Rob Versaw, Product Strategy at Dynatrace, bridging innovation and business impact.getty​In April, Uber's chief technology officer told The Information (via Forbes) the company had spent its entire 2026 budget for AI coding tools in four months. Roughly 5,000 engineers were using the program, with some "power users" running up to $2,000 a month in tokens and one internal demo burning $1,200 in two hours. The tools were clearly useful. The problem was that usage scaled faster than governance.Most teams will ask what AI costs. But I've found the better question is what that spend is actually earning.The predictions are already in. AI cost will define the conversation for the rest of the year, and the early evidence supports it. Agentic tasks can consume five to 30 times more tokens than a simple query. Budgets set in the old per-seat world no longer hold, because the meter now runs by the token, not the license, and it can accelerate quickly. By June, Uber had capped individual spending at $1,500 a month, and its own chief operating officer conceded it was hard to draw a line between token spend and features customers actually notice.​But cost is only a problem where value is not being realized. A workflow that moves a real commercial outcome earns its tokens. A thousand unmanaged experiments may create activity, but they rarely create measurable value. Deloitte's 2026 enterprise survey found 74% of organizations hoping AI will grow revenue, while only 20% are doing it today. Over one-third are already banking efficiency gains. For many organizations, revenue impact is still more aspiration than evidence. The risk is not spend itself; it is spend that does not compound into measurable value.​Australia has quiet proof that the gap can close. The Commonwealth Bank is putting roughly $2.4 billion a year into technology and capability, and that figure says little. What matters is where the money lands. At its first-half results in February, CBA tied AI to specific lines: Fraud losses down 20% year on year, driven by systems scanning more than 20 million payments a day, and a generative tool called Compass that has handled over 500,000 banker queries and lets frontline staff answer complex questions about three times faster. Chief financial officer Alan Docherty's framing was plain: Productivity and better customer outcomes feed top-line revenue, which justifies the next round of investment. The investment is easier to defend because it is tied to named workflows.​The second question is harder: Is the AI working from our context or guessing from the general internet? Harrison.ai, the Sydney medical-imaging firm still better known within healthcare than outside it, is instructive. Its radiology models were trained on more than a million clinical studies and 550 million expert annotations drawn from over 140 radiologists. That proprietary data is the advantage competitors cannot easily replicate. Independent testing has shown its chest X-ray tool lifting diagnostic accuracy by 45%, and more than half of Australia's radiologists now have access to it.I have seen this discussion from both sides—the team asking for the budget, and the leadership asking what it bought. So before rebuilding the budget, I would ask three questions:1. Which few workflows move revenue, retention or margin, and are we concentrated there or scattered across experiments nobody owns? CBA's answer is fraud detection and frontline banking. The principle scales both ways. Splose, an Adelaide allied health platform that raised $46 million in February, charges clinicians a separate fee for its AI tier, and they pay it. Willingness to pay is the cleanest signal that a workflow earns its keep. What is your answer? If you cannot name three, you may be funding activity rather than outcomes.2. Is the AI working from our data, our customers and our context or from the open internet? Mindpeak's advantage is not simply the model; it is the specialist pathology information, clinical workflow context and regulatory discipline that others lack and cannot quickly replicate. If your AI is guessing from the general web, you are paying premium prices for generic output that a rival can buy just as cheaply.3. Does every workflow carry a named owner and a visible cost before the invoice arrives? Uber discovered its number after the fact, governing it after the bill landed. Australia Post shows the disciplined version. Working with BRX, it pulled a 30% efficiency gain out of creative production, put governance and audit trails in before it automated, moved the contract from hours to outputs, and reinvested the saving rather than banking it as a cut. Same technology, opposite posture. Token usage does not wait for budget cycles.If those three questions are answered well, the cost conversation becomes much easier. Get them wrong, and no amount of optimization will make the economics work.None of this requires a new management theory. Anyone who has run a value stream map will recognize the move: Trace the work from customer request to delivered outcome, and treat every step that adds no value as waste, whatever its price. It is the same discipline a Woolworths category manager or an ASX 200 CFO already applies to any other input: tie the spend to an outcome, put a name against it and measure whether it compounds. AI has simply made that discipline urgent, because the cost is now variable, granular and quick to run away from you. Map the value stream and the unowned experiments show up for what they are: recurring spend without a clear value case.The second half of 2026 will separate the teams funding growth from the teams funding experiments. The budget conversation is where that difference becomes visible. The real question is not what AI costs. It is whether you can point, workflow by workflow, to what it earns.Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?