Why Your AI Vendor's ROI Promise Is Probably Wrong

AI vendor ROI claims are the projected return-on-investment figures that consulting firms, SaaS platforms, and automation vendors present during sales. They express expected savings or revenue gains as percentages or dollar amounts. The gap between these projections and actual measured results is the central measurement problem in enterprise AI adoption.

Gartner published data on April 7, 2026 from a survey of 782 infrastructure and operations leaders: only 28% of AI projects fully succeed and meet ROI expectations (Gartner, April 2026). One in five fails outright.

McKinsey's 2025 State of AI survey tells a parallel story: 88% of organizations use AI regularly in at least one business function, yet for most of them AI has not yet significantly affected enterprise-wide EBIT (McKinsey). Adoption is widespread, but measured returns remain scarce.

AI works. The vendor ROI projections surrounding it are structurally flawed, because they contain three systematic errors that inflate expected returns before a single line of code runs.

Error One: Baseline Neglect

The projection compares the vendor’s modelled after-state against an estimate of your before-state. Nobody measured the before-state.

This is the error that makes every other number unfalsifiable. If you do not know how many hours the workflow consumes today, what its error rate is, and what it costs per transaction, then any post-deployment figure is a claim rather than a measurement. The vendor is not usually being dishonest here. They are working from an industry assumption because you did not hand them your own number.

RAND found the root causes of AI project failure to be organizational rather than technical, with leadership misunderstanding the problem ranking first (RAND Corporation). Not knowing your own baseline is that failure in its most literal form.

The fix costs a week. Before any vendor conversation, measure the workflow you intend to automate: hours per week, error rate, rework time, cost per unit of output. Those four numbers turn every subsequent claim into something you can test.

Error Two: Cost Exclusion

The quoted price is the licence or the build. The actual cost includes integration with the systems you already run, data cleanup, staff training, the productivity dip while people adjust, and the ongoing monitoring that keeps the thing working after launch.

Vendor ROI models routinely count the first item and omit the rest. That is not a rounding difference. Integration alone regularly costs more than the tool, because it surfaces data quality problems that predate the AI project entirely.

The test is simple: ask the vendor to produce a total cost of ownership across the full first year, with integration, data preparation, training and ongoing operation as separate line items. A vendor who cannot produce that has not scoped the work.

Error Three: Attribution Confusion

The third error is the subtlest. When a process improves after an AI deployment, the improvement gets credited to the AI. Some of it belongs to the process mapping that happened during implementation, which would have produced gains on its own.

Documenting a workflow well enough to automate it almost always reveals redundant steps, unclear handoffs and unnecessary approvals. Removing those is valuable. It is not an AI result, and a vendor who books it as one is inflating their own contribution.

This matters at renewal. A business that attributed process gains to the AI will over-forecast what the next deployment delivers, and under-invest in the process work that produced part of the first result.

Five Questions Before You Commit

1. Do you require baseline measurement before deployment? A vendor who does not ask for your current numbers cannot prove they improved them. A vendor who insists on capturing them is protecting your ability to evaluate their own work, which is the opposite of a sales incentive.

2. What is in your cost model, and what is outside it? Ask for integration, data preparation, training, and first-year operating cost as named line items. The answer tells you whether the quote is a price or an opening position.

3. Can you show results from a business of comparable size and industry? Enterprise case studies do not transfer to an owner-operated firm. The relevant question is not whether the technology works but whether it has worked at your scale, against your kind of workflow.

4. How do you separate AI-driven improvement from process improvement? A vendor with a considered answer here is one who has measured their own contribution honestly. A vendor who has never been asked will say the question does not matter.

5. Will you tie any part of payment to a measured outcome? This is the question that separates confidence from projection. It does not need to be the whole fee. A vendor unwilling to put any portion of it against a number they themselves forecast is telling you how much they believe the forecast.

What the Failure Data Actually Says

Gartner’s survey of 782 infrastructure and operations leaders found 28% of AI projects fully meeting ROI expectations and one in five failing outright, leaving a majority in an ambiguous middle where the project runs and nobody has established whether it matters (Gartner). Gartner also found that 57% of these leaders reported at least one AI failure, and many of them said it came from expecting too much, too fast.

MIT’s NANDA research put 95% of generative AI pilots at no measurable profit impact, while noting that buying from specialist vendors succeeded roughly 67% of the time against about a third for internal builds (MIT, via Fortune). The technology is not the variable. The scoping, measurement and attribution around it are.

McKinsey’s finding that adoption is near-universal while enterprise-wide EBIT impact is rare points at the same place (McKinsey). A vendor ROI projection is a hypothesis. The five questions above are how you find out whether it is a testable one before you pay for it.