Can AI Do the Work Economically?

For the last few years, CIOs, CFOs and CEOs have asked a simple question: Can AI do the work? This includes things like writing code, reviewing contracts, creating customized marketing content and generating data analyses. The answer to the question is increasingly yes.

But now the big question is: Can AI do the work economically? That's a very different conversation.

The excitement over capability is what sold the first wave of GenAI. The cost is what will define the next one, and it’s already exposing flawed assumptions that don’t hold up.

Recently, reports surfaced that Uber burned through its 2026 AI coding-tools budget in just four months after expanding the use of AI coding tools. An AI consultant told Axios that one of their clients ran up a reported $500 million in a single month on one AI platform after failing to set usage limits. The company was never named and the figure is secondhand, so treat it as a cautionary tale, not a documented case. Other CIOs I know have faced similar situations with rapidly rising costs in AI, while choosing not to backfill important workers as an offset. I don’t know whether Uber’s situation turns out to be a trendsetter and that’s not really what interests me. The shift underneath it is. Leaders are starting to put AI through the same scrutiny we’d give any major cost. We’re treating AI as a serious budget line. 

In my previous CIOnews article, I argued that organizations need to establish governance, including managing costs, before AI agents become embedded in critical workflows. Once those digital coworkers arrive, leaders must understand not only what they can do, but also what they cost.

A software engineer earning $150,000 annually may cost an organization more than $200,000 after benefits, taxes, equipment, training, and overhead. That's a meaningful investment, but it's also relatively predictable and so salaries are considered a fixed cost. AI is increasingly a metered utility. Prompts consume tokens, agents consume compute, and the workflows underneath draw on APIs, storage, and infrastructure. That creates a rather awkward dynamic and a far less predictable cost: the more people use it, the higher your bill becomes. And your costs will vary extensively depending on what foundational model and version is being used. This unpredictability makes enterprise AI tools a variable operating cost. 

If an AI coding assistant helps an engineer become 30% more productive, that's a win. If that same productivity gain requires a rapidly growing stream of token consumption, premium model subscriptions, governance tools, and human oversight, the economics become more complicated.

None of this is an argument against AI. I think it will reshape nearly every function in your company. What I am suggesting is that many organizations still can’t show that AI is cheaper than the people they hope to replace.

We need to treat tokens the way we treat payroll. We need to understand who is consuming them, what work is being performed and the business outcomes being created. Most importantly, we have to understand whether the economics make sense.

What CIOs, CFOs, and CEOs should focus on now

1. Stop measuring AI success through adoption metrics. Prompt counts, usage statistics, and active users may indicate interest, but they do not indicate value. Focus instead on what a unit of real work costs: a contract reviewed, a claim processed, a customer issue resolved. Better yet, track the revenue generated.
2. Establish AI FinOps before costs become a problem. You must gain visibility into token consumption, model usage, and workflow-level economics as a core capability, not an afterthought. Build visibility and accountability into your company’s culture, too.
3. Create policy and technology to manage AI spending. Set spending limits your systems enforce on their own, with caps that roll up from the individual to the team to the company, all visible in one place. When spend tops out, it stops automatically. Decide what you track, when you review it, and what happens when limits are reached. 
4. Build a "token P&L" before making workforce decisions. Compare the fully loaded cost of human work against the fully loaded cost of AI, including supervision, governance, rework, and risk. The answer may surprise you. And the cost of building this will be far less than the cost of laying off people, only to rehire them later.
5. Remember that AI doesn't have to replace people to create value. Some of the strongest business cases emerge when AI augments talented employees rather than attempts to eliminate them. A great engineer with AI may create more value than either one independently.
   

This is all less a technological reckoning than an ordinary business one. Nobody checked whether the thing we were sure would save us money was cheaper than the people it was meant to replace, and the answer, run line by line, may turn out to be no. For now, at least. The fanfare over what AI can do is giving way to something more mundane: AI is a utility, one we need to predictably budget for like any other. 

Todd Mazza is the former CTO of Factory Mutual Insurance Company. He’s also held leadership roles at Rockwell Automation, Workday, AECOM, Levi Strauss & Co, MGM Mirage and NBC Universal. Contact Todd: LinkedIn | @ToddMazzaCIO on X