AI Agent ROI Is Barely Being Measured
AI agent ROI is the number almost nobody in marketing can produce right now, and the adoption curve has run well past the point where that is forgivable. In [Basis Technologies' AI and the Future of Marketing, 2026 Edition](https://www.basis.com/news/emarketer-marketers-ai-adoption-is-limited-by-lack-of-roi-strategies), a survey of 184 marketing and advertising professionals fielded in August 2026, 51.1% of organisations said they now deploy AI agents, a 155.5% jump year over year, and 47.8% said they have replaced human tasks with AI in the past twelve months. In the same sample, 6.5% measure their AI investments against a defined benchmark, and 65.1% either do not measure the return at all or do not know whether anyone in the business does. That is a small sample, 184 respondents, not a census of marketing. But the shape of the gap is hard to argue with: the spend went into production and the measurement never followed it.
What the Basis numbers actually say
Treat the figures as directional, because 184 respondents is a small base and the people who answer an AI survey skew towards people who use AI. With that caveat stated plainly, the internal contrast inside the sample is the interesting part, because every number comes from the same 184 people. Generative or agentic AI use sits at 98.4%. Daily use sits at 61.4%, up from 33.6% a year earlier. Agent deployment sits at 51.1%. And benchmarked measurement sits at 6.5%. You do not need a bigger sample to see that adoption and accountability moved at completely different speeds inside the same group of teams. MarTech Cube's write-up of the report landed on the same read: the constraint on AI in marketing right now is not capability, it is the absence of any ROI strategy behind the capability.
| What the 184-respondent sample reports | Figure |
|---|---|
| Use generative or agentic AI at all | 98.4% |
| Use it daily | 61.4% (up from 33.6% a year earlier) |
| Deploy AI agents | 51.1% (155.5% year-over-year increase) |
| Have replaced human tasks with AI in the past 12 months | 47.8% |
| Measure AI investments against a defined benchmark | 6.5% |
| Do not measure the return, or do not know if anyone does | 65.1% |
Nearly half this sample has already handed work that a person used to do to a system nobody is benchmarking. At that point the unmeasured thing is not a pilot or an experiment. It is production, running without a number attached to it.
Why the gap exists structurally, not through laziness
The easy explanation is that marketers are careless about measurement, and that explanation is wrong. The same teams measure paid media to two decimal places. The difference is how AI spend enters a business. A campaign arrives as a budget line with a target attached, so somebody owns the result by default. AI arrives as a scatter of tool subscriptions on a card, plus a vague promise of time saved, and it gets approved as overhead rather than as an investment. Overhead does not get a target. Nobody is ever asked at quarter end whether the overhead worked.
Agents make that worse rather than better, because an agent's output is usually an input to something else. It drafts the brief, cleans the list, writes the first version, summarises the call. None of that is a revenue event, so there is nothing obvious to attribute, and the honest answer to 'did it work' becomes 'it felt faster.' Add the review time nobody logs, which is the pattern covered in the AI babysitting tax, and even the time-saved claim stops being safe without a baseline. The result is not a measurement failure. It is an ownership failure that shows up as a measurement failure.
| Why AI spend escapes measurement | What it looks like in practice |
|---|---|
| It enters as subscriptions, not as a budget line with a target | Approved as a tool cost, so no one is accountable for a return on it at quarter end |
| The benefit is claimed as time saved, with no before number | Nobody recorded how long the task took previously, so the saving cannot be proved or disproved |
| Agent output is an input, not an outcome | The draft, list, or summary sits upstream of revenue, so there is no obvious metric to attach |
| Review and fix time is unlogged | The true cost is the subscription plus the supervision, and only the subscription is visible |
| No review date was ever set | The tool renews indefinitely because no moment exists at which it has to justify itself |
The measurable frame: four things per use case
Benchmarking AI does not require an attribution rebuild or a new analytics stack. It requires writing four things down per use case, before the automation goes live, and the reason only 6.5% of that sample has a benchmark is almost always that the tool was switched on before anyone wrote them. Pick one use case, not the whole stack, and do this for it.
- 01A baseline, captured before the tool. How long does this task take today, how many outputs per week, and what does the current result look like? If the task is already automated and nobody recorded this, reconstruct it honestly from memory and label it as an estimate rather than skipping it.
- 02One named output metric, chosen per use case and not shared across them. Content drafting gets published pieces per week or time to first draft. Lead qualification gets qualified conversations booked. Reporting gets hours to produce the report. One metric, named in advance, so the answer cannot be rewritten after the fact.
- 03Total cost including review time. Subscription plus the hours someone spends prompting, checking, correcting, and re-running. Review hours are the line item that turns an apparent win into a wash, which is exactly why they get left out.
- 04A review date, in the calendar, with a name on it. Thirty, sixty, or ninety days out depending on the cycle. At that date the use case is kept, changed, or cut, by a specific person, against the metric written in step two.
The reason to keep it to one metric per use case is the same reason behind metrics that predict revenue: a measurement everyone can see and nobody can reinterpret is worth more than a dashboard of numbers that each tell a different story. And deciding which use cases deserve a benchmark at all is the judgement call in AI marketing audit, automate versus human, some tasks are worth automating and measuring, some are worth automating and ignoring, and some should never have been automated.
What a benchmarked use case looks like written down
| Field | Example, AI drafting the weekly email |
|---|---|
| Baseline before the tool | 3.5 hours per email, 1 email per week, 22% open rate over the last 8 sends |
| Named output metric | Hours to a send-ready draft, with open rate held at or above 22% |
| Total cost | Subscription per month, plus 45 minutes per email of editing and fact-checking |
| Review date and owner | Reviewed at 60 days by the person who owns the channel, kept, changed, or cut |
That is the whole discipline. It fits in four lines per use case, and it is the difference between a stack that can justify itself and one that renews on momentum. The cost of skipping it is not only wasted subscription spend, it is the quieter damage covered in the cost of a bad marketing prompt, where unmeasured output keeps shipping because nothing in the process was ever set up to catch it.
The bottleneck is accountability, not tooling
Read the Basis figures together and the conclusion is not that marketing teams need better AI. At 98.4% usage and agent deployment up 155.5% in a year, tool access is solved. What is missing is somebody senior enough to set a target before the tool is switched on, and present enough to sit at the review date and cut what is not working. That is a leadership gap, which is why the answer here is a fractional CMO rather than another platform, someone accountable for distribution and sales outcomes who treats AI spend the way they would treat a media budget: with a number on it, a name against it, and a date when it has to prove itself. The 6.5% figure is not a technology problem. It is the predictable result of nobody being asked to own the answer.
FREQUENTLY ASKED
How many marketing teams actually measure AI ROI?
In Basis Technologies’ AI and the Future of Marketing, 2026 Edition, a survey of 184 marketing and advertising professionals fielded in August 2026, 6.5% said they measure AI investments against a defined benchmark, while 65.1% either do not measure the return or do not know whether anyone in the business does. That is a small sample rather than a census, so treat it as directional, but the contrast against 98.4% AI usage inside the same sample is stark.
Why is AI marketing ROI so hard to measure?
Because AI spend usually enters a business as tool subscriptions and a general promise of time saved, rather than as a budget line with a target, so nobody is accountable for the return. Agent output also tends to be an input to other work, a draft, a cleaned list, a summary, rather than a revenue event, so there is no obvious metric to attach. Add unlogged review and correction time and even the time-saved claim becomes unprovable without a baseline.
What is the simplest way to benchmark an AI marketing use case?
Write down four things before switching the tool on: a baseline captured beforehand, one named output metric for that specific use case, the total cost including review hours, and a review date with a named owner. At the review date the use case is kept, changed, or cut against the metric. Doing this for one use case properly beats attempting it across a whole stack.
How many marketing teams are running AI agents now?
In the Basis 2026 sample, 51.1% of organisations said they deploy AI agents, which the report describes as a 155.5% increase year over year, and 47.8% said they had replaced human tasks with AI in the previous twelve months. Again, 184 respondents is a small base and the respondents skew towards AI users, so the direction matters more than the precise percentage.
Does measuring AI ROI require new analytics tooling?
No. The missing piece in most teams is not measurement capability, it is a target set before the tool went live and a person accountable at a review date. A baseline, one named metric, honest cost including review time, and a date in the calendar can all be recorded in a few lines without touching the analytics stack.
RELATED SERVICES
Delivered in 22+ markets worldwide.
Want this done for your business?
Free audit. No pitch. 24-hour turnaround.
PROOF THIS WORKS