The EBITDA math nobody runs before buying AI

AI purchases at services firms usually start with a demo and end with a line item nobody can defend at year end. The missing step is the math. This note walks through the EBITDA arithmetic an owner should run before signing anything, and the two numbers that decide whether a system paid. Most owners can run it in an afternoon with numbers they already have.

Why do owners buy AI on time saved instead of profit?

Owners buy AI on time saved because time saved is the number vendors lead with and the easiest number to demo. The problem is that saved hours do not appear anywhere on a P&L. Unless the recovered time becomes billable work or new revenue, the saving is theoretical.

A tool that saves ten hours a week sounds compelling in a sales call. Twelve months later the owner cannot point to a single engagement, client, or margin point the tool produced. The hours were saved and then quietly absorbed by everything else.

Profit framing changes the purchase decision itself. A firm that prices AI on hours saved will buy the cheapest tool that demos well. A firm that prices AI on revenue created and margin kept will buy the system most likely to move the P&L, which is rarely the same product.

What is actually happening to professional-services margins?

Professional-services margins are compressing. Deltek’s industry study found EBITDA at professional-services firms fell from 15.4% to 9.8% over five years, a five-year low. Firms that moved early on AI are defending margin with lower growth and delivery costs. Firms that waited are absorbing the compression directly.

That gap is the real cost of waiting. A firm doing $3M at a 10% margin keeps $300K. The same firm at a 15% margin keeps $450K. The question for any AI purchase is which of those two firms it moves you toward, and by how much.

Margin pressure also explains why the AI question feels urgent to owners who ignored every previous tooling wave. Software that made delivery slightly smoother was optional. A capability that decides whether the firm keeps 10% or 15% of revenue is not.

How do you translate an agentic operator into profit terms?

Translate an agentic operator into two movements. Revenue created, meaning the engagements, reactivations, and referrals the system produced that would not have happened otherwise. And margin kept, meaning the owner and partner hours pulled out of non-billable growth work and returned to delivery or sales. Everything else is commentary.

• Revenue created. A reactivated client, a proposal that went out a week earlier, a follow-up that did not go quiet.
• Margin kept. Owner hours moved from drafting and chasing back to billable work and closing.
• Cost avoided. The coordinator or BD hire the firm did not make because the agentic operator absorbed the work.

Revenue created is the easier half to see. A dormant client who comes back after a reactivation note is visible in the bookkeeping. The discipline is recording why the engagement happened, so the quarterly review credits the system that produced it rather than general luck.

Attribution does not need to be perfect. It needs to be honest. A simple log of what the agentic operator produced, what the firm approved, and what closed is enough to run this math every quarter.

The margin side deserves equal weight. At most small and mid-sized firms, the owner is the most expensive person in the building and spends a meaningful share of the week on follow-up, drafting, and chasing. Moving ten of those hours back to billable work or closing is a margin event, not a convenience.

What is the two-number test?

The two-number test asks one question of any AI investment. Write down the revenue the system should create and the margin it should protect within a defined window. If the vendor cannot help you fill in those two numbers before the build starts, the purchase is a cost, not an investment.

The test disciplines the builder as much as the buyer. North Signal sizes every agent against the revenue and margin it should move, and defines the success criteria before the build starts. A vendor who will not put numbers next to the work is asking the firm to carry all the risk.

Notice what the test screens out. It screens out tools bought because a competitor mentioned them, pilots that run forever without a verdict, and dashboards that report activity instead of outcomes. What survives is a short list of systems with a number attached, which is the only list worth funding.

What should a firm measure in the first 90 days?

Measure four things in the first 90 days. Pipeline movement the agentic operator touched, dormant clients reactivated, proposals and follow-ups shipped against the prior baseline, and owner hours returned to billable work. Each one ties back to revenue created or margin kept, which keeps the review honest.

• Pipeline movement. Open conversations advanced, proposals out, and close rates against the prior quarter.
• Reactivation. Dormant relationships re-engaged and the revenue attached to them.
• Output shipped. Follow-ups, touchpoints, and proposals compared with the firm’s old baseline.
• Hours returned. Where the owner’s recovered time actually went, billable or otherwise.

Ninety days is long enough to be fair and short enough to force a verdict. A system that cannot show movement on at least one of the four measures in a quarter is unlikely to show it in a year, and the review meeting should say so plainly.

Running this math on your own firm takes one conversation. The Growth Audit Call walks through your pipeline, your margins, and where an agentic operator would move them, with the numbers on the table. If you would rather start with a self-serve read, the Client Loyalty Gap Audit at northsignal.studio/audit scores where revenue is leaking now.