AI Spending Faces A Tougher Test As Productivity Gains Fail To Lift The Bottom Line

Enterprises have put AI in front of their workforces, and their income statements have barely moved. McKinsey's 2026 survey found that 80% of respondents say AI has improved their individual productivity, yet only 37% attribute any EBIT impact to it. In PwC's 2026 Global CEO Survey, 56% of chief executives reported no significant financial benefit so far. Post-mortems tend to blame the model, the data or the integration, but the decision that shapes the outcome comes earlier, when a team picks the problem AI will be pointed at. With spending now under review, that choice is where leaders have the most leverage.

Siva Puvvada, Practice Director of BPM & AI at Truviq Systems, has spent more than 22 years delivering enterprise applications for Global 500 firms in financial services and insurance, including at HCLTech and EY. He's now building Truviq's AI practice with a focus on AI governance. In his conversations with customers, he keeps meeting organizations that decide where to use AI before deciding what problem it should solve.

“Nobody is asking the question: is it the right use case for me to apply AI?” Puvvada says. Everybody wants to brand themselves as using AI, in his telling, and that pressure produces project lists built for visibility. Each project then collects infrastructure, integration and training costs before anyone tests whether it moves a business number. The rest of his argument, from who approves a project to what the business case leaves out, follows from that one choice.

Saved hours and slick demos don't reach the bottom line

Pointing to McKinsey and PwC's enterprise ROI reports, Puvvada says AI gets used mostly for personal productivity. “Less than half of the organizations are actually seeing the ROI,” he says. “Where it's actually being used is with the productivity.” But saved time doesn't automatically become a return. “As a developer, by using a coding agent, if I save two hours in a day, it doesn't add up to any value at the end of the day,” he explains.

In the funding pattern he describes in his written framework, a short demo secures production budget, and the team then spends that budget looking for a business fit. In production, he writes, edge-case exceptions, latency and integration friction can make probabilistic automation far costlier than the manual process it replaced. Gartner's survey of 782 infrastructure and operations managers found that only 28% of AI use cases fully succeeded, with many failures traced to unrealistic expectations and poorly scoped projects.

His advice begins with a rule: “Don't use AI for the sake of using AI.” “Identify the use cases, business problems that the organization is currently having, and then see where AI can be used,” he says. In a process such as onboarding or KYC, that means finding the friction points that cost time or money and applying AI only there. His framework adds a test that ties the initiative to an audited metric such as unit cost per transaction, and a project whose operational line item hasn't moved by day 90 stays on the whiteboard.

A governance council can't rescue a bad use case

Getting a use case approved takes more than one team. Puvvada's sequence starts with business owners defining the problem. Data, infrastructure and architecture teams then confirm that the data exists and the systems can carry agent workloads at scale, and compliance, legal, finance and leadership approve the controls and the spend, including vector databases and tokens. “When the use case is approved, it has to touch all the departments that have a stake in the initiative,” he notes.

That breadth is why a governance council tends to arrive late, since it reviews what it's handed. “You can have AI governance, but if you're not selecting the right use cases, there is very little you can do within a governance council,” Puvvada says. “It's like doing things after you selected a bad use case that's causing problems for the organization.” His framework adds that committees can't own software, so accountability belongs to a single business-unit head whose scorecard and budget absorb the results.

Business cases leave out adoption and assume everyone shows up

The cost lines that usually make it into an AI budget are data preparation, infrastructure and integration. “One thing that I see missing many a time is the change management cost,” Puvvada says. Without enablement for the people who are supposed to use the tool, he adds, “the ROI won't be there irrespective of how good that particular initiative would be.” Buying a platform doesn't remove that cost either. “It's just that we're removing the build part of it,” he says.

The benefits side of the business case has its own bias. “We get into the optimistic side with respect to the benefits, assuming that everybody will be leveraging the initiative that's being deployed in the organization,” Puvvada says. He borrows the discipline of capital allocation, with base, conservative and optimistic cases plus an estimate of what the business unit loses if the initiative never gets built. In one conservative scenario he sketches, only a quarter of the organization uses the tool. He calls the EBITDA test the minimum foundation for any organization. “Only leverage AI where it needs to be leveraged, and that initiative should impact your EBITDA at the end of the day,” he says.

Pilots and token budgets bring the use-case question back

Approval doesn't end the risk, because a pilot rarely carries the controls production needs. “The pilot may not have all these checks and balances, which are very important when you need to take this to production,” Puvvada says. His answer is a fallback. “We can't go in and expect that it would do the right thing all the time,” he says. “So there should always be a fallback.” In the best use cases he's seen, the traditional process stays in place behind the agent while accuracy improves.

Cost pressure pushes in the same direction. Early in the year, some organizations exhausted budgets they’d scheduled for the entire year. Uber’s CTO said in April that AI coding tools had used up its 2026 AI budget, and leadership teams are now asking where the tokens go. “Organizations are now going back and checking where exactly AI is being leveraged and questioning the teams why they have leveraged AI for certain things where maybe it's just a rule engine that needs to be in place,” he notes. Engineers apply the same filter to models, since a question about a company's leave approval process doesn't need a large one. “If you just ask me what is the leave approval process in my company, I don't need a very large model to solve that particular request.”

Spending is heading toward the test Puvvada applies at the start. “Organizations are now going to tie all their spends on AI to the actual value that they're going to deliver to their boards,” he says. Anything that can't be tied to the bottom line or the top line is R&D in his view. “We're just doing a science project,” he adds.