Fix the timing, and the rest follows. Payroll stops being a guess. New builds get priced with the payment gap in mind, not just the engineering hours.
Most AI agencies start a build with a clear scope and a fixed fee. The technical work follows a plan — data pipeline, fine-tuning passes, eval sets, deployment. What doesn't follow a plan is when the money shows up.
Agencies billing fixed-fee builds carry weeks of compute cost, contractor hours, and evaluation work before the first invoice clears, and longer before the rest of it does.
When the Work Outpaces the Payment
Most founders already know cash flow is a problem. What's missing is which lever to pull first. Generic advice on managing cash flow assumes a finance team running weekly reconciliations — most 12-person AI agencies don't have one, and don't need one to fix this.
Small business cash flow problems in an AI agency rarely come from a lack of demand. They come from project timing that doesn't match payroll timing. A simple cash flow forecast for an AI agency makes that timing visible before it turns into a payroll problem.
Fix the timing, and the rest follows. This is cash flow management for small business AI agencies — not treasury work for a venture-backed lab. Payroll stops being a guess. New builds get priced with the payment gap in mind, not just the engineering hours. Four levers move this.
1. Price the build with a deposit, not a promise
A deposit before kickoff covers discovery, data prep, and the first evaluation pass — the work that happens before the client has anything to show for it. Thirty to fifty percent up front is standard for fixed-fee builds. Newer clients or smaller engagements justify the higher end. Clients with a track record of paying on time can sit closer to thirty.
The deposit isn't a signal of distrust — it keeps the agency from financing someone else's model build out of its own compute budget. Compute and contractor costs are exactly the kind that go untracked until they quietly eat the margin on the four costs most builds miss.
2. Tie remaining payments to eval gates, not time elapsed
Splitting the rest of the fee fifty-fifty — half on delivery, half at the end — puts the riskiest phase of the build on credit. That includes final fine-tuning, deployment, and the eval pass that decides whether the model actually ships.
Break the remainder into three or four milestones instead, each tied to a gate: data pipeline complete, first model version evaluated, fine-tuning complete, deployment. A ten-week build should produce a payment every two to three weeks, not two payments total. This protects the agency from doing eighty percent of the technical work while holding only half the fee.
3. Structure retainers around maintenance, not goodwill
Once a model ships, most AI agencies move into an ongoing relationship: monitoring drift, retraining on new data, maintaining the prompt library, patching edge cases. This belongs on a retainer, not a loose "we'll keep an eye on it" arrangement.
Define it in deliverables — a fixed number of retraining cycles per quarter, a capped number of eval-set updates per month. Draw a hard line between retainer maintenance and a new build, and bill anything past that line as a change order. Without that line, hours get logged but margin still disappears the same way it does on undertracked builds.
"Can you just also fine-tune it on this new dataset we got? Shouldn't take long."
That's the sentence that turns a retainer into an unpaid build, one favour at a time. It's the moment to point back at the agreement, not the moment to say yes for free.
4. Treat overdue invoices as a weekly habit, not a quarterly panic
Compute costs and contractor invoices arrive whether or not the client has paid. Agencies that recover cash fastest review aging invoices every week, not at month-end, so they never chase the same client twice. Most overdue invoices stay overdue because of a missed reminder, not because the client is refusing to pay.
A short, consistent escalation beats an apologetic one: a reminder the day after the due date, a firmer follow-up at fifteen days, a pause-of-work conversation at thirty. Write the cadence down so it doesn't depend on whoever happens to notice the invoice first.
Agency OS keeps invoice status and overdue tracking in the same workspace as the build — checked weekly, not at month-end.
What This Actually Takes
None of these four strategies need a finance hire or an elaborate cash flow projection model. They need every build priced with the payment gap in mind, and invoice status reviewed often enough that an overdue payment gets caught in week one, not week six.
Business cash flow management for a fixed-fee AI agency comes down to that habit, repeated weekly.
