AI Efficiency Without QA Is a Margin Trap. Here Are the Four Numbers That Show Where the Hours Went.

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

AI can make production faster and still make the account less profitable.

That sounds backwards until you measure what happens after the first draft.

If you run a marketing, advertising or PR agency, you've probably already got a number in your head for what AI saved you on some service line. Six hours a week on social. A day on the first cut of a deck. That number is real. It's also the only number most owners are tracking, and it's the one that lies.

The Short Answer

Track four numbers on every AI-heavy service line: hours removed, senior QA minutes, rework cycles, and error escapes. Then do the subtraction. Realized efficiency is the time removed minus the QA and the rework, and the escapes get costed separately, because they're the expensive ones.

Then put human review where a mistake is expensive, not everywhere. And don't hand every dollar of what's left back to the client, because the client isn't buying hours. They're buying the outcome, the judgment, and someone accountable when it's wrong.

What Did Brand-Side Marketers Just Say About Agency AI Work?

Digiday reported on Friday, September 25, on a town hall it hosted the day before with brand-side marketers, held under the Chatham House Rule so people could say what they actually think.

They said the quality control is bad.

One participant described "a growing acceptance of 'good enough, go,'" and said they'd seen creative executions from their agencies where the logo wasn't right. They now insist on human oversight of specific implementations before a campaign is allowed to run. Several others said their agency partners aren't as motivated to protect brand standards as the brand is, and that many agencies are leaving junior employees to oversee the platforms.

A marketer for a premium beauty brand described their own workflow: "I treat the creative output of anything that AI does as a rough draft, and it has to be a 10/80/10 formula. Ten percent of my time is upfront prompting it in the right direction; it does the eighty percent in the middle, and then I have to spend that ten percent to get good work."

Read that last sentence as an agency owner. Your client has a written standard for AI work: a human up front, a machine in the middle, a human at the end. When your agency ships the middle without the end, the client's last ten percent is your QA. They're doing it for free, and they've started saying so at industry events.

Why Is Measuring Only the Middle a Financial Mistake?

Because the middle is the only part that got cheaper.

Say a workflow removes six production hours from a deliverable. Good. Now count what it added. Two hours of senior review, because the output has to be checked by someone who knows the brand, the strategy and the facts. Another hour of rework when the first automated version missed one of the three. And when something slips through anyway, a client fire drill: the call, the apology, the make-good, the account manager's afternoon.

The six-hour headline isn't your efficiency gain. Three of those hours came straight back, and they came back at a senior rate, which is the expensive kind. In hours, the workflow netted three. In dollars, it may have netted nothing, because two hours of a creative director cost more than six hours of a coordinator. Owners net those two against each other in their heads and net wrong.

The people who noticed first were the clients. That's the part of the Digiday piece that should bother you. Not that the logo was wrong. That the brand side is now running its own QA on agency work, unpaid, and has stopped being quiet about it.

What Are the Four Numbers to Track on Every AI-Heavy Service Line?

Four, and only four. All of them come off a timesheet and a task board you already have.

1. Hours removed. How much human production time did the workflow actually eliminate? Not the vendor's claim. Yours. Compare the same deliverable before and after, on the same client, and write the difference down.

2. Senior QA minutes. How much expensive judgment did the output need before it was safe to ship? Give review its own task code, and book it at the reviewer's real cost, not the coordinator's. Minutes, because that's the unit senior people think in and the unit that disappears when nobody's counting.

3. Rework cycles. How often did the output come back because the first automated version missed the brand, the strategy or the facts? A cycle is one round trip. Count the round trips, not the hours, because the count tells you whether the prompt, the tool or the process is the problem.

4. Error escapes. How many mistakes reached the client or the market, and what did each one cost in time, trust and make-good work? Keep a log. Date, deliverable, what escaped, who caught it, what it cost to fix. A wrong logo that a client catches is an escape. A wrong claim that runs is a bigger one.

Do this per service line, not for the agency as a whole. AI-assisted social content and AI-assisted media planning will give you completely different numbers, and an agency-wide average hides the one that's bleeding.

How Do You Calculate Realized Efficiency?

The hour version first:

Realized efficiency = hours removed, minus senior QA time, minus rework time.

Then the dollar version, which is the one that matters, because the hours you removed and the hours you added are not the same price:

Realized efficiency in dollars = (hours removed at the production rate) minus (senior QA at the senior rate) minus (rework at whoever did it).

Then the escapes, costed separately and not netted into the formula, because a single escape can wipe out a quarter of gains on that line, and averaging it in hides that.

Illustration, not a client: a workflow removes six hours at $60 of loaded cost, which is $360 saved. It adds two hours of creative director review at $150, which is $300, and one hour of rework at $60. Realized efficiency in hours is three. Realized efficiency in dollars is zero. Add one escape that took an account manager four hours to make good and the line lost money on the month the tool was supposed to save it.

That's what "AI made us faster and the account got less profitable" looks like on paper. It's not a paradox. It's a subtraction nobody did.

What Does This Change About Pricing?

One thing, and it's the thing owners get backwards.

If AI makes delivery cheaper, the agency does not have to give every dollar of that efficiency back to the client. The client is still buying the outcome, the judgment, and the accountability. None of those got cheaper. One layer of the work did.

But you only get to keep that margin if the work still meets the standard. The moment the logo is wrong, the client is buying your QA with their own time, and the argument that your fee covers judgment and accountability collapses, because they just supplied both.

I've written separately about what the fee is actually made of and about what to say when a client asks for the AI savings. This article isn't that. This one is about the number you need in hand before either conversation, and the review that protects it.

Where Does Human Review Belong?

Start with my own world. In accounting, there have been different levels of AI on the bookkeeping side for years. It never removed the review process. Anyone who's worked in QuickBooks Online knows the results are only as good as the review process behind them. The automation runs the middle. The review is where the books get right, and that was true before anyone put "AI" in a pitch deck.

That's the same standard your clients just described. So, not everywhere. Human review everywhere is how you turn a six-hour saving into a six-hour cost and then tell yourself AI doesn't work.

It belongs where a mistake is expensive: brand, reputation, strategy, money, or client trust. Rank every deliverable on the service line by what an escape would cost, and put senior review on the top of that list, in full, every time. Sample the rest. A wrong headline in a paid campaign for a premium brand gets a senior pair of eyes before it runs. The fourth variant of an internal caption test doesn't.

Notice that this is exactly the standard the brand side described. Ten percent up front, ten percent at the end, and a human on both. The complaint wasn't that agencies use AI. It was that agencies stopped doing the last ten percent, or handed it to the most junior person on the account. Junior oversight of a platform is not review. It's someone watching the output go by.

The goal is a review budget that's proportional to the cost of being wrong, tracked as senior QA minutes on the ledger above, and priced into the work, because the client is already paying for judgment whether you deliver it or not.

What This Article Is Not About

It isn't the full list of what AI costs an agency; subscriptions, prompt testing and training are real and belong in a different piece. It isn't whole-agency measurement of whether AI paid off; that's a scoreboard question and it has its own article. And it isn't about whether you redesigned the work or just sped it up, which is the question underneath all of these. The ten AI questions an agency owner should ask, in order, are here.

This one is narrower. It's the subtraction that happens after the first draft, and where to put the human so the subtraction comes out positive.

Who This Is For, and Who It Isn't

This is for a marketing, advertising or PR agency owner with at least one service line where AI now does most of the middle, and a client roster that includes a brand that cares what its logo looks like. If you've quoted an hours-saved number to anyone in the last quarter, this is the article that asks what the other three numbers were.

It isn't for an agency that hasn't put AI into delivery yet; you have nothing to subtract. And it isn't a case against review. It's a case for putting review where it earns its cost and counting it where it doesn't.

AI efficiency without QA isn't leverage. It's unpriced risk.

Frequently Asked Questions

How do I measure whether AI actually saved my agency time?

Compare the same deliverable, on the same client, before and after the AI workflow, and record the production hours removed. Then record the senior review time and the rework the output needed before it shipped. Realized efficiency is the removed time minus the review and rework, and the dollar version uses each person's real cost, because senior review is priced higher than the production time it replaced.

What is the 10/80/10 rule for AI content?

It's a workflow a brand-side marketer described at Digiday's AI Marketing Strategies event in September 2026: ten percent of the human's time up front directing the tool, the tool doing eighty percent of the work in the middle, and ten percent of human time at the end refining it. The brand treats everything the tool produces as a rough draft. For an agency, the last ten percent is the review step, and skipping it hands that work to the client.

Should an agency pass AI cost savings on to clients?

Not by default. The client is buying an outcome, the agency's judgment and its accountability, and AI changed the cost of delivery, not the value of those three. The margin is yours to keep as long as the work meets the standard. If the standard slips, the client is supplying the judgment with their own review time, and the case for the fee weakens.

How much human review does AI-generated work need?

Enough to cover the deliverables where a mistake is expensive, and no more. Rank the work on a service line by what an error would cost in brand, reputation, strategy, money or client trust; put senior review on the top of that list every time, and sample the rest. Review on everything wastes the efficiency. Review on nothing turns it into risk. Accounting has run automated bookkeeping for years, and the results were only ever as good as the review process behind them.

What is an error escape and why track it?

An error escape is a mistake that reached the client or the market: a wrong logo in a campaign, an incorrect claim, a strategy miss the client caught. Track each one with its cost in time, trust and make-good work, separately from the efficiency formula, because a single escape can erase a month of savings on that service line and averaging it in hides that.

Why do clients say agency AI work is "good enough, go"?

Because they're seeing the output ship without the final human pass. Brand-side marketers at Digiday's September 2026 event described agency executions with incorrect logos, agencies leaving junior staff to oversee platforms, and having to insist on human oversight before campaigns could run. The phrase is theirs, and it describes an agency that removed the middle of the work and the end of it.

Let's Talk

If the only AI number you can quote is hours saved, you don't know yet whether the account got more or less profitable. Four numbers on one service line will tell you in a month. Start with the line you're proudest of.

I wrote a book called The 12 Biggest Tax Mistakes That Cost Agency Owners Thousands. You can request a free copy at the link below.

Request a free copy

Nobody's looking out for your money but you. Let's go look together.

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.