lands while the work is still fresh.
Being findable by the AI assistants people now ask instead of searching.
Machine-readable surfaces — an llms.txt, a full-text export, structured metadata and a real feed — so an assistant can retrieve and cite your business instead of paraphrasing a competitor. This site runs its own.
Fixing your address and hours everywhere on the internet that has them wrong.
Stale citations are the most common cause of a confidently wrong answer about you — from Google and from an AI alike.
Running it
The jobs you keep meaning to get to and never do.
They book themselves on their phone. You just see it show up.
Scheduling against real availability and real service durations, with the constraints enforced at the database rather than trusted to a form.
A list of everyone who’s ever bought from you — that you own.
Your data, exportable, on infrastructure in your name. Not a segment living inside somebody else’s marketing platform.
Texting your old customers when it’s slow.
Lifecycle and reactivation flows built off the same record, so the message knows what they actually bought.
One page a month: calls, reviews, where you show up.
Generated on a schedule by an automated workflow rather than assembled by hand — which is why it actually arrives every month.
What this costs
It depends on which of the above you actually need, so there is no number on this page and I’m not going to invent one to get you to click. Tell me what’s breaking and I’ll tell you what it takes to fix — on a call, in a straight answer.