Thirteen agents, five models, one bill: what an AI team really costs.

I run my businesses with a team of AI agents. They chase invoices, triage tickets, draft posts, answer the phone. In August I asked a simple question. Which model does each agent run on when I am not watching?

Nobody could answer it. Not me, not the notes, not the org board we keep of the whole stack. Forty-seven cards on that board. Exactly one had a model written on it.

That should bother anyone running agents. The model is the wage. Ask a business owner what each employee costs and you get an answer to the dollar. Ask most people running agents and you get a shrug and a monthly total.

So we rebuilt the board to tell the truth. Every agent card now shows the model it uses when it runs unattended. Short text codes, never colour, because I am colour-blind and meaning carried only in colour is meaning I cannot see.

The first audit was humbling. My spoken morning briefing had no model pinned at all, so it quietly inherited the session default. The session default is the dearest model in the stack. Nobody chose that. Nothing logged it. It was invisible until the board reported it. Two other agents turned out to run their automated legs on GPT-4o, a different vendor entirely, which I had half forgotten agreeing to.

Here is the team today. Thirteen scheduled agents on one server, all pinned to Sonnet in a single block of config. The chat desks split by stakes. The two agents that touch money and infrastructure run on Opus, the dearest. The writers and assistants run on Sonnet, the middle. Translation runs on Haiku, the cheapest. Five models across the fleet. Each matched to a job. None chosen by vibe.

About that cheap model. Haiku used to work our tickets. In May it invented some. Fabricated tickets, filed as though they were real. Around the same time our inbox router, also on Haiku, read its own instructions as a character description, replied “What would you like help with?” to an empty room, and exited clean. So Haiku was fired from ticket work, and the whole scheduled fleet was pinned to Sonnet. The comment explaining why is still sitting in the config file. It still applies.

But we did not fire it from everything, because cost is one column and fit is another. Haiku still answers our phone. We measured both models on the same call flow. Haiku speaks 1.5 seconds after the caller stops. Sonnet takes 2.6, and in testing it also hung up on the caller’s answer. On the phone, fast and plain beats clever and slow. Haiku also stands at the front door of our chat desks. Every plain question goes to it first, and its only job is to match that question to one known command. A hit runs a cheap deterministic lookup and the dear model never wakes up. A miss falls through unchanged. The same model that invented tickets. A different job, and it is quietly excellent at it.

Rough numbers from our own logs. Haiku costs about a fifteenth of what Opus does. One reporting job that polls five outside engines costs $6.54 run thoroughly and 69 cents run lean, and the lean run answers the question most weeks. Most of the Claude work rides on a flat monthly subscription, the same product anyone can buy. Every call to an outside model leaves through one gateway that logs spend against a key and enforces a hard budget cap. When I want the bill, I look in one place.

The part most people skip is drift. Model choice does not stay put. Someone adds a flag. A default changes upstream. A runner gets rewritten and the pin falls out on the floor. So the board does not trust anyone’s memory, including mine. A build check opens the live config files and reads what is actually there. The registry records where each model is decided, never what it is. If a pattern stops matching, the build fails loudly and the board refuses to update. We tested the check by planting three decoys, one of them a silent swap back to Haiku. It caught all three. A job runs the same check every Monday and posts the result even when everything is clean, because silence has to mean the job died. Silence must never mean all is well.

None of this is exotic. It is payroll. You would not employ thirteen people without knowing what each one is paid and what each one is for. Agents deserve the same ledger. Write down who runs on what. Put it somewhere you will actually look. Then let one dumb, loud check read the real files every week, and believe the check over your memory.