Most software bundles AI into a per-seat uplift: everyone pays a bit more, nobody knows what it cost, and there is no way to answer the one question an agency actually has — what did this client's work cost us in AI this month.
That question matters here more than in most businesses, because an agency already tracks cost per client. AI is a cost per client. If your tool cannot attribute it, you have introduced an expense you cannot put on the account it belongs to.
Why does per-seat AI pricing not work for an agency?
Because it charges you for people rather than for work, and an agency's unit is work. A five-person shop pays the same uplift whether it ran two AI tasks or two hundred, and whether those tasks were for the client on a healthy retainer or the one already over-serviced.
It also makes the cost invisible in the wrong direction. A per-seat uplift looks small on the invoice and unbounded in use, so nobody thinks about it — which is fine until AI becomes a real line item and you cannot say where it went.
Metered pricing has the opposite shape: slightly more to think about, and an answer when you need one.
How is the cost of a run actually calculated?
From tokens and tool calls, not from a seat count. Each run computes its own cost — the model's token usage, plus the tool calls the agent made, with a markup — and stores a breakdown against that run.
The practical consequence is that cost lands on the piece of work rather than on the month. A status roll-up on one client is a number. A weekly report across twelve clients is twelve numbers. That is the granularity an agency needs to make AI a billable or at least an attributable cost, rather than overhead.
If you already track hours per client, AI cost per client is the same kind of fact. Most tools cannot give it to you.
Can you use your own OpenAI or Anthropic key?
Yes, and this is the part that changes the maths for agencies with real volume. You can add your own provider key in settings, and runs route through it — you pay the provider directly, at provider prices, with no markup.
| Pooled credit | Your own key | |
|---|---|---|
| Who you pay | Us | The provider |
| Markup | Yes | None |
| Setup | Nothing | Paste a key |
| Cost visibility | Full breakdown | Full breakdown, still |
| Best when | Usage is light or spiky | Volume is steady and material |
The detail worth knowing: even on your own key, the run still records what it would have cost. So switching to a BYO key does not blind you — you keep the per-run breakdown, you just stop paying a markup for it.
What stops a runaway bill?
Three controls, and they are worth setting before you need them rather than after.
- Per-agent caps. A budget on a specific agent, so one misconfigured job cannot spend the month.
- Auto top-up. Optional. Off means you hit zero and stop; on means you keep running. Both are defensible — pick deliberately rather than by default.
- Credit purchase. You buy in advance rather than discovering the bill.
The failure mode metered pricing is accused of — a surprise invoice — is really a failure to set a cap. An agency that would never let a freelancer bill unlimited hours should apply the same instinct here.
What should you actually expect to spend?
Less than the anxiety suggests, and the honest answer is that it depends entirely on what you delegate. A status roll-up reading a handful of rows is cheap. An agent reasoning across a large sheet and calling several tools is not cheap in the same way, though still small against the hour it replaces.
The right way to find out is not a pricing calculator. Run the job you actually want automated, on one client, and read the cost on that run. You will have a real number in an afternoon, and it will be more useful than any estimate — including ours.
What you should not do is estimate from a per-seat comparison. The two models are not comparable, and the tool that looks cheaper per seat is not necessarily cheaper per job.
How does this compare with the alternatives?
| Per-seat AI uplift | Metered, pooled | Metered, own key | |
|---|---|---|---|
| Cost per client answerable | No | Yes | Yes |
| Predictable monthly figure | Yes | With a cap | With a cap |
| Cheapest at high volume | No | No | Yes |
| Cheapest at low volume | No | Yes | Roughly equal |
| Requires any setup | No | No | A key |
The per-seat column is not a straw man — predictability is a real virtue, and if you want one number that never moves, it wins. It just cannot answer the question this article opened with.
For what an agent does with those tokens, see what an AI teammate actually does in an agency, and for connecting an outside assistant, connect your workspace to ChatGPT or Claude.
Where should you start?
Set a per-agent cap before you delegate anything. Then run one real job — the weekly status roll-up is the usual candidate — and look at what that single run cost.
If the number is small and the output is right, you have a repeatable unit you can price into a retainer. If the number surprises you, you found that out for the price of one run rather than one quarter.
Set a cap, run one real job, read the cost on that run. One afternoon beats any pricing estimate, including ours. Start free