AI marketing ROI: how to measure it honestly.

← Back to all articles

AI marketing ROI is two numbers, not one

AI marketing ROI is the return on the money, time and attention a marketing function puts into AI tools, models and workflows. It splits into two halves that behave nothing alike. The efficiency half is hours removed from work that still has to happen: the brief that took four hours and now takes ninety minutes. The effectiveness half is revenue that arrived because the work got faster or better. You can prove the first one this quarter with a task log. The second runs through the same attribution problem every other marketing investment runs through, and AI does not fix attribution. It adds a variable inside it.

Most decks blend the two into a single percentage, and that is where the credibility goes. A CFO who hears "our AI programme returned 3x" will ask which part of that was hours and which part was revenue. If the answer is a shrug, the number is gone and so is the goodwill. Two numbers. Two confidence levels. Stated out loud.

The stakes here are not academic. Gartner's 2026 CMO Spend Survey, published 11 May 2026, put AI allocation at 15.3% of marketing budgets while total marketing budgets stayed flat at 7.8% of company revenue. That money came from somewhere else inside the same function. Someone traded a headcount, a channel, or an agency line for it, and sooner or later they will be asked what came back.

The public evidence is worse than the vendor decks

Start with the uncomfortable studies, because your CFO has probably read the headlines. MIT's NANDA project reported in its 2025 State of AI in Business study that roughly 95% of enterprise generative AI pilots had produced no measurable P&L return. McKinsey's State of AI survey (March 2025) found more than 80% of respondents saw no material enterprise-level EBIT impact from their generative AI use, even as 78% said their organisation used AI in at least one business function. BCG's October 2024 Where's the Value in AI? report put it structurally: only 26% of companies had developed the capabilities to move past proof of concept into value.

It is important to note what those findings do and do not say. They do not say the technology fails to work. They say most organisations cannot trace what it did, which is a measurement failure wearing a technology failure's coat. The pilots that show nothing on the P&L are usually pilots that were never instrumented to show anything: no baseline, no named task, no owner, no before-and-after. A workflow with no baseline cannot produce a return, only an anecdote.

So the honest read for a marketing leader in 2026 is this. The evidence for AI improving individual task throughput is strong and replicated. The evidence for AI moving enterprise revenue is thin and mostly self-reported. Build your ROI case on the first, and be disciplined about the second.

What is measurable: time saved, per task, against a control

The controlled research here is genuinely good, and it is worth quoting to a finance team precisely because it is not vendor research. Brynjolfsson, Li and Raymond's Generative AI at Work (NBER, 2023) tracked more than 5,000 customer support agents and found a 14% average increase in issues resolved per hour, rising to about 34% for the least experienced agents. Dell'Acqua and colleagues, in Navigating the Jagged Technological Frontier (Harvard Business School with BCG, September 2023), ran roughly 758 consultants through real tasks: those using GPT-4 completed 12.2% more tasks, moved 25.1% faster, and produced work rated over 40% higher in quality, on tasks that sat inside the model's competence.

Notice the shape of both studies. One task type. One control group. One measured output standard. That is the shape your own measurement needs to take, and it is entirely reproducible inside a marketing team: pick the recurring task, log the hours for a month before, log them for a month after, and count only the runs that shipped without a full human rewrite. That last clause is the one everyone skips, and it's the one that keeps the number honest.

The same Harvard study carries the warning that makes it credible. On a task deliberately designed to sit outside the model's competence, consultants using AI were 19 percentage points less likely to reach the correct answer than consultants working without it. Faster and more confidently wrong. Which is why "hours saved across marketing" is a meaningless figure and "hours saved on first-draft campaign briefs, at the same review-gate pass rate" is a defensible one. Measure per task or don't bother.

What is not measurable yet: revenue influenced

Revenue influenced is the number everyone wants and nobody can evidence. The reason is structural, not a tooling gap. AI sits inside the work rather than beside it, so there is no clean counterfactual: the campaign you shipped with AI assistance and the campaign you would have shipped without it also differ by timing, audience, offer, and whatever the market was doing that month. B2B attribution already strains to divide credit across a buying group of six to ten people over a cycle of several months. Adding a variable inside every touch does not make that easier.

This is where the discipline from a proper marketing measurement framework earns its place. Leading indicators, lagging indicators, and a clear statement of which tier a number belongs to. AI-assisted output belongs in the leading tier: cycle time, volume at a held quality standard, coverage of the topics your buyers actually ask about. Revenue stays in the lagging tier, reported as direction of travel with the method named, not as an attributed figure with a decimal point on it.

In my view, the fastest way to lose an AI budget line is to defend it with a number you cannot reconstruct on request. Marketing has spent fifteen years being asked to justify itself with attribution models nobody outside marketing believes. Repeating that pattern with AI, at a moment when the CFO is already sceptical, would be a strange choice.

The measurable wins and the vanity ones

Here is the split I use when a marketing team asks what to put on the slide. The test is simple: could someone outside marketing reconstruct this number from a log, without taking your word for a causal claim?

The pattern in that table is worth sitting with. Every measurable row is about a specific task with a named owner. Every vanity row is a function-level aggregate. The aggregate always looks more impressive on a slide, and it is always the first thing to fall apart in a follow-up question.

How to build an AI marketing ROI case that survives a CFO

Four moves, in this order. They take a quarter, not a year, and they work whether you're running two workflows or twenty.

Instrument before you build. Pick the three recurring tasks that eat the most hours, and log them for one month before any AI touches them. Hours, owner, output standard, rework rate. If a workflow is already live and this baseline does not exist, you can reconstruct an approximation from calendars and ticket history, but do it now rather than at budget time. Most of the pilots in the MIT study failed this step, not the technology step.

Put a named owner and a review gate on every workflow. The gate is what converts "we generated more" into "we shipped more at the same standard". Without it, an efficiency claim is unprovable and a quality regression is invisible until a customer finds it. This is one of the five components of any working AI marketing system, and it is the one teams cut first when they're in a hurry.

Report efficiency as evidence and revenue as direction. On the board slide, one line reads: three named tasks, hours before, hours after, pass rate held. The other reads: pipeline moved this way over the same window, here is what else changed, here is the method. Naming your own uncertainty reads as rigour to a finance audience, not weakness. It also keeps the claim intact when the next quarter is flatter. Pair it with the small set of KPIs a board actually trusts rather than a new AI-specific dashboard nobody asked for.

Re-baseline every two quarters. Efficiency gains are not permanent. Model behaviour changes, the task changes, and the team routes around a gate that slows them down. A number measured once in Q1 and repeated all year is a claim, not a measurement. (And yes, that applies to the systems I build as much as to anyone else's.)

Where does the audit fit in all this? Mostly at step one. A serious look at the marketing function tells you which tasks are worth instrumenting, which workflows have no owner, and where the current AI spend is buying activity rather than throughput. It also tells you where you sit on the AI maturity ladder, which changes what a realistic return even looks like. A team on the second rung measuring itself against fourth-rung claims will always look like it's failing.

The teams that will still have an AI budget in two years are the ones reporting a smaller number they can reconstruct on demand. Honest measurement is not the cautious choice here. It's the one that keeps the line item alive.

Keep reading: Marketing measurement framework for B2B · AI marketing systems for B2B · Marketing KPIs for B2B · The CMO AI readiness gap

Frequently asked questions

What is AI marketing ROI?

AI marketing ROI is the return on the money, time and attention a marketing function puts into AI tools, models and workflows. It has two halves that behave differently. Efficiency return is hours removed from work that still has to happen, and it is measurable per task within a quarter. Effectiveness return is revenue that arrived because the work got faster or better, and it runs through the same attribution limits as every other marketing investment. Reporting them as one blended number is what makes most AI ROI cases collapse under questioning.

How do you measure time saved from AI in marketing?

Measure one named recurring task at a time, with the output standard held constant. Log the hours the task took for a month before the AI workflow existed, log them again for a month after, and count only the runs that shipped without a full human rewrite. The academic work uses exactly this shape: Brynjolfsson, Li and Raymond's Generative AI at Work (NBER, 2023) tracked issues resolved per hour by customer support agents, and Dell'Acqua and colleagues in Navigating the Jagged Technological Frontier (Harvard Business School with BCG, September 2023) tracked tasks completed and time per task against a control group. Function-level averages hide too much to be defensible.

Why can't you measure AI's impact on revenue directly?

Because AI sits inside the work, not beside it, so there is no clean control group. A campaign drafted with AI assistance and a campaign drafted without it also differ by timing, audience, offer and market conditions. Attribution models already struggle to split credit between channels in a B2B buying cycle involving six to ten people over several months, and AI adds a variable inside every one of those touches rather than creating a new one to isolate. Claiming AI-influenced pipeline as a hard number means claiming a counterfactual nobody ran.

Is AI marketing ROI worth reporting to the board yet?

Yes, provided it is reported as two separate lines with two stated confidence levels. Report efficiency as evidence: named tasks, hours before and after, the review-gate pass rate. Report revenue as direction of travel with the method named, not as an attributed figure. That framing survives a CFO question, and it also protects the budget line, because a smaller number that can be reconstructed on request beats a larger one that cannot. Public evidence supports the caution: MIT's NANDA project reported in its 2025 State of AI in Business study that around 95% of enterprise generative AI pilots had produced no measurable P&L return.