You are not paying for AI. You are paying for the prompt you stuffed.

People ask what AI costs.

Wrong noun.

Models are cheap enough that the argument moved. The expensive part is everything you shove into the run: the handbook, the last twenty emails, the brand voice, the vendor list, the thing legal said in 2024. Context is the bill. A teammate that carries five departments in its head on every invoice chase will light money on fire with a smile.

One generalist is how you go broke. Not because it is "too smart." Because it is too loaded.

The stuffed prompt is the silent salary

You have seen this. Someone pastes the company into Custom Instructions, then asks for a reminder email. The model rereads the constitution of Northwind Studio to say "hi, invoice 4412 is late."

On a ChatGPT plan that looks like nothing. It is still a cap you chew through. On an API meter it looks like a line item that makes no sense until you open the trace.

I will not quote token prices. They change. The shape does not. Long briefs cost more than short ones. Repeating the brief on every run costs more than remembering the job.

Clerk does not need the content calendar. Mail does not need the warehouse route. Maker does not need Bell & Rowe's payment history. If they share a brain and a prompt, they all pay for all of it, every time.

Specialists plus a short brief is not a management fad. It is how you keep the bill boring.

$99 versus BYO versus a meter

Three ways to pay for the brain.

OSS in the $99. Models on our pipe, a computer, a board. Fine for a lot of recurring work. Assign a cheap one to the rote lane and a stronger one to the thing that needs judgment-shaped drafting.

Bring your own. ChatGPT, Claude, Grok, the plan already sitting in the browser. Work runs there. We do not mark up the task. Plus is about $20. Pro and Team are higher. Still a plan. Flat versus a meter is its own argument.

API. Right when you resell usage, spike, or need a model the seat does not offer. Wrong as the default for a teammate that runs every weekday while you sleep. Agents blow human usage intuition. That is not a moral failing. That is physics.

All-in for most shops is Octopus plus the plan they already have. $99 plus ~$20 is still a rounding error next to a junior ops seat. BLS still has benefits around 30.1% of compensation. A $4,000 salary loads to about $5,200-$5,700 a month. Three years of $99 is $3,564. Three years of that hire is roughly $180k-$206k. The teammate still cannot do judgment, relationships, showing up, or taste. I will keep saying that until it sticks.

Per-run visibility or you are guessing

If you cannot see Tuesday, you cannot manage cost.

The board should answer, without a treasure hunt:

• Who ran
• Which job
• Which model
• What they touched
• What waited for a human
• Whether the brief was fat

That is not analytics theater. That is how you notice Scout is rereading the same ten pages because nobody wrote "stop after the summary."

Approvals keep the expensive mistakes in Needs you. The vault keeps keys out of the thread so you are not paying to re-paste secrets. Own email means the follow-up lands with the teammate that started it, instead of a shared inbox that gets pasted back into context next week.

Dedicated workspace. We never train on your data. Seven-day money-back. Dedicated VM for agencies: talk to us.

Microsoft's 2025 Work Trend Index: 82% of leaders expect digital labor to stretch capacity soon. Anthropic 2026: 57% already on multi-stage workflows, 16% crossing teams. The 16% are the ones who learned that "one agent for the company" is a cost center with a name.

How a cheap team stays cheap

Write jobs like you would write them for a person who is good, not omniscient.

Clerk: unpaid over 14 days, weekday 9, draft, stop before send. Do not attach the handbook.

Mail: sort, draft the routine four, flag the supplier. Do not attach last year's brand offsite.

Maker: calendar, then wait. Do not attach payroll.

Watch: the thing that broke last month. Short.

Chief: route. Chief should not do the work. Chief should not eat the other briefs.

When two jobs need the same fact, put the fact in one place they can both read. Do not paste it into both prompts "just in case." Just in case is how a $99 plan feels tight in week three.

Stack accounts when a plan is out of room. Fallback mid-task so you do not rerun the fat context from scratch on a new model because the first one timed out. Reruns are a tax you levy on yourself.

The product is not a cheaper token. The product is a desk where a short job stays short.

Compare that to the first ops hire, not to a chat window.