Your AI Budget Is Being Measured Against The Wrong Number
Shweta Gummidipudi, Chief Information Officer at Twilio.
gettyI get asked some version of this question that most CIOs get asked: Are our AI costs justifying the ROI? Like most of my peers, I’ve spent all year grappling with that question because it rests on a comparison that doesn’t exist yet.
Think about how we got comfortable with cloud spend. It took years, but the industry eventually produced something a finance team could work with. You could look at infrastructure cost as a percentage of revenue, compare it against companies of similar shape and form a reasonable view. That figure was never precise, but it didn’t need to be. It gave two people in a budget conversation a shared reference point.
AI has no equivalent. We’re three years into meaningful enterprise adoption with no credible answer to what a company of a given size should be spending. Every budget conversation right now is a negotiation without a yardstick. That usually means the loudest opinion in the room wins.
The closest thing to a benchmark is descriptive rather than prescriptive. Boston Consulting Group’s January 2026 report found that technology companies planned to increase AI spending to around 2.1% of revenue. That tells you what companies are doing but says nothing about whether they’re right. Given that a 2026 Gartner Inc. report highlighted that 59% of AI initiatives never reach production, averaging the behavior of a group still mostly experimenting is a strange thing to anchor to.
The deeper problem is that revenue is the wrong denominator entirely.
Two companies with identical revenue can be in completely different positions. One has a unified data platform, governance in place and knowledge workers building agents against trusted systems. The other has data scattered across a dozen warehouses, no observability into model usage and pilots that will never scale. If both spend the same percentage of revenue, one is investing, while the other is burning money. The percentage won’t tell you which is which.
What determines whether a dollar of AI spend produces anything is operating maturity; that’s what we should be benchmarking against. There’s no published benchmark for what each stage should spend, which is part of the problem. Let me sketch one anyway because the shape is more instructive than any single number. If the tech sector average sits somewhere around 2% of revenue, here’s roughly how I would expect maturity to spread it.
Legacy is the 20-year-old enterprise that grew through acquisition and never finished integrating. Half a dozen systems that each think they own the customer record, no shared definition of anything, and AI infrastructure that barely exists. Spending heavily here is the most common way companies waste money on AI because there’s nothing underneath to make the spend productive. The money belongs in the foundation, not in models.
AI-assisted is where most enterprises actually sit today. Everyone has a co-pilot license, a few motivated teams have built something useful, and the data foundation is partly there. Spend is moderate and should be climbing because this is the stage where the infrastructure that unlocks the next big win gets funded.
AI-enabled means the data is unified, governance and cost controls hold and a knowledge worker can build an agent against systems the company trusts without filing a ticket. Spend rises meaningfully here, and it should because the infrastructure around it means the money converts into output rather than evaporating.
AI-first is where that capability is mature and the organization itself is built around it. Spend is highest, and so is risk appetite. When an enterprise is truly here, the budgets are big, and the ability to absorb a bad bet is just as big. Very few enterprises are actually here, and many claiming to be are describing an ambition.
The practical consequence is that two companies sitting two stages apart shouldn’t be spending the same amount. The one further along should be spending more. Judged against a revenue benchmark, that company looks undisciplined. Judged against its own maturity, it’s doing exactly what it should be doing.
None of this solves the harder question underneath, which is what AI spend returns. The gap there is stark. A 2025 IBM report found that 29% of surveyed executives said they could confidently measure AI ROI, while 79% believed they were seeing productivity gains. According to a 2026 CIO.com survey, only 19% of responding IT leaders said their AI initiatives had met or exceeded business goals. We feel the benefit and can’t yet prove it in terms finance will accept, and any CIO who claims otherwise is describing impact and calling it return.
This is why we put controls on consumption before setting budget targets. You can enforce a ceiling today, but you can’t set a credible budget until you know what a unit of that spend produces. Plenty of vendors will sell you a framework that assumes otherwise.
When your CFO asks whether the spend is justified, you’re going to feel the pull toward a percentage. Let me save you the trouble: Resist it. Ask instead what stage your organization is in, whether the spend matches that stage and what would have to be true to reach the next one. That conversation is harder and considerably more useful, and you can have it honestly today.
The benchmark will come eventually. It usually does. However, it’ll be built on maturity rather than revenue, and the companies that get there first will be the ones that were honest about where they stood while everyone else compared percentages.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
