KPMG’s newest quarterly AI survey produced a statistic that has spread across social media and business press over the past week: 49 percent of senior leaders said their organisation had scaled back, narrowed, delayed or paused AI agent deployments because operating costs began to outweigh the benefits. That figure comes from KPMG’s Global AI Pulse for the second quarter of 2026, a survey of 2,145 senior leaders across 20 countries. It is a striking number, and it is the one every headline has led with.
It is also the less useful number on its own. A separate, companion KPMG survey, fielded across 204 US-based C-suite and senior business leaders at organisations with at least a billion dollars in annual revenue, found something more revealing: only 26 percent of organisations report full, real-time visibility into what their AI systems actually cost to operate, even though two-thirds already have monitoring dashboards in place. Read that finding alongside the global pullback figure and the story changes shape entirely. The 49 percent pullback stops looking like a loss of enthusiasm for AI. It looks like the predictable result of asking finance teams to manage a cost that most of them cannot actually see until the invoice arrives.
2Data’s role in this is not to help organisations avoid AI. It is to help them avoid being part of that 49 percent, and understanding why that pullback is happening is the place to start.
The Companies Not in the 49% Are Not Smarter
They can simply see the meter. KPMG’s Global AI Pulse for the second quarter of 2026, drawn from 2,145 senior leaders across 20 countries at organisations with more than fifty million dollars in annual revenue, found that AI remains a top investment priority for 79 percent of leaders, with average planned spending holding at 188 million dollars. Confidence in AI is not collapsing. Spending is not collapsing. What is collapsing, for roughly half the market, is the gap between what an agent was expected to cost and what it actually costs once it is running in production.
That gap exists because the underlying pricing model changed while most organisations were still building their budgets around the old one. Flat, predictable, per-seat subscriptions are being replaced, module by module, with consumption-based pricing tied to tokens, credits, or compute minutes. A seat licence is easy to forecast. A metered bill that scales with how autonomously an agent behaves is not, and the organisations struggling with AI economics right now are largely the ones that have not yet built the instrumentation to see that second kind of cost before it lands.
The Structural Problem Behind a Behavioural-Looking Number
It would be easy to read the 49 percent figure as a story about executives being reckless with AI budgets. The more accurate reading is that the cost problem is structural, not behavioural. Nobody scaled back an agent deployment because they were careless. They scaled back because the pricing model shifted from a subscription they could budget against to a meter they had no way of reading in real time, and by the time finance noticed, the spend had already happened.
This distinction matters enormously for what actually fixes the problem. A behavioural problem gets solved with better discipline or tighter approval processes. A structural visibility problem gets solved with instrumentation: knowing, before an agent is deployed, roughly what it will cost to run at the volume the business actually needs, and knowing, while it is running, whether that estimate is holding or drifting.
This is precisely where 2Data comes in. As AI shifts from a flat, licensed feature to a metered, consumption-based part of the software estate, the work is no longer just knowing what a Copilot seat costs. It is understanding what an organisation is actually paying for, where each layer of that cost originates, and how to manage it commercially as AI becomes a larger and more permanent line item inside existing software agreements, not a separate, unmanaged category sitting outside them.
Where the Two-Layer Cost Problem Actually Lives
For organisations running on Microsoft’s ecosystem specifically, there is a concrete and frequently misunderstood mechanism behind this gap. A Microsoft 365 Copilot seat licence covers Copilot usage inside the applications it was built for, Word, Excel, Outlook, Teams, and the rest of the core suite. That is a flat, predictable, per-user cost, and it is the cost most organisations have already budgeted for and feel comfortable with.
The moment an organisation deploys an autonomous agent through Copilot Studio, or increasingly through newer agent-orchestration surfaces Microsoft is building out, an entirely separate consumption clock starts running. Copilot Studio is priced by consumption, metered through credits pooled at the tenant level, at roughly one cent per credit on a pay-as-you-go basis or a prepaid capacity pack of twenty-five thousand credits for two hundred dollars a month, and that meter runs on a different bill, tracked through a different admin surface, with no spending cap applied by default.
The scale of that second meter is easy to underestimate. Every Microsoft 365 Copilot licence already includes some baseline internal-agent Copilot Studio access, which means the real cost exposure sits specifically with custom agents built for broader distribution or higher message volumes, priced separately from the seat licence entirely. An organisation that has budgeted carefully for its Copilot seat renewal can still be caught completely unprepared by what a single well-used custom agent costs once it is live, because that cost was never part of the seat conversation to begin with.
Why This Is a Governance Problem, Not Just a Finance One
It is tempting to treat AI cost visibility as purely a finance function, something that gets solved once the right dashboard or reporting cadence exists. The KPMG data suggests the gap runs deeper than reporting. It is a governance gap: who approves a new agent before it goes live, what estimate of cost and value that approval is based on, and who is accountable for checking whether the estimate held once the agent has been running for a month.
Organisations that treat this purely as a reporting problem tend to build a dashboard that shows what was spent after the fact, which helps explain last month’s invoice but does nothing to prevent next month’s surprise. Organisations that treat it as a governance problem build the cost estimate into the approval step itself, before an agent is deployed, and then treat any material deviation from that estimate as a trigger for review rather than a fact absorbed quietly into the next budget cycle.
Why the KPMG Pullbacks Are Happening Now
This handoff point, the exact moment an organisation moves from paying for a seat to paying for consumption, is precisely where the KPMG pullbacks are concentrated, even though the survey itself does not use Microsoft-specific language to describe it. The pattern holds across vendors and platforms: predictable licensing gives way to metered billing exactly at the point where AI moves from an assistant a person actively supervises to an agent operating with some degree of autonomy, and autonomy is what drives token and compute consumption up sharply.
Nobody explained that handoff point clearly enough, and that is the gap organisations are now paying to discover the hard way. A finance team that approved a Copilot renewal at a known per-seat cost has no natural reason to expect a second, unrelated, unbounded bill to appear once the same team starts experimenting with agents inside the same platform. The bill is a surprise not because anyone did anything wrong, but because nobody flagged that the commercial model itself changes the moment usage crosses from supervised to autonomous.
The GitHub Copilot Moment Already Happened to Developers
The clearest, most visceral illustration of this shift did not happen inside an enterprise AI programme. It happened to individual developers. GitHub moved Copilot from flat-rate subscriptions to usage-based billing on June 1, 2026, and coverage from the first day of the transition captured one developer’s real-time experience discovering the new economics: a workflow that consumed the vast majority of a monthly credit allotment within hours, projecting toward a bill of roughly one hundred and eighty dollars for a plan that had been a flat ten dollars a month the day before.
GitHub itself was transparent about why the change happened. The company’s own announcement explained that Copilot now supports far more complex, agentic workflows than the flat-rate model was ever designed to price, and that the shift to credits metered by token consumption reflects what the product has actually become over the past year rather than what it was when the flat-rate pricing was originally set. The lesson generalises well beyond a single developer tool. Wherever a flat-rate AI product evolves toward autonomous, agentic behaviour, the pricing model tends to evolve with it, and the transition rarely comes with enough advance warning for budgets to adjust in time.
For most Microsoft 365 Copilot customers, the equivalent moment has not happened yet, simply because agentic deployment through Copilot Studio is still relatively early at enterprise scale. That is not a reason for comfort. It is a reason to get instrumentation in place before the moment arrives rather than after, which is precisely the window GitHub’s own developer base did not have.
What Rephasing With Discipline Actually Looks Like
KPMG’s own framing of the pullback is worth taking seriously rather than reading past. The report describes this as a market maturing rather than a bubble bursting, with organisations rephasing investments for greater financial discipline and strategic value rather than abandoning AI altogether. That framing matters, because it draws a sharp line between two very different responses to the same underlying pressure.
One response is reactive: cut whichever agents are visibly expensive once finance flags the bill, with no clear view of which of those agents were actually delivering value per credit spent and which were not. The other is structured: know, before deployment, roughly what an agent will cost to run at the volume the business needs, monitor that cost against the estimate continuously once it is live, and make the keep-or-cut decision based on value generated per unit of spend rather than on which line item happened to be large enough to notice.
The difference between those two responses is entirely a question of instrumentation, not intent. Every organisation in the KPMG survey wants to spend AI budget well. The 49 percent who pulled back are disproportionately the ones who could not see the cost clearly enough to manage it proactively, and had to manage it reactively instead once the invoice made the problem impossible to ignore.
What This Actually Looks Like for the Organisations We Work With
This is not an abstract industry statistic for the clients we support. A meaningful share of the Microsoft estates we review already have Copilot Studio agents live, in pilot, or planned for the next two quarters, and the pattern KPMG describes shows up in miniature well before it reaches the scale of a headline statistic. An agent that started as a small internal pilot, approved without anyone estimating what it would cost to run at full rollout, is the exact shape of decision that produces a surprising bill six months later.
The clients navigating this well are not the ones with the smallest AI ambitions. They are the ones who can already answer, with real consumption data rather than a general impression, what their Copilot Studio spend looks like against their Copilot seat spend, and who treat that as one commercial picture rather than two unrelated bills. Clients without that visibility today are the ones for whom KPMG’s data is not a future risk. It is a description of where they already are.
What This Means for Budget Owners This Quarter
For anyone holding budget responsibility for a Microsoft 365 Copilot renewal or a planned Copilot Studio agent rollout, the practical response to this data does not require waiting for a formal AI governance programme to stand up before taking action. A short, honest inventory of what has already been deployed, which agents are live, what they were expected to cost, and whether anyone has actually checked the real consumption against that estimate, tends to surface the gap within a single afternoon of work rather than a lengthy audit.
That inventory alone often reveals the same pattern KPMG’s data points to at scale: a handful of agents quietly consuming far more than expected, sitting alongside several that are underused relative to their cost, with nobody having compared the two lists side by side until now. Making that comparison explicit, and doing it before the next renewal or budget cycle rather than during it, is the single highest-leverage step available immediately.
The Question Worth Asking Before the Bill Arrives
The practical question every organisation running or planning agentic AI deployments should be asking right now is a direct one. Has the equivalent of the GitHub Copilot moment already happened inside this organisation, quietly, in a smaller way that has not yet reached the size where finance notices it. And if it has not happened yet, is there an actual instrumentation layer in place, tracking Copilot Studio credit consumption, agent message volume, and per-agent cost against value, that would catch it early rather than after the fact.
For most organisations still early in their agentic AI journey, the honest answer to the second question is no. That gap is exactly what closes the distance between an organisation that ends up in the 49 percent, cutting agents reactively once costs become visible, and one that scales AI deliberately because it built the visibility in from the start.
How This Connects to the Work We Already Do
This is precisely the kind of gap our own commercial reviews are built to close, and it is not a new discipline invented for AI specifically. It is the same work already applied to a Microsoft 365 Copilot renewal or an Azure commitment: get an accurate, independent picture of what is actually being consumed, map that consumption back to the specific agent, team, or business owner responsible for it, and use that picture to negotiate and govern the commercial relationship rather than reacting to it after the invoice arrives.
Applied to Copilot Studio specifically, that means auditing current agent consumption against the seat licence renewal it sits alongside, quantifying which agents are earning their cost and which are not, and folding that picture directly into the next Microsoft renewal conversation rather than treating AI spend as a separate, unmanaged category sitting outside the rest of the commercial relationship.
KPMG’s Four Recommendations, Applied to a Microsoft Estate
KPMG’s own report does not stop at diagnosing the problem. It closes with four recommendations for closing the visibility gap, and each one lands differently once it is applied specifically to an organisation running Microsoft 365 Copilot and Copilot Studio rather than treated as generic AI advice. Translating the four into what they actually mean inside that environment is where the KPMG findings turn into something a budget owner can act on this quarter.
Get a Meter on Both Layers, Not Just One
KPMG’s first recommendation is building real-time cost visibility rather than waiting for the invoice. Inside a Microsoft estate, that has to mean visibility into two separate meters, not one. Seat-based Copilot spend is already visible, because it is a fixed number on a familiar renewal. The consumption meter running underneath Copilot Studio agents is the one that typically has no equivalent dashboard anyone is checking regularly, and it is precisely the one KPMG’s data says only a quarter of organisations can actually see in real time.
The practical version of this recommendation is not a generic AI cost dashboard. It is confirming, specifically, who has visibility into Copilot Studio credit consumption today, how often that consumption gets reviewed, and whether an alert fires before a monthly allowance is exhausted rather than after.
Make the Seat-Versus-Consumption Distinction Common Knowledge
KPMG’s second recommendation is treating AI cost literacy as a leadership discipline rather than a technical detail owned solely by IT. The single most useful piece of literacy to build first, inside a Microsoft environment, is the seat-versus-consumption distinction covered above. Finance and business leaders who approve a Copilot renewal need to understand explicitly that the approval covers seat usage only, and that any agent deployed through Copilot Studio draws against an entirely separate, uncapped budget line by default.
That distinction rarely gets taught during a standard renewal conversation, because the Copilot seat purchase and the Copilot Studio agent decision are typically made by different people at different times. Closing that gap is less about a training programme and more about making sure whoever approves a new agent understands, in plain terms, that they are opening a second bill, not adding a feature to the first one.
Price the Agent Before It Is Built, Not After
KPMG’s third recommendation is embedding a cost review into the approval process itself, so no agent goes live without a projected cost-to-value case attached. Applied to Copilot Studio specifically, that means estimating expected message volume and the mix of standard versus premium message types before an agent reaches production, since those two factors alone determine most of the gap between what an agent was expected to cost and what it actually costs once real users are driving it.
This does not need to be a heavyweight governance process to be effective. A short, mandatory estimate, reviewed by whoever owns the Microsoft relationship, before an agent moves from sandbox to production, catches the large majority of the surprises this article’s data describes, simply because it forces the cost conversation to happen before deployment rather than after the first invoice.
Rephase the Way You Would Renegotiate a Vendor Contract
KPMG’s fourth recommendation is rephasing investment toward the agents generating the strongest returns rather than retreating from AI altogether. That is, in effect, a licensing and renewal discipline applied to AI spend rather than to a software contract, and it is exactly the kind of exercise a structured commercial review is built to support: rank existing agents by value delivered per credit consumed, keep and expand the ones that clearly earn their cost, and either redesign or retire the ones that do not.
Framed this way, rephasing is not a defensive response to a KPMG statistic. It is the same discipline already applied to a Microsoft 365 Copilot renewal or an Azure commitment, extended to cover the newer, faster-moving consumption layer sitting on top of it.
Conclusion
The 49 percent of executives who pulled back on AI agents did not fail at AI. They failed at cost visibility, and only 26 percent of organisations currently have the real-time instrumentation that would have caught the problem before it became a headline statistic. That is a solvable gap, not an indictment of agentic AI itself, and KPMG’s own data backs that reading directly: spending is holding, confidence is rising, and the organisations pulling back are rephasing rather than retreating.
The practical work ahead of that rephasing is understanding exactly where a Microsoft 365 Copilot seat licence’s coverage ends and where a separate, unbounded Copilot Studio or agent consumption meter begins, building the visibility to track that second meter before it generates a surprise, and making the keep-or-cut decision on agents based on value per credit rather than which bill happened to be large enough to notice first. Organisations that build that instrumentation now will be having a very different conversation at next year’s KPMG survey than the one dominating this year’s.
None of this is about steering organisations away from AI. It is about making sure they are not the ones discovering the cost of AI agents the hard way, after the fact, as part of next quarter’s 49 percent.
