AI Automation Agencies: When You Need a Controller, Not a CFO

Agency A bills clients flat monthly retainers and pays a predictable amount for compute each month, because its workflows run on fixed volumes. Agency B bills on a mix of project fees and outcome-based pricing, and its compute cost scales directly with how much a client's workflow actually runs that month, sometimes doubling with no warning. Agency A can get by with lighter financial oversight because its numbers are stable and easy to forecast. Agency B needs someone watching margin by client, weekly if needed, because a single client running an inference-heavy workflow can quietly erase the profit on three other accounts.

Treating API and compute spend as one undifferentiated software expense line hides the real story, because the cost of serving Client A might be ten times the cost of serving Client B even at similar contract value. A controller sets up cost tracking that ties compute spend back to the client or workflow that generated it, so gross margin per account becomes visible instead of guessed at. Without that discipline, an agency can be growing revenue every quarter while margin quietly erodes, because the newest, most complex clients are also the most expensive to serve and nobody has connected those two facts.

Once margin visibility exists, the fractional CFO question becomes about pricing strategy: should outcome-based contracts include a cost-of-compute clause that adjusts pricing if usage runs far above the estimate, should certain workflow types be priced differently because they are structurally more expensive to run, and how much cash cushion does the agency need given that its costs can spike faster than its revenue can respond. A fractional CFO also owns the model of how many new client engagements the team can take on before headcount or compute budget becomes the constraint, which is a different question from whether last month's numbers reconciled.

A controller in this business closes the books monthly, allocates compute and API spend to the client or project that drove it, reconciles contractor and subcontractor payments if the agency uses outside specialists for overflow work, and flags any client whose margin has slipped below a set threshold before it becomes a quarterly surprise. Sales and marketing spend for an agency at this stage often runs near 37% of revenue at growth-focused, VC-style companies, though a bootstrapped automation shop usually runs leaner than that¹. A controller tracking spend against a realistic benchmark, rather than an inherited SaaS-company number, is doing the job right.

A monthly controller checklist for an automation agency looks like this:

• Close the books each month and allocate compute and API spend to the client or project that drove it.
• Reconcile contractor and subcontractor payments when outside specialists handle overflow work.
• Review gross margin per client, since a single inference-heavy workflow can quietly erase the profit on several other accounts.
• Flag any client whose margin slips below a set threshold before it becomes a quarterly surprise.

A full-time controller carries a real cost: the national median salary for accountants and auditors sits at $83,680, before benefits and payroll taxes². For an agency still under a dozen clients, a fractional CFO who also builds simple per-client margin tracking can often cover the gap without that fixed hire. Once the client roster and the variance in compute cost per client both grow, the case for a dedicated controller who owns that allocation daily becomes hard to argue against, because a fractional CFO reviewing the numbers monthly will always be a step behind a margin problem that can show up mid-month.

Say two clients each pay the agency the same monthly retainer for what looks like a similar automation workflow. Client A's workflow processes a steady, predictable volume, so its compute cost stays flat and the account runs a healthy margin every month. Client B's workflow scales with that client's own customer volume, and after a strong sales month for Client B, the agency's compute bill for serving them roughly doubles even though the invoice to Client B did not change. On a blended, company-wide basis this can look like a normal month. Once compute cost is tagged to each client, the picture is obvious: Client A is effectively subsidizing Client B, and either the contract needs a usage clause or the account needs repricing before it happens again. This is the exact scenario that pushes an agency from casual bookkeeping toward a controller who tracks cost at the client level as routine practice, not a special project.