Revenue gets logged everywhere. Cost gets logged nowhere consistent. Which means nobody can say, with any confidence, which clients are actually profitable and which ones are being quietly subsidised by the rest of the agency.
Most AI agencies can tell you exactly what a client paid. Fewer can tell you what that client actually cost to serve.
A chatbot build that closed at a healthy fee can quietly bleed margin through inference costs nobody priced into the quote. A fine-tuning pass that ran three times longer than scoped does the same thing. The revenue line looks fine. The cost line was never built.
The problem with tracking revenue alone
Most AI agencies run some kind of project tracking software — a board showing what's in progress, what's blocked, what shipped. That tells you about delivery. It tells you almost nothing about cost.
By the time the invoice goes out, the team has already moved to the next build. Nobody goes back to check whether the eval sets took twice as long as estimated. Nobody checks whether the RAG pipeline needed a second vector database instance that never made it into the original scope.
Revenue gets logged everywhere — contracts, invoices, the deal the founder closed. Cost gets logged nowhere consistent. It's the same gap behind agencies that log hours but still lose margin — the hours exist, they just never connect to a cost figure. Which means nobody can say, with any confidence, which clients are actually profitable and which ones are being quietly subsidised by the rest of the agency.
What changes once you can see the cost side
Pricing gets sharper. A retainer that looked profitable on paper but consumed twice the inference budget of a comparable client becomes visible, and the next quote accounts for it.
Project type decisions get easier, too. If every RAG deployment runs thinner margins than every standalone chatbot build, that's worth knowing before the pipeline fills with more of them.
And scope creep stops happening silently. A client asking for "one more retraining pass" stops being a favour and starts being a line item someone can actually see.
The four costs most AI agencies don't track
Time. Not the hours estimated at the proposal stage — the hours actually logged against the build, including the data cleaning and debugging nobody quotes for upfront.
Contractor fees. Freelance ML engineers or specialists brought in for a specific deployment. Easy to log as a general expense, hard to trace back to the project that needed them.
Tools and infrastructure. Model API usage, inference cost, cloud compute, and any evaluation tooling licensed specifically for that client's build. This is the category AI agencies miss most, because it scales with usage instead of sitting still like a flat subscription fee.
Overhead. Project management time, internal review cycles, account management — work that never shows up as billable hours but still draws down the project's margin.
The simple model
Project margin is revenue minus those four costs, calculated per project and per client — not once a quarter across the whole agency.
Project Revenue − Time cost (hours × fully-loaded rate) − Contractor fees − Tool & infrastructure cost − Overhead allocation = Project Margin
Run this per client and per project type. A multi-platform RAG deployment and a single chatbot build carry very different cost profiles, and averaging them together hides exactly what you're trying to find.
Agency OS links project records directly to cost and invoice data, so margin is visible per client and per project type without a separate tracker to update by hand.
Where this usually breaks down
Most agencies have an invoice tracker that shows what's billed, a task management tool that shows what's done, and a project management dashboard that shows status. Three systems, three different pictures, and none of them showing cost.
Project tracking alone won't surface this. Cost has to live as its own layer, connected to the same project record, or it stays invisible until the quarter-end review.
For a small AI agency, this is also where cash flow management gets harder than it needs to be. You can't tell which retainers fund the business and which are quietly draining it.
What to do this week
Pick your three highest-revenue clients. For each, total the four cost categories against their last completed project. You'll likely find at least one where the margin was thinner than the invoice suggested.
Once that's visible, the next quote — and the next client you take on — gets priced on what the work actually costs, not what it looked like it should cost.
