Sohail Kasmani. Brand and marketing strategist in London: pharma brand management at Cipla and Alembic, now marketing strategy and AI adoption at ACTAsia. MBA, Brunel University London.
How to measure the ROI of AI in marketing (with a worked example)
To measure the ROI of AI in marketing, you count three kinds of value separately (efficiency savings, incremental revenue or profit, and reduced risk or better quality), subtract the full cost of the AI including the human review time it needs, and divide by that cost. The effectiveness part has to come from a controlled test against a baseline, because time saved tells you nothing about whether the work still sells.
Most organisations can’t show that return yet. PwC’s 2026 Global CEO Survey of 4,454 chief executives found that 56% have seen no significant financial benefit from AI to date, and only 12% report both cost and revenue benefits.
How do you measure the ROI of AI investments in marketing operations?
You start by refusing to let one number do all the work, because each of the three value types needs its own measurement method.
- Efficiency value: agency fees you no longer pay, production costs avoided, hours freed and actually redeployed. This is the easy one to count, which is exactly why it gets overcounted.
- Effectiveness value: extra gross profit that would not have happened without the AI, from better targeting, email personalisation or faster creative testing. You can only see this with a holdout or a controlled experiment.
- Risk and quality value: fewer compliance errors, fewer factual mistakes going out under your brand, faster catch of a campaign going wrong. It also runs the other way, because AI can add risk (a chatbot confidently inventing a returns policy is a cost, just one that turns up in the complaints inbox rather than the invoice).
The ROI formula
AI ROI (%) = (Efficiency value + Incremental gross profit + Risk/quality value - Total cost of AI)
/ Total cost of AI x 100
Total cost of AI = licences + integration and data work + training
+ ongoing human review and correction time + change management
Two lines in that formula decide whether the final number means anything. The effectiveness line uses gross profit, not revenue, because an extra £100,000 of sales at a 40% margin is worth £40,000 to the business and your finance director knows that even if the campaign deck doesn’t. And the cost line includes the review time, which is the bit vendor ROI calculators tend to forget, presumably out of modesty.
Why measuring only efficiency hides losses
The clearest recent evidence comes from Adam Peruta, an associate professor at Syracuse University’s Newhouse School of Public Communications, who co-led a study with the research firm Ipsos, published in HBR as “Research: AI-Generated Ads Perform Worse Than Human-Made Ones, Even When Customers Can’t Tell Them Apart” (3 September 2026). They showed 3,000 US consumers 20 video ads for major brands, half made by people and half rebuilt entirely with AI from briefs based on the originals. Viewers struggled to tell them apart, but the AI versions were 14% weaker on short-term sales potential and 17% weaker on long-term brand equity.
Peruta’s point is that the metrics a team naturally reaches for, like cost per asset and production time, would have made the AI ads look like a triumph. A CFO sees lower production costs, a CMO sees faster campaign cycles, and nobody sees the 14% until the sales line softens and we all blame the category, the pricing, the media mix and probably the weather. AI did hold up on conventional, product-led work, though: the AI versions of the Febreze and Herbal Essences ads (both Procter & Gamble brands) did at least as well as the human versions on short-term sales potential, which argues for testing by format rather than banning anything.
Masha Shunko (University of Washington) and Serguei Netessine (Wharton) make the same argument from operations management in HBR’s “Stop Automating Old Processes. Design New Ones Instead.” (14 September 2026). One of their four failure modes for AI pilots is “optimizing the wrong metric”, where the dashboard improves while outcomes deteriorate, like a customer service agent that raises ticket closure rates by ending conversations quickly and creates repeat contacts. Anyone who has watched an email team celebrate open rates while unsubscribes climb will recognise the pattern.
Klarna, the Swedish buy-now-pay-later company, shows both sides. In May 2024 it said it had cut its spending on external marketing suppliers by 25% (translation, production, CRM and social agencies), and that AI accounted for about $10 million of annualised sales and marketing savings, including a $6 million reduction in image production costs. On the customer service side, where it had pushed hard on an AI assistant, a Klarna statement published by Forbes in May 2025 said chief executive Sebastian Siemiatkowski had acknowledged that an overemphasis on cost, not AI itself, led to lower quality, and the company described a “dual-track” approach that adds human support back alongside the AI. If your only customer service metric is cost per contact, that kind of quality drop shows up as a lovely downward line on the dashboard right up until it shows up in the churn numbers.
A worked example with illustrative numbers
The numbers below are illustrative, made up to show the arithmetic, and not taken from any real company. Imagine a UK direct-to-consumer brand that moves its paid social creative production to an AI workflow for a year.
| Line | Efficiency-only view | Full view (AI on all creative) | Full view (AI on tested formats only) |
|---|---|---|---|
| Licences, integration, training | -£65,000 | -£65,000 | -£65,000 |
| Human review and correction time | -£15,000 | -£15,000 | -£15,000 |
| Total cost of AI (the two lines above) | £80,000 | £80,000 | £80,000 |
| Agency and production savings | +£90,000 | +£90,000 | +£65,000 |
| Team hours redeployed at £30 an hour | +£30,000 (1,000 hours) | +£30,000 (1,000 hours) | +£25,000 (about 830 hours) |
| Incremental gross profit vs human creative (from holdout) | not measured | -£48,600 | £0 |
| Fewer rework cycles and approval errors | not measured | +£3,600 | +£3,600 |
| Total value | £120,000 | £75,000 | £93,600 |
| ROI | 50% | -6% | 17% |
The middle column assumes a holdout test showed the AI creative drove 9% less incremental revenue than the human creative, on a paid social channel producing £1.2 million of incremental revenue a year. That is £108,000 of lost revenue, and at a 45% gross margin it costs £48,600 of gross profit. The efficiency-only view reports a 50% return on a project that is losing money.
The right-hand column is what happens when the team only uses AI for formats where the test showed parity (product demos, resizes, localised versions) and keeps people on the hero work. The savings shrink, but this time the 17% return survives contact with a holdout test. It still leaves out brand equity, which you won’t see in a single year’s numbers, so I’d report it next to the ROI as a tracked guardrail rather than pretend it’s in the sum.
What should an AI ROI measurement plan include?
- Write the outcome sentence first. Shunko and Netessine suggest every AI pilot completes this: “We will improve [outcome] in [end-to-end process] by redesigning [workflow], using AI to [role], with humans retaining authority over [judgment/control], exceptions routed through [escalation path], success measured by [primary KPI], and performance protected by [guardrails].” If the team can’t fill in the guardrails, it isn’t ready to measure anything.
- Take a baseline before switching anything on. I’d record at least six to eight weeks of the metrics you will judge the AI on (cost per asset, cycle time, conversion rate, complaint rate), because baselines taken after launch have a habit of flattering the launch.
- Run a holdout or A/B test for the effectiveness line. For paid media, Google Ads Conversion Lift splits audiences into exposed and control groups (by user or by geography), though it isn’t available on every account and you have to ask your Google representative. For email and CRM, hold back a random 10% of the audience from the AI-personalised version. For creative, test AI and human versions against the same brief and the same spend.
- Separate incrementality from attribution. Attribution reports which touchpoints were present before a conversion; incrementality tells you which ones caused it. Google Analytics 4 now offers only data-driven attribution and two last-click models, having retired first-click, linear, time-decay and position-based models in November 2023. Data-driven attribution only sees trackable clicks, so it cannot tell you whether an AI change created demand or just got credit for it.
- Use marketing mix modelling for the channel-level and long-term picture. Google’s Meridian and Meta’s Robyn are both open-source MMM packages that can be calibrated with experiment results (Meridian with geo experiments, Robyn with geo tests and Meta’s own lift studies). MMM needs years of weekly data: Google’s Meridian guidance on data volume works through an example with three years of it, so it suits established brands rather than a six-week pilot.
- Count the full cost, including the hours people spend checking, correcting and re-prompting. If a £20-a-month tool needs a senior marketer to fix its output for five hours a week, it is not a £20-a-month tool.
- Give each value type its own time horizon and review monthly by use case. Efficiency shows up within a quarter, conversion effects within a test cycle, brand effects over a year or more in your brand tracker. Peruta’s rule for creative works for the rest too: if the AI version performs as well or better against the same brief, scale it, and if it falls short, shelve it before it wastes media money.
If you are using an AI assistant to help read the test results, give it the design and the raw numbers and ask it to be sceptical, something like this:
Results of a 4-week holdout test. Treatment: 45,000 customers got AI-personalised emails. Control: 5,000 got the standard version, randomly assigned. [paste conversions, revenue, unsubscribes for each group] 1. Calculate incremental conversion rate and revenue per customer, with 95% CIs. 2. Is the difference statistically meaningful at this sample size? 3. What in the design could bias the result (timing, list selection, novelty), and what data would rule it out? Do not round, and say plainly if the test is underpowered.
Which metrics should you track for each AI use case?
| Use case | Efficiency metric | Effectiveness metric | Risk/quality guardrail | Best test design |
|---|---|---|---|---|
| Content and creative | Cost per asset, time from brief to approval | Conversion rate, pre-test scores, brand lift | Brand tracker, error and correction rate, disclosure compliance | Same brief, AI vs human, matched spend |
| Email and CRM | Hours per campaign, send volume per marketer | Incremental revenue per recipient, repeat purchase | Unsubscribe and spam complaint rate | Random holdout of 5-10% of the list |
| Paid ads (bidding, targeting, creative) | Hours on campaign management, agency fees | Incremental conversions, incremental ROAS | Brand safety incidents, frequency, new vs returning customer mix | Conversion lift or geo experiment |
| Segmentation | Analyst time per segmentation refresh | Response and margin by segment vs old segments | Segment stability over time, data protection compliance | Old segments vs AI segments on split audience |
| Analytics and reporting | Hours saved on reporting, time to insight | Decisions changed and their measured outcome | Accuracy of AI-generated numbers against source data | Audit a sample of outputs against the raw data |
| Chatbots and conversational AI | Cost per contact, share handled without a human | Conversion from assisted sessions, retention | Repeat contact rate, escalation rate, satisfaction, factual errors | Randomly route a share of traffic to human-only service |
The guardrail column is where most of us get lazy. Our guide to AI customer segmentation goes into the segment-level testing in more detail, and how to build an AI marketing workflow covers where the human checks sit.
How do AI analytics platforms improve campaign performance tracking?
AI in marketing analytics helps most with the tedious parts of tracking. Google Analytics 4’s Analytics Insights uses machine learning to flag unusual changes and emerging trends automatically, and lets you set up to 50 custom rules per property, including a “has anomaly” condition that doesn’t need a manual threshold. So a broken tracking tag or a campaign that falls off a cliff on a Saturday gets noticed on Saturday instead of at Tuesday’s meeting.
Natural-language querying helps too (asking “which campaigns lost conversion rate this month” rather than building a report), but these tools only summarise the data they are given. If your tracking undercounts offline sales, or the platform is marking its own homework on attributed conversions, the AI will produce a very fluent summary of the wrong number, which is why the holdout tests above still matter.
What are the top benefits of using AI workflows in marketing analytics?
The one that matters most for ROI is more testing: cheaper creative variants and faster analysis make it realistic to run controlled tests on every use case instead of one a year. After that come faster reporting, earlier warnings on spend and tracking problems, brand managers querying the data without joining the analytics queue, and open-source MMM that mid-sized teams can run because the software is free (the analyst time is another matter). Every one of those comes with a checking cost, and I’d budget for it from day one. There is more on how that plays out day to day in the ACTAsia case study on AI adoption in marketing.
Can AI improve ROI? What the current evidence says
It can, but only a minority of firms can show it, and the big 2025-26 surveys agree on roughly how small that minority is.
- McKinsey’s “The State of AI in 2026: On the Road to ROI” (1,719 respondents across 97 countries, surveyed May to June 2026) found 37% attribute at least some EBIT impact to AI use, about the same as the year before. Only 6% qualify as high performers, meaning they attribute 5% or more of EBIT to AI and report significant value from it. Marketing and sales (along with product development) is where respondents most commonly report revenue gains.
- BCG’s “The Widening AI Value Gap” (1,250 senior executives, September 2025) classed 5% of companies as “future-built” and 60% as laggards. The future-built group expects twice the revenue increase and 40% greater cost reductions than laggards in the areas where they use AI, and BCG puts 70% of AI’s potential value in core functions including sales and marketing.
- PwC’s CEO survey, above, has 12% getting both cost and revenue benefits and 33% getting one or the other.
- Gartner’s 2026 CMO Spend Survey (401 marketing leaders in North America, the UK and Europe) found CMOs put 15.3% of marketing budgets into AI while only 30% report mature or fully developed AI readiness.
Spending keeps climbing regardless: BCG’s January 2026 AI Radar survey of 640 CEOs, cited in Shunko and Netessine’s HBR article, reports that companies expect to roughly double AI spending this year, from 0.8% of revenue to about 1.7%. So marketing teams are putting 15.3% of their budgets into AI while most of the organisations signing off the spend can’t yet prove a return. More of our figures are in AI in marketing statistics 2026, and if you want the basics first, start with what AI in marketing actually covers.
A short guide to selecting an AI platform for marketing data analysis
- Data connection: does it connect directly to your analytics, CRM and ad platforms, or does someone export CSVs every Monday forever?
- Traceability: can every number it produces be clicked back to the source query or table? If not, you can’t audit it.
- Causal tools: does it support holdout design, lift analysis or MMM calibration, or only attribution reporting?
- Independence: is it built by a media seller grading its own channel? Meridian and Robyn come from Google and Meta, so that’s not disqualifying, but you should know whose homework it is.
- Data protection: where is the data processed, is it used to train the vendor’s models, and does that fit UK GDPR for your customer data?
- Exit: can you take your models, prompts and data with you if you leave?
The question I’d put to any marketing team right now is simple to ask and awkward to answer: of the AI tools you are paying for today, which one has a holdout test showing it made more money than it cost, and if none do, which one are you going to test first?
FAQs
What is a good ROI for AI in marketing?
There is no standard benchmark, because published surveys measure whether firms see any financial benefit rather than a typical return. PwC’s 2026 CEO survey found 56% of CEOs saw no significant financial benefit from AI. A sensible internal test is that the full-cost ROI, including human review time and measured through a holdout, beats the return you would get from spending the same money on media or people.
How long does it take to see ROI from AI in marketing?
Efficiency savings such as lower agency or production costs usually show within a quarter. Effects on conversion need a controlled test, and four to eight weeks per use case is a sensible plan for most email and paid media tests. Brand effects take a year or more to appear in brand tracking, which is why the Ipsos study’s 17% lower brand equity for AI-made ads is a risk most short-term ROI calculations would miss entirely.
What is the difference between attribution and incrementality in AI marketing measurement?
Attribution assigns credit for a conversion to the touchpoints that came before it, using models such as GA4’s data-driven attribution or last click. Incrementality measures whether the conversion would have happened anyway, by comparing an exposed group with a randomly held-out control group. To judge an AI investment you need incrementality, because attribution can give credit to activity that caused nothing.
Should AI time savings count towards marketing ROI?
Only when the freed hours are redeployed to work with a value, or when they reduce real costs such as agency fees or contractor spend. Hours saved that disappear into more meetings are not a financial return. Count them at a loaded hourly cost, and subtract the time people now spend reviewing and correcting AI output, which is often larger than expected.
Can marketing mix modelling measure the impact of AI?
Yes, at channel level and over time, if you have enough history. Google’s Meridian documentation works through an example using three years of weekly data. Open-source tools such as Meridian and Meta’s Robyn can be calibrated with lift test or geo experiment results. MMM is less suited to judging a single AI tool in a short pilot, where a holdout test gives a faster and cleaner answer.
Sources
- Harvard Business Review, Adam Peruta, “Research: AI-Generated Ads Perform Worse Than Human-Made Ones, Even When Customers Can’t Tell Them Apart”, 3 September 2026. https://hbr.org/2026/09/research-ai-generated-ads-perform-worse-than-human-made-ones-even-when-customers-cant-tell-them-apart
- Harvard Business Review, Masha Shunko and Serguei Netessine, “Stop Automating Old Processes. Design New Ones Instead.”, 14 September 2026. https://hbr.org/2026/09/stop-automating-old-processes-design-new-ones-instead
- PwC, “PwC 2026 Global CEO Survey” (press release), 19 January 2026. https://www.pwc.com/gx/en/news-room/press-releases/2026/pwc-2026-global-ceo-survey.html
- McKinsey & Company (QuantumBlack), “The State of AI in 2026: On the Road to ROI”, 25 August 2026. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
- Boston Consulting Group, “AI Leaders Outpace Laggards with Double the Revenue Growth and 40% More Cost Savings” (press release on The Widening AI Value Gap), 30 September 2025. https://www.bcg.com/press/30september2025-ai-leaders-outpace-laggards-revenue-growth-cost-savings
- Gartner, “Gartner 2026 CMO Spend Survey Finds CMOs Allocate 15.3% of Marketing Budgets to AI, But Only 30% Are Ready to Scale AI Capabilities”, 11 May 2026. https://www.gartner.com/en/newsroom/press-releases/2026-05-11-gartner-2026-cmo-spend-survey-finds-cmos-allocate-15-point-3-percent-of-marketing-budgets-to-ai-but-only-30-percent-are-ready-to-scale-ai-capabilities
- Klarna, “AI helps Klarna cut marketing agency spend by 25% and run more campaigns”, 28 May 2024. https://www.klarna.com/international/press/ai-helps-klarna-cut-marketing-agency-spend-by-25-and-run-more-campaigns/
- Forbes, Gene Marks, “Business Tech News: Klarna Reverses On AI, Says Customers Like Talking To People”, 18 May 2025. https://www.forbes.com/sites/quickerbettertech/2025/05/18/business-tech-news-klarna-reverses-on-ai-says-customers-like-talking-to-people/
- Google Ads Help, “About Conversion Lift”, accessed September 2026. https://support.google.com/google-ads/answer/12003020
- Google Analytics Help, “Get started with attribution”, accessed September 2026. https://support.google.com/analytics/answer/10596866
- Google Analytics Help, “Analytics Insights”, accessed September 2026. https://support.google.com/analytics/answer/9443595
- Google for Developers, “Meridian”, accessed September 2026. https://developers.google.com/meridian
- Google for Developers, “Amount of data needed” (Meridian documentation), accessed September 2026. https://developers.google.com/meridian/docs/pre-modeling/amount-data-needed
- Meta Marketing Science, “Robyn”, accessed September 2026. https://facebookexperimental.github.io/Robyn/