Skip to content

Guide · Agent costs

How to bill clients for AI agent costs

Your coding agents now do part of the work you bill for. Here is how to record what they cost per client, pick a pricing model, set a markup and put it on the invoice without surprising anyone.

By the eBRBe team · Published

Why AI costs need their own line

When Claude Code, Cursor, Codex or a ChatGPT agent writes part of a deliverable, two things change. The hours go down, and a new cost appears: tokens, billed by the model provider per million tokens in and out. If you only bill hours, the faster you work, the less you earn, and the token bill comes out of your margin.

The industry hasn't settled how to handle this. Digiday's reporting describes agencies weighing whether to absorb AI costs or pass them on, with no single answer yet: some meter token costs to clients case by case, others fold them into technology or subscription fees, and others absorb them. Omnicom's leadership has said it wants these costs to be part of the client conversation rather than absorbed quietly (Digiday).

Whatever you decide, you can only decide it well if you know what each client's agent work actually cost.

Step 1: record the cost per client

For every agent session or run, keep:

  • Who and what: the client and project it was for, the person who ran it, and the tool (Claude Code, Cursor, Codex, your own agent).
  • Usage: the model, input and output tokens, and cache reads and writes, which many providers price differently.
  • Cost: worked out from the model's list price, or the exact amount when your provider reports it.
  • Billable or not: internal experiments and rework you choose to absorb shouldn't land on a client invoice.

Reconcile with the provider's invoice now and then: batch discounts, negotiated rates and regional pricing mean list price and your real bill can differ. Flat-rate seats (a ChatGPT or Claude subscription) have no per-token bill at all; decide up front whether those are overhead or a cost you recover.

A worked example with list prices: one Claude Opus 5.5 run with 120,000 input and 8,000 output tokens, at $4 and $20 per million tokens, costs $0.64. A hundred such runs in a month is $64.00 for that client. Try your own numbers in the AI Agent Cost & Markup Calculator.

Step 2: choose a pricing model

Ways to bill AI agent costs
ModelHow it worksGood when
Pass-through at costThe client pays exactly what the tokens cost, on its own line.You want maximum transparency, or usage is small and you'd rather not argue about margin.
Cost plus markupCost plus a percentage (or a fixed fee) for choosing, running and checking the agents.Agent work is a real part of delivery and you carry the setup, review and the risk of failed runs.
Blended into the rate or a fixed feeNo separate AI line: your hourly rate, project fee or retainer already covers expected AI use.Usage is predictable, and the client buys an outcome rather than inputs.
AbsorbedYou treat AI as overhead, like software licences.Costs are tiny compared with the fee, or it's a competitive edge you don't want to price.

All four are in use. In Digiday's look at agency AI bills and its reporting on token economics, examples range from an agency passing token costs to clients on a metered, case-by-case basis, to one building tokens into its tech-and-subscription pricing, to agencies that put AI work on flat fixed fees, include token costs at cost without a markup, or absorb them. Pricing doesn't have to be the same for every client either: a metered line suits a client running heavy agent workloads, a blended rate suits a small website retainer.

Even if you blend, keep recording the cost. It tells you whether a fixed fee still covers the work, and it's your evidence if a client asks what the AI part of the job cost.

Step 3: set a markup you can explain

There is no industry benchmark for marking up tokens yet; Digiday notes that clients and procurement teams have little to compare against. Ranges you'll see quoted come from consultancies rather than market data: for example, blckalpaca describes a 30–50% margin on passed-through, tracked costs as standard practice for agencies re-billing agent infrastructure. Treat any such figure as one firm's view, and base yours on what you actually do around the runs: prompt and tool setup, reviewing output, re-running failures, and paying the bill before the client pays you.

Watch the difference between markup and margin. A markup is added to the cost; the margin is what you keep as a share of the price, so it is always smaller (AgencyAnalytics explains the same difference for ad-spend markups):

Markup on cost and the margin it gives
Markup on costMargin on the price
10%9.1%
20%16.7%
30%23.1%
50%33.3%
100%50%

Continuing the example: $64.00 of agent cost with a 30% markup bills $83.20, and you keep $19.20.

Step 4: show it on the invoice

  • One line per project and model, next to the hours for the same work, e.g. “AI agent runs · Claude Opus 5.5 · 42 runs”.
  • Show cost and markup if you agreed to pass-through, so the client sees exactly what they pay for. With a blended rate, there is no AI line to show.
  • Bill only what's billable: leave out runs you marked as internal, and runs already on an earlier invoice.
  • Mind the currency. Providers bill in US dollars; if you invoice in another currency, say which exchange rate you used.

Step 5: put it in the contract

A short clause avoids the awkward conversation later. Adapt something like this with your own lawyer:

“We use AI tools in delivery. Model usage for your project is billed monthly at the provider's list price plus [X]%, itemised on the invoice. We'll ask before monthly AI costs exceed [amount].”

A monthly cap or an alert threshold keeps a runaway agent loop from becoming a billing dispute.

Checklist

  1. Every agent run is recorded against a client and project, with tokens and cost.
  2. You know which runs are billable, and you reconcile against the provider's invoice.
  3. Each client has a pricing model: pass-through, cost-plus, blended or absorbed.
  4. Your markup is written down, and you know the margin it gives.
  5. The invoice shows AI costs the way the contract says it will.
  6. There's a cap or an alert for unusual months.

How eBRBe does it

The Claude Code hook in the ebrbe CLI, MCP, the REST API or a CSV import log each run against a client, priced from list prices, and the Free plan shows the spend per client. On Pro, markup rules per workspace, client or project price those runs on the invoice next to the human hours, and agent cost reports show cost by project, model and tool.

Sources

Checked on 29 October 2026. Quotes and practices are as reported by each source.

See what AI costs per client, free

Free for up to 3 people, with the AI cost ledger included and a 14-day Pro trial. No card needed.