How to Calculate AI Marketing ROI (With a Template)

Most AI marketing ROI calculations are wrong in the same direction. They count every hour the tool saved as money earned, they leave out the hours spent reviewing what the tool produced, and they credit the tool for results the team would have got anyway. The number that comes out is large, the finance team does not believe it, and the next budget conversation starts from zero trust.

This is the calculation I use instead. It has three parts: a gain side that only counts what changed, a cost side that includes the costs vendors never mention, and a template you can copy into a spreadsheet. It produces a smaller number. It also produces one you can defend. (Which metrics feed the gain side in the first place is a prior decision - the KPI selection framework is how I make it.)

Why the honest number matters now

AI is no longer a line item anyone can wave through. CMOs now put 15.3% of the marketing budget into AI, while only 30% say they are ready to scale it, according to Gartner’s 2026 CMO Spend Survey. Duke’s CMO Survey found generative AI use up 116% year over year, but deployed in just 15.1% of marketing activities. Spend is running ahead of proof. Sooner or later someone asks what the 15.3% bought, and a calculation built on hours-saved-times-hourly-rate does not survive the question.

The formula

The formula is the ordinary one. The work is in what you allow into each term.

AI marketing ROI = (Gain - Cost) / Cost

Calculate it per workflow, not per tool and not for “AI” as a whole. A tool used in four workflows has four ROIs, and averaging them hides the one that is losing money. A workflow is a repeatable job with a start and an end: drafting the weekly newsletter, producing ad variants for a campaign, writing product descriptions for a catalogue.

The gain side: three kinds, counted three ways

Gains come in three kinds, and each one has a rule about what counts.

1. Time reclaimed, valued only when redeployed. The hours the workflow no longer takes are real. Their value is not the hourly rate times the hours. It is whatever those hours produced instead. If a writer saves six hours a week drafting and spends them on two extra briefs, the gain is the value of two briefs. If the six hours dissolve into meetings, the gain is zero. I ask the person what they did with the time. If the answer is vague, I count half, and I write down that I did.

2. Output lift, measured against a baseline. Where AI changes results rather than effort, measure the result against what the same workflow produced before. More ad variants tested, a higher subject line open rate, more pages published that rank. The rule is that the baseline has to be from before the change and measured the same way. If you cannot state the before number, you cannot claim a lift.

3. Cost avoided, at the real replacement price. Work you would otherwise have bought: a freelancer’s draft, an agency’s variant set, a translation. Count it at what you actually paid last time, not at a list price, and only if you would in fact have bought it. A translation you would never have commissioned is not an avoided cost.

Two things never go on the gain side. Results that would have happened without AI, and results in a channel where AI was one of several changes made at the same time. The attribution problem does not go away because the input was a model. If the AI change landed in the same month as a new offer and a redesign, you do not have a clean number, and the calculation should say so. When a clean number is worth the effort, a holdout test is how I get one.

The cost side: the six lines vendors leave out

The subscription is the smallest line. Here is the full list I use.

Licences and usage. Seats, API tokens, add-ons. The easy line.
Setup time. Prompt writing, voice guidelines, templates, integration. One-off, but count it in the first period.
Review time. Every hour a human spends checking, correcting or rejecting AI output. In my experience this is the line that most often turns a positive ROI negative, and the one most calculations omit entirely. Measure it. Ask reviewers to log a week.
Training and ramp. Time for the team to get competent, and the lower output while they do.
Errors that shipped. Off-brand copy, a wrong claim, a hallucinated statistic that had to be corrected in public. Estimate the cost to fix and any measurable damage. Zero is a legitimate number if nothing shipped wrong, but it has to be a checked zero.
Tool churn. The cost of the tool you stopped using and the migration to the new one. Cheap if you piloted before buying. Expensive if you did not.

Sum these for the same period as the gain. If the period is a quarter, include a quarter of the setup cost only if you expect the workflow to run for a year. Spreading a one-off cost over the life of the workflow is fair. Leaving it out is not.

The template

Copy this into a spreadsheet, one sheet per workflow. Fill in the period first and keep every line inside it.

Line	What goes here	Rule
Workflow	One repeatable job	Not a tool, not “AI”
Period	Month or quarter	Same for every line below
Baseline	What the workflow produced and cost before AI	Measured before the change, same method
Gain: time redeployed	Hours saved x share redeployed x value of what they produced	Zero if not redeployed
Gain: output lift	Result now minus baseline, in money	Same measurement as the baseline
Gain: cost avoided	Work you would have bought, at last paid price	Only if you would have bought it
Cost: licences and usage	Seats, tokens, add-ons	Actual invoices
Cost: setup	Hours x loaded rate, spread over expected life	Include in first period
Cost: review	Logged reviewer hours x loaded rate	Log a real week, do not estimate
Cost: training and ramp	Hours plus reduced output during ramp	Estimate and say so
Cost: errors shipped	Fix cost plus measurable damage	Checked zero allowed
Cost: tool churn	Old tool’s remaining cost plus migration	Zero if none
ROI	(total gain minus total cost) / total cost	Report with the assumptions
Confidence	High, medium or low	Low if any gain line is estimated

I added the last row on purpose. A ROI figure reported without its confidence is the one that gets quoted back at you a year later.

A worked example (hypothetical)

Imagine a four-person B2B marketing team using an AI writing tool for one workflow: drafting their two weekly blog posts. Before AI, each post took the writer about eight hours from brief to publishable draft. After three months with the tool, it takes about four, including the writer’s own edit.

Time reclaimed is eight hours a week. The writer spent about half of it on a second newsletter that did not exist before and the other half went nowhere in particular. Gain from redeployed time: four hours a week of newsletter work, valued at what the team would pay a freelancer for the same, which in this scenario is a few hundred a week. Output lift: the team published the same number of posts, so none. Cost avoided: they stopped buying two freelance drafts a month, at the rate they actually paid.

On the cost side: the licence, about ten hours of setup in month one spread over a year, a review log that showed the editor spending ninety minutes a post checking claims and fixing voice, a week of ramp, one post that shipped with an unsourced statistic and had to be corrected, and no tool churn.

Run through the template, the ROI in this scenario lands somewhere modestly positive, with the review time eating most of what the hours saved were worth. The interesting output is not the percentage. It is that the review line is the biggest cost, which tells the team exactly what to fix next: better briefs and a stricter source rule in the prompt, so the editor’s ninety minutes drops. For content specifically, this workflow view is half the story - the content ROI method adds the performance side, rankings to conversions, with a cohort comparison against human-written work.

All the numbers in this scenario are invented. The shape of the result is what I see in practice: real but modest gains, review time as the hidden cost, and the calculation being most valuable as a diagnostic.

The mistakes that inflate the number

Hours saved times hourly rate. The most common gain-side error. Saved time has no value until it is spent on something.

Vendor ROI calculators. They value every hour at full rate, count no review time and assume full adoption. I use them to understand the vendor’s assumptions, then throw them away.

Crediting AI for a shared result. If three things changed, AI gets a share at most, and the honest report says which share and why.

Measuring the pilot and reporting it as the program. The enthusiast who ran the pilot is not the median user. ROI drops when the workflow spreads, and the calculation should be rerun after it does.

Skipping the review log. Estimated review time is always too low. Log it.

Never rerunning the number. ROI changes as prompts improve, review gets faster and licences get renegotiated. Recalculate quarterly, using the same template, so the trend is visible.

Reporting it upward

Report one number per workflow, its confidence, and the single assumption that moves it most. In my experience a CMO would rather see three workflows at modest, defensible returns than one program at a spectacular number nobody can reconstruct. The CMO’s view of AI spend is about risk as much as return, and a calculation that shows its own weaknesses reads as lower risk.

For the manager doing the reporting, the same rule applies that governs any AI-assisted status report: the numbers come from the log and the spreadsheet, and the model only helps write the narrative around them. An AI tool asked to estimate its own ROI will do exactly what the vendor calculator does.

FAQ

What is a good AI marketing ROI? There is no benchmark worth quoting, because the number depends entirely on what you let into the gain side. A modest positive figure with high confidence beats a large one with low confidence in every budget conversation I have seen.

Should I calculate ROI for the whole AI program? Only as the sum of workflow-level calculations. A program-level number built top-down is a guess with a percentage sign. The program level does add real costs of its own - overlapping subscriptions, abandoned pilots, rework nobody logs - and those hidden costs have their own audit. When the question is what a customer costs rather than what a workflow returns, the revised CAC calculation is the metric-level view of the same discipline.

How long before the number is meaningful? One full period after the ramp. For most workflows that is a quarter. Month one is setup and training, and its ROI is negative by design - what to expect month by month is its own guide.

What if the gain is impossible to put in money? Report it as a measured non-money result, such as posts published or variants tested, next to the money lines. Do not convert it with an invented rate. A gain with no money value is still a gain. A gain with a made-up money value is a liability.

This post belongs to the analytics and ROI hub. The attribution half of the problem, and why last-touch cannot see AI-influenced buyers, is in marketing attribution in the zero-click era. For the executive framing of AI budgets, see AI for CMOs, and for the wider map, the complete guide to AI marketing.