Report · Thursday 3 September 2026
You asked three questions on Tuesday. Two now have real answers, paid for and billed, not guessed. The quality control agent you asked for is built: one agent writes, a different agent checks it, and no person is in that loop. It has been shown 27 bad drafts and it stopped all 27. And the twelve emails you thought were still owed turned out to be written already — by you, on 26 August.
You asked what the newsletter test would cost. Here it is, off the bill.
One client, one month, both newsletters
$0.2610 cents to write the two letters — one to their past customers, one to
their agents.
16 cents to check them both before they go anywhere.
That is not an estimate. It is what the bill said, for a real run, on a real month, using your own plan. And it is per client — so:
| If we have… | Newsletters cost | A year |
|---|---|---|
| 2 clients (today) | 52 cents a month | $6 |
| 50 clients | $13 a month | $156 |
| 250 clients | $65 a month | $780 |
One number in the last report was wrong. I am withdrawing it.
On Tuesday I told you the four agents would cost about $1.28 each, "corrected" from a measurement. That correction came from one job. I have since measured a second one, and my method was wrong in the opposite direction — too high, not too low. There is no single fix for both. So the corrected figures are gone, and I now quote only what has actually been billed.
There are two kinds of cost in this business, and it helps to keep them apart.
Costs that go up every time you sign a client. The phone calls, and the agents. Sign ten more inspectors and this goes up ten times. That is the table below.
Costs that stay the same whoever signs. The server, GoHighLevel, Hamming, and the hours people work. Two clients or two hundred, these barely move. Beth is working those out now.
Only the first kind is in this report. Here it is.
| What goes up with each client | One client, a month | Share |
|---|---|---|
| Josh — the phone calls | $30.71 | 96% |
| All five agents together | $1.44 | 4% |
| One client | $32.15 |
So Chad and Ted together cost us about $64 a month right now. Fifty clients would be about $1,600 a month. Two hundred and fifty, about $8,000.
That $1.44 is five agents added together. When you first read this report, two of the five were my estimate. Both have now been run. Here is what they actually cost, beside what I had guessed:
| The two that were guesses | I guessed | It cost | How wrong |
|---|---|---|---|
| The Brain Keeper | $0.18 | $0.50 | nearly 3x too cheap |
| The Microsite Keeper | $0.35 | $0.14 | 2.5x too dear |
Wrong in both directions again, which is why they were worth running rather than arguing about. Nothing in the table below is a guess any more.
| Agent | What it does | A month | Real? |
|---|---|---|---|
| The Checker | Did the service actually happen? | $0.54 | billed 2 Sep |
| The Newsletter Writer | Writes the month's two letters | $0.10 | billed 3 Sep |
| Quality Control | Checks work before it goes out | $0.16 | billed 3 Sep |
| The Brain Keeper | Do Josh's four records agree? | $0.50 | billed 4 Sep |
| The Microsite Keeper | Keeps a town page current | $0.14 | billed 4 Sep |
| All five | $1.44 |
The Brain Keeper paid for itself on its first run
Fifty cents. It read the four descriptions of Chad's business side by side — what a caller hears, Chad's form, the backup copy in the code, and a fortnight of real calls — and found seven places they disagree.
One of them was a price. On 25 August you ruled that the final new-construction visit is priced by square footage, like any other inspection, not at the flat $425 the earlier phase visits cost. That ruling reached the record Josh reads from. It never reached the backup copy, which still said $425 flat.
No caller was ever quoted wrong — the record wins, and Chad has one. But the backup is what Josh falls back on if that record is ever lost or reset, and on a 5,000 square foot home the difference is $175 out of Chad's pocket, every time. It is fixed, and there is now a test that fails the moment it drifts back.
Every conversation we have had about "what will the AI agents cost at 250 clients" has been about 4 cents in every dollar. The phone is the business. Inside the phone, the voice alone — ElevenLabs, the thing that makes Josh sound like a person — is 63% of a call. More than the phone line, the recording, the listening and the thinking put together.
And you already made the right call on that
There is a cheaper voice. It would cut the cost of a call by about a third. You listened to both on real calls and kept the dearer one, because the cheap one made a mess of street names. That was the right decision and this report does not reopen it.
Please do not set a price off $32
$32 is only the first kind of cost, and the second kind is bigger than it — much bigger, today.
GoHighLevel is $497 a month whether we have two clients or two hundred. With two, that is $249 a client. Then the server, then Hamming, then every hour you, Beth or I spend.
Beth is working that side out now. Until she has, a price built on $32 would sell the business at a loss.
The second kind of cost does not go up when you sign someone. So it gets cheaper for every client, every time you add one. GoHighLevel on its own shows the shape:
| Clients | GoHighLevel, per client |
|---|---|
| 2 — where you are today | $249 |
| 12 | $41 |
| 50 | $10 |
| 250 | $2 |
Same bill every month. The tenth client costs almost nothing to serve and pays the same as the second. That is where the money in this business is — not in trimming what the agents cost, which is $1.44.
You said this one first, and you said it should go in before a person reads anything. It is built. It has been run three times.
| What it checked | Drafts | Bad ones | How many it stopped |
|---|---|---|---|
| Review replies — your own checklist | 12 | 10 | 10 |
| Google Business posts | 13 | 11 | 11 |
| Newsletters | 8 | 6 | 6 |
| Total | 33 | 27 | 27 |
Nothing bad got through. Among the things it caught, in a client's own name: a service the company does not sell · a town forty miles outside their area · a price from the wrong price list · our own subscription price, on a client's Google profile · the words "passed" and "failed" about a house · and a promise to text a homeowner.
It costs about 2 cents to check one item.
The interesting part
Three times now, it has found a fault I had missed when writing the test. It was right each time and I was wrong. Once it noticed that a post pitching our service in their name was not just a price problem — it was selling a service the inspector does not offer. I had only marked the price.
It reads the checklist more carefully than the person who wrote it. That is the whole argument for putting it in the gate.
This is the idea underneath the whole thing, and it is your idea, not ours. You said it on Tuesday:
Your words, 2 September
"If we have to human review all of these things at scale, this product will be unaffordable."
"Scaling human labour to meet 250 or 1,000 — that's not doable."
You were right, and we had drawn it wrong. Our first plan had a person saying yes to every post. That works at two clients and is impossible at 250.
So one agent writes, and a different agent audits it. The writer never gets to send anything. It hands its work to a checker, the checker reads it against a written checklist, and only work that passes goes anywhere. No person is in that loop.
| Step | Who does it |
|---|---|
| Decide the month's theme | A person — once a month, not once a post |
| Research | Agent |
| Write it | Agent |
| Check it against the checklist | A different agent |
| Publish it | GoHighLevel |
The person moves from approving every item to choosing the theme for the month. That is a management decision, made once, and it does scale.
It is a fair thing to doubt, so here is the honest answer.
The best evidence is not the 27 out of 27. It is that the checker has now caught three faults that I missed when I wrote the tests. I marked a draft as having one problem; it found two, and it was right all three times. Once it spotted that a post selling our service in the client's name was not just a price mistake — it was offering a service that inspector does not sell.
It reads the checklist more carefully than the person who wrote the checklist. That is the argument for putting it in the gate, and it is a better argument than any score.
But this is not "no human" yet, and we should not pretend it is
You set the rollout yourself and it is the right one: "Before we check manually, I want to put that quality control agent in there, so what we see has already been reviewed by the AI agent."
So the agent checks first, and a person still looks — until the people are comfortable. Then the person stops looking. The person steps out at the end of that, not at the start.
Three things have to be true before that day: Beth's real checklist replaces our stand-in · it runs beside her for a month and agrees with her · and nothing it passes turns out to be wrong. Twenty-seven for twenty-seven is a good start. It is not a month.
One more thing worth saying plainly. The sending is now blocked in the software, not by anybody remembering. A post that has not been through the checker cannot be sent — not "should not", cannot. And a pass has to name which checklist judged it and which run, so "it passed" is something a person can go and check rather than something we say.
I had this on your list as the thing blocking the newsletter. It was not.
The twelve for clients and the twelve for agents were written on 26 August, from your own instruction on the call — the real estate coach, the theme for the month, and let the article be picked to fit. It has been in the system since. I listed it as owed because I did not check first.
So September for a Texas inspector already reads: the autumn walk-round — closing the house down for the season, with the local angle that the cooling system has just done six months, so service it now. Delaware gets a different one. That is your design, and it is what the test ran on.
Your radon rule works too
December is the radon month. A firm that sells radon keeps it. A firm that does not gets a different December instead — not a hole in the year, and no email asking a homeowner about a service their inspector cannot provide. We tested both and both behaved.
Two of the five in that table are new this week. Here is why each one exists.
The Brain Keeper. Nobody asked for this one. It reads four descriptions of one business — what a caller actually hears, what the client typed on the form, what our code says, and real transcripts of real calls — and reports where they disagree.
It exists because of the pre-drywall fault in August. The $425 was corrected in the wrong place, so Josh went on refusing a job Chad sells, and nobody knew for a week. That fault takes one person one minute to fix — once somebody knows.
The Microsite Keeper. Keeps a town page current for the month and fixes anything on it that disagrees with the client's own record. It proposes the change. A person still says yes.
A way to publish a blog post. Dil found that another setup is doing keyword → blog post → Google Business post, and it works. We can do the same here, and the piece that was missing is now built.
Worth saying plainly: GoHighLevel cannot do this part. It has no way for software to create a page — only a person clicking. WordPress does. We wrote that down in June and never acted on it.
Five things. Four of them are one sentence each.
The Google Business Profile free offer. Still the oldest one on the list. Parker is still promising it to people right now.
The call to action on a bad review. You wrote the review-reply checklist yourself and it says every reply asks for a referral. Should a reply to an angry review still do that? We have guessed no. One word from you settles it.
Beth's checklist for Google posts. You said it: she knows what she checks before she posts. We wrote a stand-in overnight so testing could carry on, and five of its rules are ours, not hers — each one marked as ours in the file. Her real one replaces it.
Does a newsletter ask for anything? Your review checklist has a call to action because you specified one. Nobody has said whether a monthly newsletter should.
Three bookings need a human eye. The first agent we ran found three calls where Josh told a caller they were booked, and no booking was saved. That is a week old now and still nobody has opened the CRM to look. It is not agent work — it is ten minutes with the records.
| Thing | Why we left it |
|---|---|
| Publish anything | Everything the agents wrote is a draft. Nothing reached a customer. |
| Set a fee | Yours. We built the cost side; the price is a decision, not a sum. |
| Quote a price anywhere | Nothing is agreed. The site still says pricing is being finalised. |
| Touch Chad's or Ted's records | The agents find and report. They never change a value. |
| Re-run the first agent | Already measured on Tuesday. Spending again to learn the same thing is waste. |
You asked on Tuesday whether GoHighLevel should be doing this work instead of the agents, to hold the cost down. You were right about the shape and the numbers now say how right. GoHighLevel publishes and schedules; the agents write and check. And the writing and checking is four per cent.
Which means the cost question we have all been circling is the wrong question. Nobody should be trying to save money on the agents. The money is in the phone, and the only two levers there are how long calls run and which voice we use — and you have already looked at the voice and said no.
The one that would change the answer is the quality control agent, and not because of what it costs. It is what lets a person stop reading every post. Twenty-seven for twenty-seven is a good start and it is not proof. What would make it proof is Beth's real checklist instead of our stand-in, and then a month of it running beside her while she still checks. If it agrees with her for a month, she can stop.
Booked Solid Inspector · report written 3 September 2026.
Costs are from the Anthropic bills for five sessions run on 2 and 3 September, and from the
per-minute phone costing of 16 August. No price for Booked Solid appears in this report
because none is agreed.