Most companies cannot measure AI ROI because nobody owns the question. Tools are owned by teams, pilots are owned by departments, and the return is owned by no one. The fix is not a bigger budget, and it is not another platform. It is a three-column inventory, built in two weeks, that lists what you spend, who is accountable for each initiative’s business result, and what each one measurably produces.
That is the short answer. The rest of this article explains why the gap exists, what it quietly costs, where AI spending hides when you go looking for it, how to build the inventory, what should count as a measurable result, and the objections that will come up in the room when you propose doing this. It ends with a five-minute step you can take today.
Why can’t leadership teams answer what their AI investment returned?
Because the spending and the question live at different altitudes, and they arrive on different clocks.
AI spending accumulates bottom-up. A subscription here, because a team wanted to try it. A pilot there, because a department head saw a competitor mention something similar. An API bill that grows a little each month, because usage grows and nobody has set a ceiling. Each purchase made sense locally, was approved locally, and is reported locally, if it is reported at all.
The return question arrives top-down, and it arrives rarely. It shows up at budget time, in a board meeting, or when a CFO notices that the software line has grown and wants to know why. By then, the spending is twelve months of accumulated local decisions, and the person being asked the question was not in the room for most of them.
So the answer comes back as a list. We rolled out these tools. We have pilots running. The teams are experimenting. Everyone present recognizes that a list of activities is not an answer, and the meeting moves on with the question still open, which is the worst outcome available: the question was asked, nothing changed, and it will be asked again next quarter with more spending behind it.
The pattern is structural, not personal. In most companies, ownership of an AI initiative means responsibility for running the tool: managing seats, onboarding users, and fielding complaints. It does not mean accountability for a business result. Those are different jobs, and the second one is usually unassigned. Until someone owns the result, the return stays unmeasured, and the question stays unanswerable no matter how competent everyone in the room is.
What does unmeasured AI spending actually cost?
More than the subscriptions, because it corrupts the decisions around them. Three mechanisms do the damage, and each one compounds the others.
First, budget allocation defaults to impressions. When no initiative can prove its return, funding decisions fall back on how each initiative is presented. The team with the most confident champion, the most polished demo, or the most fashionable use case wins the allocation. Confidence is not correlated with results, and in my experience the correlation sometimes runs the other way: the teams doing careful, measurable work tend to present it with caveats, and the caveats lose to the demo.
Second, renewals run on inertia. Canceling a tool feels like admitting the purchase was a mistake, so tools survive on the absence of evidence against them rather than the presence of evidence for them. Nobody is measuring, so the evidence against never materializes, so the renewal goes through. A tool can survive three renewal cycles this way while being opened by four people a month.
Third, and most expensive: the initiative that would produce real revenue sits unfunded. When nothing is measured, a genuinely valuable proposal cannot distinguish itself from the incumbents. It cannot prove it would outperform them, because they have no performance on record to outperform. The cost of unmeasured spending is not only what you pay for. It is what you never build, and that cost appears on no invoice, which is exactly why it persists.
Where does AI spending actually hide?
The first surprise of building the inventory is usually the spend column itself. Companies that expect to find a modest number often find a materially larger one, because AI spending hides in places the software budget line does not show.
It hides in seats: licenses provisioned during a rollout and never reclaimed when people changed roles. It hides in usage-based API costs that grow monthly without a purchase decision ever being made. It hides inside platforms you already pay for, where an AI feature tier was added to the renewal and absorbed without a separate line. It hides in duplication, where two departments independently bought tools that do the same thing under different names. It hides in expense reports, where individuals pay for AI subscriptions on cards because procurement felt slow. And the highest hidden cost is not a subscription at all: it is the people-time consumed by pilots, the hours of your most capable staff spent evaluating, configuring, and attending meetings about tools that never reached production.
None of these items is scandalous on its own. Together they are why “What do we spend on AI?” is a research question rather than a lookup and why answering it is the first genuinely useful act of measurement.
How do you build the AI ROI inventory?
Three columns, one page, one owner, two weeks. Resist every temptation to make it bigger than that, because the failure mode of measurement projects is scope: the inventory that tries to become a dashboard never ships.
Column one, spend. Every AI tool, platform, subscription, API, and initiative with its actual annual cost, including the hidden categories above. Estimates are acceptable where invoices are ambiguous; blank cells are not.
Column two, owner. A named person accountable for each initiative’s business result. Not the administrator who manages seats; the person who would answer for the outcome in a budget meeting. If no such person exists for a row, write “none,” and do not fix it in the spreadsheet. An ownerless initiative is a finding, and findings should stay visible until they are resolved in the organization, not in the document.
Column three, measurable product. What each initiative demonstrably produces. Revenue attributed, cost removed, hours saved, or “nothing yet.” “Nothing yet” is a legitimate entry and often the most valuable one on the page, because it is honest, and the entire exercise exists to replace impressions with honesty.
Then hold one meeting. The inventory owner walks the page, row by row, with the people who approve budgets. No slides. The decisions that fall out are usually obvious once the page exists: consolidate the duplicates, cancel the unopened, assign owners to the ownerless, and give the “nothing yet” rows a date by which they will have a number or a shutdown.
Companies that complete this exercise usually discover they need less AI than they thought, connected better than it is. The problem is rarely too few tools. It is assets that were never wired together or measured.
What should count as a measurable result?
A number that someone outside the initiative would accept. That is the whole test, and it disqualifies most of what currently passes for AI results.
Revenue-facing initiatives should show attributed pipeline or closed revenue, with the attribution logic stated plainly enough to be challenged. Cost-facing initiatives should show a cost that existed before and does not exist now: a vendor replaced, a contractor not renewed, a process retired. Time-facing initiatives are where honesty goes to die, so hold them to a harder standard: hours saved only count when they are converted into something real, headcount redeployed, output increased, or backlog reduced, or they should be labeled soft and weighted accordingly. “The team feels faster” is a sentence, not a result.
Adoption metrics, seats active, and prompts sent are operating data, not returns. They tell you whether a tool is being used, which matters, and nothing about whether the use produces value, which matters more.
The objections were answered plainly
“We are investing to learn. Demanding ROI this early kills experimentation.” The inventory does not demand ROI; it demands honesty about where each initiative stands. “Nothing yet, learning, review in Q3” is a fully acceptable row. What the inventory kills is the permanent pilot: the experiment with no owner, no date, and no defined result that would end it. Experiments deserve protection. Experiments without end conditions are just spending with better branding.
“Measurement overhead will slow the teams down.” One page and one meeting is the entire overhead, and it removes more work than it adds, because the current state, in which every initiative periodically defends itself ad hoc when someone asks, is itself a measurement system, just an expensive and unreliable one.
“Our AI value is intangible: better decisions, faster teams.” Some of it genuinely is, and the inventory has a place for that: labeled soft, and not used to justify spend on its own. The moment intangible value becomes the primary justification for a line item, that line item has stopped being managed. Intangibles are a bonus on top of measured value, not a substitute for it.
“Isn’t this the CFO’s job?” The CFO can own the spend column tomorrow. Columns two and three require judgment about what each initiative is for, and that judgment lives with whoever sponsors AI in the business. In practice, the inventory works best with a business owner and CFO support, which is also why it fits on one page: it has to be readable by both.
Why two of your four leaks will not appear in this inventory
The inventory measures what happens inside your company, and that is only half the picture. Two of the places AI investment leaks are structurally invisible to internal measurement.
The first is AI visibility. When your buyers ask ChatGPT, Claude, Perplexity, or Gemini for vendors in your category, those engines answer, and either you are in the answer, described correctly, or you are not. Nothing about that shows up in your analytics, your CRM, or your inventory, because the buyer who reads the answer and chooses someone else never contacts you. We wrote a full issue on what that failure looks like in practice, including a documented court case, in Cited, and Wrong, Is Worse Than Uncited.
The second is data readiness. Your proprietary data and internal knowledge are the one input competitors cannot buy, and their condition determines whether every initiative in your inventory runs on your knowledge or on generic knowledge. A degraded data foundation quietly caps the return of every row on the page, and no row will say so.
Measuring those two takes different instruments than a spreadsheet, which is why they are two of the four areas in the scorecard below.
What is a good first step this week?
Score yourself before you inventory anything. We built the Executive AI ROI Scorecard for exactly this: 20 statements across the four areas where AI investment compounds or leaks, spend and return, AI visibility, data readiness, and workflow implementation are scored in five minutes with one calibration rule that keeps the number honest: a practice only scores high if it is documented and would survive the person who runs it leaving.
The total tells you how far along you are. Your lowest section tells you what to fix first, and it is usually not the area anyone has been discussing in meetings.
Get the Executive AI ROI Scorecard
Elizabeta Kuzevska is Co-Founder of Revenue Experts. Revenue Experts is an AI consulting and implementation firm with a rare mix of executive experience and technical depth. Our founders have built companies, run large-scale marketing and sales teams, and developed winning GTM strategies.
We help established B2B & B2C companies activate proprietary data; improve AI visibility through AEO; automate high-value workflows; and build secure RAG systems on two paths: cloud-based on managed platforms for speed & scale or sovereign inside your own infrastructure, where your data never leaves your environment and the models, knowledge bases, and systems built on it are assets you own outright, all measured against revenue, efficiency, and ROI.