7 Places AI Spending Actually Pays Off

Global AI spending is on track to pass $2.5 trillion this year. What almost nobody can produce is a clean answer to the follow-up question: is it working?

MIT’s Project NANDA study put the number that’s been quoted in every boardroom since — roughly 95% of enterprise AI pilots deliver no measurable impact on the P&L. Morgan Stanley found that by late 2025, only about a fifth of S&P 500 companies could point to a measurable AI benefit at all. IBM’s CEO study landed in the same place: about a quarter of initiatives hit their expected return.

That’s not an argument against spending. It’s an argument about where. The companies in the small group getting returns aren’t buying better models than everyone else. They’re pointing the same tools at different problems.

Here are seven places the money tends to come back, and the question to ask before you sign off on the next one.

1. Work that already has a number attached to it

If a process doesn’t have a baseline, you can’t prove anything changed. That sounds obvious. It’s also the single most common reason a pilot dies quietly after two quarters — the team can show the tool works, but nobody can show the business moved.

Start where you already keep score. Claims processed per adjuster. Days sales outstanding. First-response time. Cost per ticket. Anything with a year of history behind it gives you a before and an after that a CFO will accept.

Ask: what metric will this move, and what has that metric done for the last four quarters?

2. The plumbing underneath the model

The model is the cheap part. It’s a line item you can swap out next quarter for something better. What you can’t swap out quickly is a permissions system that knows who’s allowed to see what, a data layer that agrees with itself across departments, and a record of what’s actually true right now.

Companies that skip this end up with a very fluent assistant confidently working from a document somebody deprecated in 2023. Then trust collapses, adoption drops, and the license renewal gets cancelled. The infrastructure spend feels like it’s delaying the fun part. It’s the part that decides whether the fun part survives contact with real work.

Ask: if this tool gives someone a wrong answer, how would we find out?

3. High-volume, low-stakes work with a human at the end

The best early wins are boring. Drafting the first version of a proposal. Triaging inbound tickets into the right queue. Summarizing a two-hour call into a page. Pulling structured data out of unstructured documents.

These share a shape: they happen constantly, they eat hours from people you pay well, and a mistake gets caught before it reaches a customer. That last part matters more than people expect. It’s what lets you deploy fast without wrapping the whole thing in six months of risk review.

Ask: if this output is wrong, who catches it, and what does the mistake cost?

4. Finding what the company already knows

Every organization past a certain size pays a quiet tax. Somebody solved this problem eighteen months ago, wrote it up carefully, and nobody can find the document. So the work gets done again from scratch.

Internal search and retrieval is unglamorous, doesn’t demo well, and has a short path to real savings. It also compounds. The more institutional knowledge becomes findable, the faster new hires get useful and the less senior time gets spent answering questions that were answered before.

Ask: how many hours a week do our people spend looking for something that already exists?

5. Getting your people to actually use it

There’s a gap between seats purchased and seats used, and it’s usually wider than anyone wants to report. A license nobody opens is pure cost. A license someone uses badly can be worse than that.

The spend that fixes this isn’t more software. It’s time. Structured training tied to the specific work people do, a few internal examples of what good use looks like, and someone visible who’s willing to answer dumb questions. Companies that fund this get returns from tools they already bought. Companies that don’t keep buying new tools hoping one of them will finally stick.

Ask: what percentage of the licenses we bought last year get opened weekly?

6. One workflow all the way through, instead of ten pilots

Ten pilots feels like momentum. It’s usually the opposite. Each one demands attention, none gets enough to reach production, and the portfolio produces a lot of slide decks and no change in the numbers.

Pick one workflow. Take it end to end — intake, processing, handoff, exception handling, the ugly edge cases nobody puts in the demo. Getting one thing fully into production teaches you more about your own organization than ten experiments will, and it gives you a repeatable pattern for the next one.

Ask: which of our current pilots would we bet the quarter on, and why aren’t we?

7. The ability to prove any of this

Measurement gets treated as overhead. It’s closer to a financing cost. Citi found that debt markets have started charging a spread premium to companies classified as AI adopters versus AI enablers — meaning the market is already pricing the difference between spending on AI and demonstrating return from it.

Build the ability to answer the question before you’re asked it in a board meeting. Task-level measurement, a before-and-after you didn’t construct after the fact, and an honest accounting of what got shut off. If you can’t show what the spend replaced, you didn’t save anything. You added a line item.

Ask: if the board asks for our AI return next quarter, what document do we hand them?

The pattern underneath all seven

None of these are about the technology. They’re about whether the organization around the technology is set up to convert it into something a CFO can see.

The 5% that get returns tend to do the same unglamorous things: they pick work that was already measured, they fix the data underneath before they build on top of it, they finish one thing, and they keep receipts. The other 95% buy the same tools and skip those steps.

The next AI decision in front of you probably isn’t which model to license. It’s whether you’ve built the conditions where any model would pay off.