AI cost is starting to move out of the background. For a while, many organisations could treat AI as experimentation: a few licences, a few pilots, some team-level testing and some useful shortcuts. That phase has not disappeared, and it shouldn’t. But the cost profile is changing.
Gartner forecasts worldwide AI spending will reach $2.59 trillion in 2026, up 47% year on year. That figure needs careful handling because Gartner says much of this spending is dominated by vendors and hyperscalers, rather than direct enterprise operating budgets. Even so, the direction is clear. AI is becoming a much more visible commercial category.
The practical question for leadership teams is no longer whether AI should be explored. It is whether the organisation can see enough of its AI usage to understand where money, effort and attention are actually going. That is not only a finance problem. It is an operating visibility problem.
The invoice is only the visible cost
The visible cost of AI is usually the easiest part to understand. Licences, cloud usage, infrastructure, model access, consulting support and training budgets may all become significant, but they can usually be found somewhere. They appear on invoices, contracts, budgets or procurement records.
The harder costs sit inside the work itself. Two teams pay for overlapping AI tools because neither knows what the other is using. A pilot continues after the original sponsor has moved on because nobody has made a clear decision to stop it. A support team saves time drafting customer replies, but managers spend more time reviewing tone, accuracy and context.
A customer success team starts using AI summaries, but different people trust those summaries in different ways. A manager reviews an AI-assisted customer update without knowing whether it was built from notes, copied from a tool, edited properly or sent with unchecked assumptions.
None of this necessarily appears as a clean AI line item. It appears as rework, review effort, duplicated spend, poor handoffs and inconsistent customer experience. It appears as work that looks faster in one team but creates more effort somewhere else.
That is why the cost of AI isn’t just what shows up on the invoice. It’s also what happens when usage spreads faster than ownership, review standards and evidence of value.
Poor foundations make AI more expensive
AI doesn’t remove weak operating foundations. It exposes them.
IBM reports that over a quarter of organisations estimate they lose more than 5 million dollars annually due to poor data quality, with 7% reporting losses of 25 million or more. That matters because AI depends heavily on the quality, context and reliability of the information it works with.
If the data is inconsistent, duplicated, incomplete or poorly understood, AI doesn’t magically solve the problem.
It can help people move faster with the wrong input. It can produce more polished versions of flawed assumptions. It can spread weak data into more places, and it can make a poor process look more efficient than it really is.
This is one of the hidden cost traps. A business may think it’s paying for AI, when in reality it is also paying for unresolved data issues, unclear workflows, inconsistent review standards and weak ownership. The tool is new. The operating problems are often not.
AI adoption is spreading faster than operating control
McKinsey’s 2025 State of AI report found that 88% of respondents say their organisations use AI in at least one business function. That sounds like progress, and in many cases it is. But McKinsey also found that most organisations have not yet scaled AI across the enterprise, with roughly only one-third reporting that they have begun to scale their AI programmes.
That gap matters. AI usage is becoming common, but enterprise-level value is still uneven. McKinsey also reports that only 39% of respondents see enterprise-level EBIT impact from AI, which reinforces the distance between activity and measurable commercial return.
This is where leadership teams need to be careful. Activity can look like adoption. Adoption can look like value. Value can look obvious until someone asks where it is actually being measured.
Inside the business, the picture is often less tidy. One team uses AI to write first drafts. Another uses it to summarise meetings. Another uses it to analyse customer feedback. Another has quietly built a useful workaround that nobody else knows about. Another avoids AI completely because the manager is unsure what is allowed.
The organisation may be “using AI”, but that doesn’t mean it has a clear operating view. It may not know which workflows have improved. It may not know which tools are duplicated. It may not know which outputs are being reviewed properly. It may not know where AI has actually reduced effort and where it has simply moved effort into checking, correcting and explaining.
That’s the real cost question. Not just how much AI costs, but whether the business can see what it is getting back.
Unmanaged experimentation creates hidden costs
Experimentation matters. Organisations shouldn’t try to control AI so tightly that people stop learning, because a new technology only becomes useful when people can test it against real work. But experimentation becomes expensive when it remains invisible.
When every team experiments separately, the business loses the ability to compare what is working. When every manager sets their own standard, review quality varies. When every department buys or trials its own tools, spend fragments. When pilots keep running without value evidence, activity becomes a substitute for progress.
This is how AI cost builds quietly. Not always through one large, failed programme, but through many small, reasonable decisions that are never joined up. A licence here, a pilot there, a prompt library in one team, a separate tool in another, a new review burden for managers, a few customer communications that need correcting, a few reports that look polished but contain assumptions nobody checked.
Gartner reported that in infrastructure and operations, only 28% of AI use cases fully succeed and meet ROI expectations, while 20% fail outright. That finding is specific to infrastructure and operations, so it shouldn’t be treated as a universal claim about every AI project. But the underlying message is still useful: AI does not deliver value simply because it exists. It delivers value when it fits the work, has clear ownership, is reviewed properly and is measured against outcomes that matter.
The related 19.2 Insights article on fragmented AI adoption explores how this variation becomes commercially visible through inconsistent behaviours, weak manager visibility and hidden operational strain.
Cost control should not mean slowing useful adoption
The wrong response to rising AI costs is to shut down experimentation. That may reduce visible spend in the short term, but it also slows learning. It pushes useful adoption back into the shadows. It encourages teams to hide what they’re doing. It makes AI feel like a compliance issue rather than a practical capability.
The better response is managed experimentation. That means allowing people to explore AI in real work, while making the important parts visible. Where is AI being used? What task is it supporting? Who owns the use case? What tool is being used? What does it cost? What risk does it introduce? Who reviews the output? What evidence shows that the work has improved?
Those questions don’t kill experimentation. They make experimentation more useful. They help leaders distinguish between activity and value. They help managers guide teams without pretending to be technical experts. They help organisations avoid turning AI into another layer of uncontrolled software spend.
Most importantly, they protect the space for useful AI adoption by making it easier to defend.
What leaders need to make visible now
Leaders don’t need to start by shutting AI usage down. They need to make the current reality visible.
That means understanding where teams are already using AI, which tools are being paid for twice, which customer-facing tasks are now AI-assisted, and who checks the output before it reaches a customer. It also means looking beyond the invoice and asking which workflows are genuinely faster, which workflows have simply moved effort into review or correction, and which pilots are still running because nobody has made a decision to stop them.
These are practical operating questions. They sit close to the work. They show whether AI is improving the business or simply spreading through it.
This is also where governance needs to be understood properly. Good governance shouldn’t be a brake on useful AI use. It should make useful AI use easier to repeat. It should help people understand what’s allowed, what needs review, what should be escalated and what good looks like in the flow of work.
The organisations that get this right won’t be the ones with the most slogans about AI. They will be the ones who can see how AI is changing work, where it’s creating value and where it is quietly adding cost.
WAIA is designed around this practical operating layer: shared standards, manager support, role-aware guidance and visible adoption signals for workplace AI adoption.
Closing point
The answer isn’t to slow useful AI adoption. The answer is to stop treating experimentation as invisible.
AI is not a strategy, it's a tool.
Like any tool, it needs context, ownership and evidence of value. The organisations that manage this well won’t simply be the ones that spend more carefully. They’ll be the ones that can see where AI is helping, where it’s duplicating effort, where it is increasing review burden and where it is changing work faster than the operating model can support.
That’s the real cost question now.
Not whether AI is worth using, but whether the business can see enough to know where it’s worth using well.