The AI ROI question is coming.
Most finance teams don't have an answer.
This quarter, finance leaders are going to keep hearing a version of the same question on their earnings calls: you’ve spent two years investing in AI. So what has it returned?
For a lot of teams, the honest answer is a pause. Not because the AI isn’t working, but because the work it was pointed at was never the kind you can count.
That pause is about to get expensive. Through 2024 and 2025, saying “we’re investing in AI” was enough on its own. It signaled ambition, and ambition was the currency. In 2026, with Q2 results rolling in and analysts running the same line of questioning across every sector, ambition isn’t the currency anymore. A number is.
The question most AI can't answer
The problem usually isn’t the technology. It’s where the technology was deployed.
Most finance teams bought general-purpose AI and pointed it at diffuse “productivity”: a faster way to draft an email, summarize a document, or get a first pass at an analysis. Useful, often. But the return is unattributable. “The team feels a bit faster” doesn’t go on a slide next to a dollar figure, and it doesn’t survive a follow-up question.
There’s a category difference hiding underneath that. Generic AI applied to finance is not the same thing as AI built for how finance work actually gets done. The first makes individuals marginally quicker at tasks. The second takes measurable operational work off the team’s plate entirely, and measurable is the whole point.
What "measurable" actually looks like
The finance teams who will answer the ROI question cleanly all did the same thing: they pointed finance-native AI at the operational layer, the high-volume, rule-bound, repetitive work where the outcome is a real, countable number.
Accounts payable is the clearest example, because almost everything about it can be counted: how many invoices moved through without a human touch, how accurately, how much faster the books closed, how much capacity came back.
Here’s the part most AI conversations skip. The materiality isn’t in any single invoice. It’s in the volume. At enterprise scale, a small improvement repeated across every invoice compounds into a number that shows up on the P&L.
So picture it. A top-tier logistics operation that has driven the handling time on an invoice down to about a minute and a half, with roughly three of every four flowing through untouched. A multi-entity property manager clearing more than 171,000 invoices in a single quarter, over 80% of them no-touch. An industrial distributor that trimmed its monthly close from four days to three while nearly nine in ten invoices never needed a person. A finance team whose no-touch rate started in the low single digits and climbed past 40% on its own, because the system kept learning its own data.
Those aren’t hypotheticals. They’re real finance teams, running this today. And the reason each one lands is simple: it’s a number a CFO can say out loud on a call, because the work behind it was countable to begin with. That’s the difference between AI you can point to and AI you can only talk about.
Why the right AI compounds
There’s a second reason finance-native matters, and it shows up over time rather than on day one.
Generic automation is configured once and then slowly drifts out of date. AI built for finance operations does the opposite. It learns from the operational data running through it and gets more capable the longer it runs. The finance team above, whose no-touch rate climbed past 40% on its own, didn’t reconfigure anything; the system learned the shape of its own invoices and vendors. That isn’t a tool you bought. It’s infrastructure that improves while you sleep.
For the ROI question, the distinction matters: a one-time efficiency bump is hard to defend a year later. A return that compounds is a story that gets stronger every quarter.
The mid-year gut-check
Half the year is gone. Before the next board meeting, three questions are worth asking of every AI investment on the books:
- Can I attach a defensible number to this (hours returned, accuracy, close days, headcount I didn’t have to add)?
- Is the return concentrated in countable operational work, or spread across vague “efficiency”?
- Is this AI built for finance, or general-purpose AI pointed at finance and hoping?
The teams that can answer those didn’t necessarily buy more AI. They bought the right kind, and they pointed it out at work, which produces evidence.
The real return
The honest version of the return isn’t only about cost. The reason any of this matters is what it frees the team to do. When the processing layer runs itself, accurately and with the audit trail to prove it, the people who used to key invoices spend their time on judgment, analysis, and the decisions that move the business. That is the shift worth defending on a call: not that finance got cheaper, but that finance moved from processing to strategy.
The AI that can’t be measured will have a hard quarter. The AI pointed at the work that counts will have a number, and a story that keeps getting better.
.svg)
