Companies That Watch Their AI Bill Are 5x More Likely to See Returns

KPMG surveyed 2,145 senior leaders at organizations above $50 million in revenue, across 20 countries. Almost half, 49%, had scaled back, narrowed, delayed or paused agent deployments when operating costs ran past the value produced. Only 7% reported established returns.

The 7% is the number everyone will quote. It’s the least useful one in the study.

The useful number is the split underneath it. Among organizations with full visibility into what their AI actually costs to run, 15% reported established ROI. Among those without that visibility, 3% did. Same technology, same year, five times the hit rate, and the variable separating them was simply whether anyone could see the bill.

Only 35% said their operating costs were fully visible and actively monitored.

That tracks with what’s happening to agent projects generally. Gartner now expects more than 40% of agentic AI projects to be cancelled by the end of 2027, and runaway cost sits among the top reasons, usually showing up as a visibility problem before it shows up as a spending one.

Rob Fisher, Global Head of Advisory at the firm, drew the conclusion the industry has been slow to:

“AI is now as much a financial management priority as it is a technology one.”

Why the bill is hard to see

Prashanthi Kolluru, founder of KloudPortal, wrote up the mechanism in Forbes on August 10, and it’s the clearest short description of the problem we’ve read. The core error is applying software budgeting to something that doesn’t behave like software:

“A production agent is constantly at work: retrieving enterprise data, evaluating prompts, invoking models, connecting to APIs, applying business rules, logging activity for governance and, in many organizations, routing sensitive responses through human review.”

None of that is a license fee. It’s a meter, and it runs faster as the thing succeeds. Kolluru names the trap precisely:

“As business teams gain confidence, they start asking for more…Every one of those requests makes sense on its own, but together, they mean more model calls, larger context windows, more retrieval operations and heavier governance requirements.”

That’s the shape of the problem for a marketing team specifically. Nobody approves a budget overrun. What happens is the content agent works, so someone points it at the blog archive, then at the email back catalog, then at every product page, and each of those decisions is obviously correct on its own.

The gap she identifies in most budgets is the running cost, not the build cost:

“Most technology budgets account for infrastructure, licensing and implementation. What they tend to leave out is everything it takes to keep AI agents running responsibly: monitoring output quality, evaluating model performance, updating prompts as business processes shift, maintaining data connections, enforcing governance policies and continuously measuring outcomes.”

The context that makes this a marketing problem now

This lands the same week Google started giving away analysis that used to be billable, and a month into a stretch where marketing teams have been absorbing work rather than adding headcount. The direction of travel is more agent-shaped work sitting inside marketing, arriving through free features rather than through a procurement cycle that would have forced a cost question.

Free to switch on is exactly how you end up in the 65% with no cost visibility.

Update, August 17. A second dataset points the same way, and this one asked marketers directly. TransUnion commissioned UTA Advisory to survey 100 senior marketing and technology leaders at major US brands, and published the results on August 5. Of those 100, 89% expect AI marketing investment to rise over the next 12 to 24 months and 64% are confident of hitting their AI goals. Against that: 48% say they have enough visibility into platform-level AI to make optimization calls with confidence, and 36% rate their data and process readiness as high.

That’s the KPMG finding restated in marketing’s own vocabulary. Confidence and spending intent run well ahead of the ability to see what the spending does. Matt Spiegel, EVP of TruAudience growth strategy at TransUnion, put it as a sequencing problem:

“Marketers are increasingly confident in AI’s ability to drive business results, but many are still working to build the foundations needed to scale it effectively.”

Two limits on that sample. A hundred leaders at major US brands is a small, large-company panel, so it says little about a 20-person team. And TransUnion’s release does not disclose when the fieldwork ran, which means the survey could be considerably older than its August 5 publication date.

Two caveats on the KPMG data. KPMG’s fieldwork ran April 28 to May 25 and published June 24, so this is a spring read, and it covers agent deployments broadly rather than marketing ones. Kolluru’s piece is a Forbes Council post, which is a contributed format rather than reported journalism, and she runs a company that sells AI implementation work. The mechanism she describes is verifiable against the KPMG numbers, which is why it’s here, but the incentive is worth naming.

The move

You probably can’t get full cost visibility this quarter. You can get the thing that matters most of it.

Before switching on the next AI feature, write down what it costs to run at 10x current volume, and who sees that number monthly. If nobody owns it, you’ve just joined the 65%, and the study says that’s where the returns aren’t.