AI and Agency Profitability: Ten Questions, in the Order That Matters

By Craig S. Cody, CPA, Certified Tax Coach.

Here's the answer up front. There are ten questions that decide whether AI ever reaches your bottom line, and the order you answer them in matters more than the answers themselves. Most agency owners start at question three, "is this paying off," without having answered one and two, "what is it costing us" and "are we giving the savings away in how we price." That's a ratio with a broken denominator. You can't measure a return against a cost you never totaled.

So this is the map. Ten questions, four stages, and the reason each stage has to come before the next one.

I work with more than 70 marketing and advertising agency owners every month. The AI conversation comes up in nearly all of them now, and it almost always starts in the middle.

The Ten Questions at a Glance

Read this list first and find yourself on it. Wherever you can't answer a question, that's your starting point, not the one further down that feels more urgent.

Stage one, what it costs. The denominator. Nothing after this works without it.

1. What does your AI actually cost, all in?
2. Is your pricing model handing the savings straight to the client?

Stage two, what it returns. Three different audiences: your P&L, your team, your clients.

1. What return are you actually measuring?
2. Where did the freed-up capacity go?
3. Did the client ever feel it?

Stage three, what it does unsupervised. Three questions that sound like one and aren't.

1. What can it spend, and on whose authority?
2. What data can it touch, and who can get in?
3. What happens when nobody's watching?

Stage four, what it costs you to leave. The two nobody asks, and the two that show up in what your agency is worth.

1. Can you get out?
2. What should never have left the building in the first place?

Ten questions. Four stages. If you only ever answer one, make it the first one.

Why Does the Order of These Questions Matter?

Because each stage produces the input the next one needs.

Stage one gives you a cost number. Without it, every efficiency claim in stage two is a feeling. Stage two tells you whether the tool earned its place, which is what tells you whether it's worth governing in stage three. And stage three tells you how deep the tool is wired into your business, which is exactly what stage four asks you to unwind.

Run them out of order and you get busy work. I've watched owners build a careful policy for a tool that turned out to be costing them more than it returned, which is a well-governed waste of money. I've also watched owners measure a beautiful efficiency gain on a workflow they should never have moved outside the agency in the first place.

The windshield beats the rearview mirror here, same as it does everywhere else in your numbers. These questions are meant to be answered before the money moves, not reconstructed in March.

One more thing before the list. Adoption isn't the story anymore. A UK Treasury select committee found in January that more than 75% of City companies already use AI. Everybody has the tools. What separates agencies now is whether anybody can say what the tools did to the money.

Stage One: What Is AI Actually Costing You?

You can't run a return calculation without this. It's first for a boring reason: it's the denominator.

Question 1: What Does Your AI Actually Cost?

Most owners picture the subscription. That's the smallest part of it. The real bill includes prompt testing, the senior human hours spent checking output, rework when the output is confidently wrong, training time, and the tools nobody remembers buying.

The hardest line to see is the one that vanishes from the timesheet. Hours disappear from the record without disappearing from the business.

The fix is to stop treating AI as free magic and start treating it as a cost center with a name and a number attached. That's the whole argument in AI Is Not Free Labor: Here's the Bill Hiding in Your Agency's Margin, which is where to go for the full cost list and the six questions that build it.

Question 2: Is Your Pricing Model Handing the Savings to the Client?

This is the one that quietly undoes everything else.

If you bill by the hour and AI cuts the hours, you just gave your client a discount you never agreed to. The work got better and faster, and your invoice got smaller. That isn't an AI problem. It's a pricing problem that AI exposed.

Agencies that price on outcome and value keep the savings. Agencies that price on time hand them over automatically. Answer this before you celebrate a single efficiency gain, because an efficiency you can't keep isn't an efficiency, it's a rate cut.

Stage Two: What Is It Giving Back?

Now you have a cost. These three questions turn it into a return, and each one measures a different audience: your P&L, your team, and your clients.

Question 3: What Return Are You Actually Measuring?

Faster isn't a return. Faster is a feeling until it turns into one of four things: more revenue, higher margin, shorter delivery time, or a better client outcome.

Roughly 61% of agencies now use generative AI, against something closer to 17% of in-house client teams. That gap gets quoted as proof agencies are winning. It isn't. It's a measure of adoption, not return, and those are different words for a reason.

There's a specific number that answers this properly, and rather than rebuild it here I'll point you at it: Your Agency's AI ROI Isn't Faster Work. It's Higher AGI Per Employee lays out the five numbers that tell you whether AI reached the bottom line.

Question 4: Where Did the Capacity Actually Go?

The layoff headlines don't fit a lean agency. Big companies cut bloat with AI. A 10 to 50 person shop doesn't have bloat, so cutting heads cuts muscle.

The right question isn't how many humans you can remove. It's what more your existing humans can carry now, and whether that extra capacity got sold or just absorbed. Capacity that nobody sold is capacity that shows up as a quieter afternoon and the same monthly bill.

AI Isn't Coming for Your Agency's Jobs. It's Coming for Its Ceiling is the full version, including the discipline that stops "capacity" from becoming a word people say in meetings.

Question 5: Did the Client Ever Feel It?

Every measure in question three and four is inward facing. This one looks the other way.

If AI genuinely improved your work, something should show up in what clients do, not just in what your team reports. Did they renew. Did they expand the scope. Did they stop pushing back on price. Those are verdicts. Hours saved is testimony.

An agency can look efficient on every internal dashboard and still be losing clients, and if that's happening, the internal dashboards are measuring the wrong thing.

Stage Three: What Can It Do Without a Human in the Room?

Once a tool has earned its place, it usually gets more rope. This stage is about how much rope, and it's three different questions that sound like one.

The clean way to keep them apart: question six asks what it may spend, question seven asks who can get in, question eight asks what happens when nobody's watching.

Question 6: What Can It Spend, and on Whose Authority?

The risk here isn't fraud. Fraud sets off alarms. The risk is spend that goes through exactly as designed, because the permissions were too wide and nobody wrote down what "too wide" meant.

For an agency this bites hardest on client media budgets, where you're moving other humans' money through your own accounts. "The tool decided" isn't an answer that survives a client asking why 40% of the month went somewhere they never approved.

When AI Spends Your Agency's Money, Who Approved It? has the one-page sheet that settles this before the tool goes live.

Question 7: What Data Can It Touch, and Who Can Get In?

Different question, different failure. This one is about what the software can see and who can reach it, not what it can buy.

Write down which tools are approved, what data may go into them, who has access, what has to be reviewed by a human, what gets logged, and who owns an incident when one happens. A rule your systems enforce is a control. A rule in a document nobody opens is a hope.

Half of leaders report a formal AI risk program. Barely a third of the people actually doing the work agree one exists. A program the floor can't see isn't governing anything.

Question 8: What Happens When Nobody's Watching?

Increasingly the software doesn't stop when it hits something it doesn't understand. It retries. That's a design choice made by the people who built it, and it's precisely why you need to write down when it must stop instead.

Four things need an answer: what does it do when it isn't sure, what gets queued for a human rather than decided, who's awake and how fast do they have to respond, and who can shut it off and from where.

The bill for a bad decision made at 2am doesn't arrive at 2am. It arrives in a client conversation six weeks later, or as a line on your P&L you can't explain.

Stage Four: Can You Get Out, and What Should Never Have Moved?

The last two questions are the ones almost nobody asks, and they're the two that show up in what your agency is worth.

Question 9: Can You Leave?

Question seven asked who can get in. This one asks whether you can get out.

Your dependency is probably sitting in five places: the model provider, cloud storage, the CRM, the measurement stack, and the automation layer quietly stitching the other four together. The question isn't whether those vendors are good. Most are excellent, which is why you bought them.

The question is what happens if one changes price, policy, access, or availability. Only one of those four looks like an emergency, and it's the only one most agencies have a plan for.

This is also the leg that stops being a margin argument and becomes a valuation argument. A workflow that exists only inside a vendor account you can't move is a diligence finding, not an asset. Your AI Stack Has a Single Point of Failure covers the six questions to answer per critical workflow.

Question 10: What Should Never Have Left the Building?

The last question isn't about software at all.

Before you automate or outsource anything, decide what stays. Judgment, client context, and the relationships that make you hard to replace belong inside the agency, and if you move them for a cheaper unit cost you've made the agency cheaper and less valuable at the same time.

Cost reduction is a one-time subtraction. Differentiation is what gets multiplied when somebody buys you. Outsourcing Is a Strategy Only If You Know What Not to Outsource works through how to decide.

Where Most Agencies Actually Are

I'll be honest about what I see rather than dress it up as a statistic.

Across the agency P&Ls that cross my desk, the pattern is consistent: owners are somewhere in stage two, trying to prove a return, with stage one unanswered. They can tell me which tools they use. They can't tell me the total, and without the total there's nothing to divide into.

The second most common spot is stage three without stage two, which is a policy written for a tool nobody checked the value of.

Almost nobody is in stage four. That's the one that costs the most when it's skipped, because it doesn't show up as a bad month. It shows up as a lower multiple on the day you sell.

Here's how to place yourself in about a minute.

Can you say, out loud, what your agency spent on AI last month, including the review hours? If no, you're at question one, whatever else you've built. Go there first.

Can you name one number a specific tool was bought to move, and say whether it moved? If no, you're at question three, and the honest answer is that you own a subscription rather than an investment.

Can any tool in your business touch a payment method, a client's data, or a live campaign without a human clicking approve? If yes, you're in stage three whether you've addressed it or not. The tool didn't wait for your policy.

And the stage four test, which takes the longest to answer and is worth the most: if your single most important AI-assisted workflow went away on Monday, how many days until you're delivering normally again? If the answer is a shrug, that's your number, and it's the one a buyer will eventually ask for.

None of this requires new software. It requires somebody sitting down with the actual numbers for an afternoon, which is why it keeps not happening.

Who Should Ignore This Article?

If AI in your shop is one person using it to draft first passes, and nothing you own can touch a payment method or a client's data, you're fine. Bookmark this and come back when that changes.

If you're looking for someone to tell you AI is magic and you should buy three more tools, I'm not your guy.

And if you want to be told AI is a fad, I'm not that either. The efficiency is real. Whether any of it reaches your bottom line is a separate question, and it's the one this whole list exists to answer.

Frequently Asked Questions

Does AI actually improve agency profitability?

It can, but adoption doesn't guarantee it. Roughly 61% of agencies use generative AI while far fewer can state what it returned, so the widely quoted adoption figures measure uptake rather than profit. AI improves profitability when the saved capacity converts into revenue, margin, faster delivery, or a better client outcome, and when your pricing model lets you keep the savings instead of passing them to the client as a smaller invoice.

What order should an agency answer its AI questions in?

Four stages. First establish what AI costs you in total, including senior review time and rework, and confirm your pricing model doesn't hand the savings away. Second, measure the return internally and check whether clients felt it. Third, decide what the software may spend, what data it may touch, and what happens when nobody is supervising. Fourth, confirm you could leave the vendor, and confirm the work you moved was work that was allowed to move.

Why can't I just measure AI's ROI first?

Because ROI is a ratio and you'd be dividing by a number you don't have. Most agencies underestimate AI's cost by counting only subscriptions, leaving out prompt testing, human review hours, rework, and training. A return measured against an understated cost overstates the result, and the hours that disappear from the timesheet don't disappear from the business.

Is AI going to replace agency staff?

That framing fits large companies with bloat to cut, not a 10 to 50 person agency where cutting heads cuts muscle. The more useful question is what additional work your existing humans can carry now, and whether that extra capacity was actually sold. Capacity nobody sold is a quieter afternoon and the same monthly bill.

What are the biggest AI risks for a marketing agency?

Three distinct ones that often get treated as a single "controls" problem. Unauthorized spend, where a tool buys something within permissions nobody scoped properly, which hits hardest on client media budgets. Data and access exposure, where the wrong material goes into the wrong tool or too many humans can reach it. And unsupervised action, where the software keeps going through a problem instead of stopping and escalating to a person.

Does AI vendor dependency affect what my agency is worth?

Yes, and this is the least-asked question on the list. A buyer is purchasing your ability to keep delivering. A critical workflow that lives only inside a vendor account you can't move reads as a discount rather than an asset, the same way anything else a buyer can't independently verify does. An operating stack with no documented exit path is a diligence finding.

How much should an agency spend on AI tools?

There isn't a benchmark percentage worth quoting, and anyone giving you one is guessing. The useful test isn't the size of the spend, it's whether each tool has a named number it was supposed to move and whether it moved it. A tool with no assigned metric is a subscription, not an investment, however small the monthly charge looks.

Where to Start

Don't try to answer all ten this quarter. You won't finish, and half-answers are worse than none because they feel like progress.

Do this instead. Add up every AI tool your agency pays for, across every team member and every credit card. Then add an honest estimate of the senior hours spent reviewing and reworking what those tools produce. That single number is question one, and most owners have never seen it.

If it surprises you, you've just learned why the rest of the list matters.

The agencies that flourish over the next few years won't be the ones using the most AI. They'll be the ones whose humans can say exactly what it did to the money. That's not a technology decision. It's a numbers decision, and it's the same discipline that decides everything else on your P&L.

If you'd like the tax and profit side of this in one place, I wrote a book for agency owners on keeping more of what you make. You can request a free copy at the link below.

Request a free copy of my book

Craig S. Cody is a CPA, Certified Tax Coach, and retired NYPD Lieutenant. His firm works with more than 70 marketing and advertising agency owners every month, helping them keep more of what they make through proactive tax planning.

This article is general education, not advice for your specific situation. Confirm your own facts with your advisor before acting.