per Josh minute — Telnyx, Deepgram, Claude and ElevenLabs. That is 43¢ a booked call.
Confirmed on the live serverBooked Solid Inspector · Cost register · 27 August 2026 · tool costs now real
Every cost line this business has, in the shape Ken asked for: divided by minutes. The calls are settled, the engines are settled, and the tool bill is now real — $2,391 a month. One line is still missing, and it is the one that decides the price.
I have to understand what my total costs are. It's not just Josh… If I price blind and I underprice, I lose money on every transaction. Why would I do that? Ken Compton, 26 August
The answer on one page
per Josh minute — Telnyx, Deepgram, Claude and ElevenLabs. That is 43¢ a booked call.
Confirmed on the live serverper client, per month — all four engines, every email, the drift monitoring, the number rental. Rises slowly with the forever list.
Computed from live ratesa month of tools — Hamming, GoHighLevel, two Claude subscriptions, the server. Payroll and the rest of the overhead are still missing.
Half known · the rest is Ken and BethAn average client costs about $33.75 a month to serve, before a penny of overhead. 480 Josh-minutes at $0.0641 is $30.77, plus $3.00 for everything else we run for them — about $35.20 by year five, as the retention list grows.
The technology is 7 cents a minute. The tool bill is $2,391 a month — a fixed number that does not care how many clients it is spread across. Everything between 7 cents and the 50¢–$2 range on the call is that $2,391, plus payroll, divided by too few clients.
| 2 clients | 6 | 12 | 20 | 30 | |
|---|---|---|---|---|---|
| Cost per minute | $2.56 | $0.90 | $0.49 | $0.32 | $0.24 |
| Cost per client, per month | $1,229 | $432 | $233 | $153 | $113 |
| of which the tools are | $2.49 | $0.83 | $0.41 | $0.25 | $0.17 |
480 Josh-minutes per client per month — the measured average from four months of call records. Vendor cost per client is $33.75.
$2,391 ÷ ($597 − $33.75) = 4.2 clients. At $697 it is 3.6; at $797, 3.1. So every month past the fifth client is contribution — before anybody is paid.
Which means the whole distance between five clients and Ken's "thirty clients is profitability" is payroll. That is not a criticism of the number — it is the strongest evidence yet that the missing line is the only line that matters.
| Payroll + other overhead | Total fixed | $/min at 12 | $/min at 30 | Break-even at $597 |
|---|---|---|---|---|
| Nothing — tools only | $2,391 | $0.49 | $0.24 | 5 clients |
| $5,000 | $7,387 | $1.35 | $0.58 | 14 |
| $10,000 | $12,387 | $2.22 | $0.93 | 22 |
| $14,500 | $16,887 | $3.00 | $1.24 | 30 — Ken's own number |
He said thirty clients is where this turns profitable at $597. Run that backwards against the real tool bill and it implies about $14,500 a month of payroll and other overhead beyond these five tools.
If that is roughly right, the price is nearly built. If it is high, profitability arrives a lot earlier than thirty — at $5,000 of payroll it is fourteen clients, and at nothing but tools it is five.
Bucket A
ElevenLabs is 63% of a call — more than the line, the recording, the hearing and the brain put together. The cheaper voice would save about $116 a month across a twelve-client book, and Ken already overruled that on audio quality. That stands — a cheaper voice that sounds worse is a failed test, not a saving. It is recorded here only so nobody re-proposes it as a discovery.
Josh's brain costs 1.1¢ a minute with prompt caching working. Without it, the same brain costs 6.9¢ — six times — and nothing announces the change. One line in the server log says which state we are in, on every single call.
Bucket B
Ken asked what it costs to run Engine 3 per minute. Engines 2 and 3 have no minutes — they are per send, and the answer is cents.
| Line | What runs | $ / client / month |
|---|---|---|
| Engine 1 · Reputation | 5 emails per completed job, plus the booking text to the inspector's own mobile | $0.36 |
| Engine 2 · Agent | one touch a month to ~100 referring agents — flat, because agents churn | $0.07 |
| Engine 3 · Retention | one touch a month to every past client — ~480 at go-live, and it now grows forever | $0.32 |
| Content research | the monthly article pull — 8 web searches and one model read per theme | $0.14 |
| Drift monitoring | a model reads real calls against that client's own source of truth | $1.11 |
| Telnyx number rental | per client line | $1.00 |
| CRM record + transcript + report | work order, call transcript and report filed to the contact | $0.00 |
| Total | on top of the calls, at go-live | $3.00 |
Engines 2 and 3 together cost about 28 cents per client per month in AI and email. The bottleneck Ken named himself — "that's my problem in the red" — is a person approving each piece before it sends. If that review is fifteen minutes a month, human time is a hundred times the AI cost of the engine being reviewed.
"As long as they're in business, they should keep marketing those prior clients… clients will stay there forever." The retention list stops expiring at twelve months and starts compounding — and Ken split the two audiences himself: agents churn ("a handful will stay, the majority will go out of business"), so only the client list grows.
| Past clients | Emails / month | Email cost | Share of that client's call cost | |
|---|---|---|---|---|
| Go-live · a 12-month import | 480 | 580 | $0.39 | 1.3% |
| Year 1 | 935 | 1,035 | $0.70 | 2.3% |
| Year 3 | 1,798 | 1,898 | $1.28 | 4.2% |
| Year 5 | 2,600 | 2,700 | $1.82 | 5.9% |
| Year 10 | 4,371 | 4,471 | $3.02 | 9.8% |
Additions and unsubscribes balance at about 13,000 contacts — decades away at 40 jobs a month. The list does not run away.
"Will it be extra emails?" Yes — 39¢ a month at go-live, $3.02 at year ten.
"Will I have extra storage?" Effectively no. Contacts are not metered on the agency plan, Telnyx recording storage is $0 and recordings delete at twelve months, and transcripts are about 13 MB per client per year.
"What would be the terms?" The real question, and it isn't financial: unsubscribe has to work across every engine, cancellation has to stop the sends, and a retained contact is never texted.
A GoHighLevel sub-account climbs a daily sending ladder starting at 1,000 emails a day. A 2,700-name list fired in one batch clears that ceiling by year five. Our sends are 1-to-1, which HighLevel says keep delivering past the cap — worth one live test. Stagger the monthly send across a few days and the question never arises.
And the content does not cost more. A twelve-month sequence that runs forever does not need twelve new themes a year: the puller picks a current article against each month's theme at send time and remembers what has already gone out, so the same theme yields a different article every year. Ken's "every year we'd add a different group of stuff in there" already happens by itself, at no extra cost.
Going out to the internet is cheap, if it is on a leash. Anthropic's web search is $10 per 1,000 searches — a penny each, with the results costing three to five times that in tokens. Our content puller is capped at eight searches a pull, and its 44 pulls a year serve the whole client book for about $20. An agent with no cap is the way to lose money here; every one that touches the internet needs a limit written into the code before it is switched on.
Bucket C · now half real
These are the actual subscriptions, supplied 27 August. None of them grows when a client is added — which is why the per-minute number falls so steeply with client count.
| Tool | $ / month | Share | $/min at 12 clients | at 30 | What it is |
|---|---|---|---|---|---|
| Hamming AI | $1,200.00 | 50.3% | $0.208 | $0.083 | Call testing. Half the entire tool bill — and the one line here that is a dial rather than a subscription |
| GoHighLevel Agency Pro | $497.00 | 20.8% | $0.086 | $0.035 | The CRM. Flat at any client count — 50 sub-accounts cost the same as 2 |
| Claude — Dil | $400.00 | 16.8% | $0.069 | $0.028 | A seat, not the product. Development and agent work |
| Claude — Ken | $200.00 | 8.4% | $0.035 | $0.014 | Same — a person's seat |
| Rose Hosting | $89.99 | 3.8% | $0.016 | $0.006 | The server everything runs on. This closes the "server cost unknown" line |
| GitHub Pro | $4.00 | 0.2% | $0.0007 | $0.0003 | Added 2026-09-09 — $48 a year. Actions is $0 billable: 648 of 3,000 included minutes used in September |
| Total tools | $2,390.99 | 100% | $0.415 | $0.166 | Flat — it does not move when a client is added |
Dil: "we purchase it for credits." 2¢ a credit; a minute of test call is 15 credits — 30¢. $1,200 a month is roughly 4,000 test-call minutes if it is all used, and nothing if it is not.
So the honest way to carry it is not "$1,200 of fixed cost" but "how many test minutes are we buying this month?" — a number Ken sets. ⚠️ And the sixty-day export says it runs hot when nobody is watching: 106,245 credits, 69,165 of them in forty-eight hours. Two days were 65% of two months. The control is a credit budget and a balance alert.
None of which argues for dropping it — one missed inspection is $700–$800, and it is the gate in front of every booking.
"$497 per month even if we have 50 accounts of clients." Confirmed — a new client adds $0 of CRM.
💰 The billing screen shows monthly billing and offers two months free on annual — $4,970 instead of $5,964. That is $994 a year for one click, and it spends money, so it is Ken's or Beth's call.
⚠️ "Free AI ends August 31." Anything of GoHighLevel's own AI switched on becomes metered in four days. We don't depend on it — review replies are written on our server. Confirm nothing is activated rather than activating under a deadline.
💡 That menu also carries Domain Purchase — the microsites accelerant needs a registrar we don't have.
The Anthropic API bill for the product. The two Claude lines above are seats for Ken and Dil. Josh's brain, the drift reader, Parker, James and the content puller all run server-side on an ANTHROPIC_API_KEY — a separate metered account. At 30 clients Josh's brain alone is about $163 a month. It has to be arriving somewhere.
The usage vendors. ElevenLabs, Deepgram and Telnyx are per-minute, so they correctly sit in Bucket A rather than here — but their invoices are what confirm the 6.4¢. Pull one month of each and the whole model stops being a model.
One split that is easy to get wrong and expensive if you do: a vendor's plan fee is fixed, not per-minute. Only the usage half belongs in Bucket A. Dividing a plan fee by today's small minute count inflates the per-minute rate enormously — and would price us out of the market on a bookkeeping error.
The accelerants
Ken asked what an accelerant costs in AI, and named what was needed to answer it — "we need workflows for the accelerants… if he doesn't get the workflow, he's not gonna know how many turns." Every deliverable below is quoted from the live sales pages, and every step is costed three ways.
At $50 an hour that is $371 of labour against $3.63 of tokens — a hundred to one. AI is under 1% of what an accelerant costs to deliver. The question is not what the tokens cost. It is how many human minutes each promise takes, and how many of those a build can remove.
| Accelerant | AI / mo | Hard / mo | Human min | Cost / mo at $50/hr | What is missing before it can be delivered |
|---|---|---|---|---|---|
| Paid Ad Management | $0.15 | — | 100 | $83 | No Google Ads or Local Services integration exists. A person in Google's own screens until one is built 2–4 days |
| Converting Website | $0.05 | — | 20 | $17 | Page generator and publish pipeline. Setup is ~6 hrs and includes the domain cutover — the one step that can lose a business its rankings 1–2 weeks |
| GBP Management | $0.31 | $0 | 100 | $84 | No Business Profile API. The review reply is already built and better than GoHighLevel's — nothing can post it ~1 week |
| Local Microsites · 5 towns | $0.50 | $6.25 | 75 | $69 | No registrar, no multi-domain hosting. Scales worst by hand — twelve towns is three hours a month 1–2 weeks |
| Client & Agent Newsletter | $0.57 | $0.39 | 60 | $51 | Assembly, an approval screen, a send button. Everything upstream already exists 3–5 days |
| Trusted Home Pro Network | $2.00 | ? | 90 | $77+ | Not priceable yet — vendor ad budget, the 30-second vendor video, and card processing to take vendors' money 2–3 weeks |
| Authority Builder · one-off | — | ? | 150 once | — | Not priceable yet — the paid syndication fee is the entire economics and nobody has a quote 1–2 days |
| All seven, one client | $3.63 | $6.64 | 445 | $381 | AI is 0.9% of it · plus a one-off ~$31 of build tokens |
The rates are a vendor price list: $5/$25 per million tokens on Opus 5, a cache read at 10% of input, $10 per 1,000 web searches, $0.000675 an email. The quantities are my judgement — nothing has been measured, because no accelerant has ever been run.
🔴 Dil found a real error. A website and a microsite are agentic builds — an agent writing a page, checking it, revising it — and every turn pays for its context again. I costed them as single API calls. One build turn is 3.4¢ on Sonnet, 8.5¢ on Opus; a fourteen-page site is 80–200 turns. Converting Website: $1.90 → $3–$17. One microsite: $0.91 → $0.85–$5.10. The one-off across all seven: $7.65 → about $31. Four times what I published — and still 4% of the setup labour sitting beside it.
On searching the internet: the search fee is a penny; what the results cost in tokens is three to ten times more. A capped eight-search pull is $0.19–$0.95, so the working figure moves from $0.45 to $0.60 — and 44 pulls a year still serve the whole book for about $26. The real control is the cap, not the model: ours stops at eight. An uncapped research agent can spend dollars a run with nothing looking broken.
How to stop estimating: build one microsite for a real town and read the Anthropic console before and after — that single delta replaces the whole column with a fact. Permanently: every response already returns its own usage and nothing captures it. Half a day to log it per job.
Human effort is given in minutes on purpose — Ken has the loaded labour rate and I do not. One number turns every minute above into dollars. The $50 column is illustrative.
The live page promises it "at no charge… for as long as your subscription is active." That is about 100 minutes per client per month at zero revenue — roughly $2,500 a month at thirty clients, and it never converts to paid for those fifty.
Not an argument against the offer. An argument for building the Business Profile integration before the fifty fill up, and carrying that cost in the base price.
From the archive note: every specialist carried the same toolset — search the web, fetch a page, read and write repo files. None could post to a channel, log into a platform, read an analytics number, or run on a schedule.
That is exactly what an accelerant needs. Bringing an agent back does not deliver one. Building the integration does — and then the agent is optional.
Said once, then it is Ken's call: the two nearest to delivering are the Newsletter and Google Business Profile. The two furthest are the Home Pro Network and the Converting Website. If accelerants need to earn this quarter, that ordering is the cheapest route to it — and it is close to the opposite of the order they appear in on the site.
Not client cost, but real money
The two Claude seats on the tool list — $600 a month between Ken and Dil — are this bucket, and they are flat. What sits outside them is the metered API the product runs on. Q and the four specialists, the development sessions, and the three agents on the public website. Parker, James and web Josh cost about 17¢ a session — roughly $250 a month at 500 sessions, $2,500 at 5,000. Nothing records website traffic, so that is a range rather than a fact; GoHighLevel or site analytics closes it in minutes. The development seats are now a known $600 and are flat — which is the good news. The metered API account behind Josh, drift, Parker, James and the content puller is the part still unmeasured, and the Anthropic console shows it, split by key, in about two minutes.
The Parker and James chat endpoints are public, with no per-visitor rate limit, no captcha and no daily spend cap. Each request costs about 1.7¢; one request a second, sustained, is about $1,400 a day. Nobody has to be malicious — a badly written scraper would do it. Two cheap fixes: a per-IP limit, and a spend alert on the console so a surprise bill announces itself instead of arriving.
How to finish this
Pull eight invoices for one dated period — Telnyx, Deepgram, Anthropic, and the P&L (ElevenLabs, GoHighLevel, Hamming, the server and GitHub are now read) — split each into plan fee (fixed) and usage (per-minute), and divide usage by the minutes in that period. The 2,097 test minutes already burned through these vendors are a paid-for sample of exactly the workload being priced: here, the test traffic is a gift rather than a problem.
The permanent version is a small job and mostly already built: the bot posts a record for every call, and the voice pipeline is already producing token and character counts that nothing captures. Wire those into the call log and every call costs itself, forever. Worth doing once the price is set — not worth waiting for.