Money wasted on AI is climbing because almost no one is measuring it

AI costs keep moving in the wrong direction, and the reason is uncomfortably simple. Most teams can't see what their AI tools cost, so they can't tell when the cost stops being worth it.

Flexera's latest report puts figures on it. In the Flexera 2026 State of ITAM Report, which surveyed 512 technology professionals, only 31% of organisations reported having accurate visibility into their AI software, while 59% said wasted AI spending had increased year on year (Source: Flexera, 2026 State of ITAM Report). Costs are rising, and the people responsible for them are mostly working without a clear view.

The visibility gap is the root cause, not a side effect#Link to this section

It is tempting to treat waste and poor visibility as two separate problems. They are the same problem. You can't cut waste you can't see, and you can't defend costs you can't measure.

Flexera also found that complete IT asset visibility has dropped to 36%, which it attributes to growing complexity across AI, SaaS and cloud environments (Flexera). AI makes this harder because it doesn't sit neatly in one category. It spans cloud, SaaS and devices while adding new layers of models, agents and data, each with its own economics (Flexera). A per-seat licence is easy to reason about. Token-based billing that scales with every prompt is not.

Uber is a clear example of what happens next. The company used up its entire annual AI budget in roughly four months, which its chief technology officer revealed in April, largely because agentic coding tools bill by the token rather than a flat monthly licence (Source: TechCrunch). The response was a monthly cap of US$1,500 per employee for each agentic coding tool, tracked through an internal dashboard that employees can exceed with permission (TechCrunch). That is a sensible control. But a cap limits the cost; it doesn't measure the value.

What visibility actually requires#Link to this section

Seeing your AI costs is the first step, and most organisations stop there. Invoices, seat counts and token dashboards tell you what you paid. Whether the payment produced anything is the question around them. Real visibility needs three layers, and each is measurable.

Cost per active user, not cost per licence. Start with what you pay, then divide by the people who actually use the tool in a given month, not the people who were assigned a seat. If you bought 500 seats and 180 people use the tool weekly, your effective cost per active user is nearly three times the number on the contract. The gap between seats purchased and seats used is often the largest single pool of waste, and it is invisible on any invoice.

Adoption depth, not adoption presence. A login is not usage. Track how often people reach for a tool for real work, whether that has grown or stalled since rollout, and which teams have quietly abandoned it. Flexera found that 45% of organisations are now actively looking for unapproved, or shadow, AI use (Flexera), a sign that official adoption numbers and actual behaviour have drifted apart.

Perceived impact against paid cost. This is the layer almost no one has. Ask the people using the tool whether it saves them time, on what tasks, and how much. Set that self-reported value beside the bill. When a tool costs a lot and the people using it can't point to what it changed, you have found waste that no cap will fix.

What good looks like is a simple, defensible ratio: cost per active user, trending flat or down, beside a steady or rising measure of time saved and confidence from the people doing the work. When those two lines diverge, costs rising while impact flattens, you have your answer before the budget runs out, not four months after.

Technology, process or accountability#Link to this section

So which is the real problem? The likely answer is that the technology exposed a gap that process and accountability were always meant to fill. Token billing didn't create waste. It removed the flat-fee cushion that hid it. The fix is not another dashboard of raw usage. It is a regular, lightweight measure of whether the spending is working, owned by someone who has to answer for the number.

Measuring impact now, while the costs are still climbing, is what makes next year's figures something you can explain.