How to measure AI ROI and secure financial returns

Gautier Cloix, co-founder and CEO at H, gave a blunt answer on how to measure AI ROI: don't accept productivity as the return. He told companies to demand measurable results and drop vendors that don't deliver them within six or seven weeks.

Ameya Kanitkar, co-founder and CTO at Larridin, told a story that shows why. A CIO surveyed staff and presented a 10 percent time savings to the board. The board turned to the CFO: “Okay, we are saving 10%, let's get those savings back. Let's either cut 10% or let's cut future headcount from 10%.” The CFO rejected the request, according to Kanitkar.

Step 1: Define clear objectives for AI initiatives

Cloix argued that KPIs based on adoption rates or the number of AI use cases deployed are the wrong measures. Companies investing 100 million need to know what they get in exchange and whether they can measure it fast. Efficiency on its own isn't a return, he said. When employees get more efficient, they simply spend more time at the coffee machine.

Step 2: Choose AI ROI metrics tied to business goals

Kanitkar admitted that when he ran developer productivity for 4,000 engineers at LinkedIn, his team fell into the same trap: “And to be honest with you guys, we had bunch of vanity metrics. How many lines of code change, how many PRs, how many commits, things like that.”

Better metrics track an operational change. Brandon Card, founder and CEO at Terzo, said one of the biggest food and beverage companies in the world plans to cut its back office reporting team from 1,000 to 100 people: “They want to lay off 90% of that department to automate reporting, to automate audits, to automate all of these things across finance and the back office function.”

Arvind Mathur of Amazon Web Services reported that Syngenta achieved a 5 percent crop yield increase. Syngenta built agents that give farmers targeted advice using cleaned, semantically structured agricultural data.

Step 3: Know where the value actually comes from

Ambica Rajagopal of Michelin estimated that copilots and personal productivity tools bring only about 20 percent of AI's total value to an enterprise. They matter for building trust in the technology, but “the remaining 80% of the value from AI to an organization actually comes from building custom models that model the internal work processes.”

Colin Jarvis of OpenAI described a semiconductor company where “an engineer at this company probably spends 80% of their time dealing with old technical debt.” Every morning, overnight jobs threw up 1,000 bugs for engineers to work through over the week. The company built an intelligent debug service to do root cause analysis and propose patches.

Step 4: Measure financial impact after implementation

Cloix's advice on vendors was direct: “Kick these companies, including ours, out after six, seven weeks if you don't see any results. You don't need to wait for two years. That would be my advice.”

Step 5: Account for hidden costs and risks

Card said big enterprises deployed Microsoft Copilot directly on top of SharePoint and corporate data without cleansing it first, and got 60 to 70 percent accuracy. “Well, I can tell you there's a reason why Copilot's being handed out for free right now. Because they're not getting great results because the data cleansing process was skipped.”

Terzo is analyzing 36 repositories of corporate data at a Fortune 100 company, and Card said the team found highly sensitive employee data feeding into the models: “We have found so much personal information in these corporate file stores or these repositories, like people's mortgages, people's divorce papers, people's kids' college tuitions, all feeding the LLMs.”

Kanitkar's CIO story is the clearest guide to how to measure AI ROI. A time savings estimate didn't survive contact with the board. Concrete operational changes, like the reporting team Card described or Syngenta's crop yields, do.

To hear more from the people putting AI into production, see the HumanX 2027 program preview.