The Business Impact of AI Adoption: Where Cost, Efficiency, and Operational Risk Actually Show Up

McKinsey's tenth annual State of AI survey, published on August 25, 2026, and based on 1,719 respondents across 97 countries, includes two numbers that don't sit comfortably together. Eighty percent of respondents say AI has improved their individual productivity. Only 37% say AI has contributed anything at all to their organization's EBIT – essentially unchanged from a year earlier. Just 6% clear McKinsey's "high performer" bar of attributing 5% or more of EBIT to AI.

A third number from the same survey explains why this article exists: about one in five respondents report that AI-related operating costs, including token costs, have constrained how much AI they use. McKinsey is careful to call this a meaningful consideration rather than a widespread constraint – 60% still expect to increase AI investment over the next year – but a cost line that changes behavior before anyone has proven the return is the definition of a measurement problem.

So the picture is a technology people can feel working, a cost line that is starting to bite, and a bottom line that has not moved. This article looks at the impact of AI on business in that narrow, practical sense: where the cost shows up first, what the efficiency gains actually cost, how the picture differs by industry, and how to measure it all before the spend gets ahead of the outcome.

We treat "AI impact on business" as a financial and operational question – cost, efficiency, output quality, and operational risk, the things a CFO or a VP of Engineering has to answer for. That is a different question from AI's societal or environmental footprint, which gets most of the headlines and none of the budget scrutiny. 

‍

What AI Impact on Business Means

The business impact of AI adoption falls into four measurable categories, and the most common reporting mistake is mixing them.

•      Cost. The most visible: compute, licensing, integration, and the people running it. Measured as spend per team, per feature, and per environment – not as one monthly total.

•      Efficiency. Throughput per dollar or per hour: tickets closed, code shipped, reports produced. Measured against the same team's pre-AI baseline on comparable work.

•      Output quality. Accuracy, consistency, and whether AI-assisted work needs less rework than before. Measured as revert rate, review rounds, and defect escape rate.

•      Operational risk. The slowest to show up and the most expensive when it does: compliance exposure, model drift, vendor lock-in, and ungoverned AI agents making decisions nobody signed off on.

Operational risk is the category teams most often declare unmeasurable, but it is. It has countable proxies: the share of AI traffic passing through a governed gateway (coverage), the number of registered agents running without cost, time, or recursion limits, the number of models and MCP servers in use without a named owner, and time-to-revoke when access has to be pulled. If none of those numbers exist, the risk is not low. It is unmeasured.

Deloitte's 2026 State of AI in the Enterprise survey, covering 3,235 director-to-C-suite leaders across 24 countries, shows the same activity-versus-outcome gap from another angle: 66% report productivity gains, but only 20% report revenue growth today, while 74% hope to get there. 

‍

Where AI Impact on Business Operations Shows Up First: Infrastructure and Compute

Before AI touches a single business outcome, it touches the infrastructure bill. For most companies running AI in production, inference quickly becomes the least predictable cost line. Unlike a fixed SaaS seat license, inference cost scales with usage, model choice, and prompt complexity, which makes it structurally harder to forecast than traditional cloud spend.

The macro picture sets the context. Gartner's July 2026 forecast puts total worldwide IT spending at $6.37 trillion in 2026, up 14.2%, with data center systems – the infrastructure AI models run on – growing faster than any other category, by 62.5% to $822 billion. That figure covers all IT, not AI alone. Gartner's separate May 2026 forecast puts worldwide AI spending at $2.59 trillion in 2026, up 47%, with more than 45% going to infrastructure rather than applications.

As Gartner analyst John-David Lovelock put it: "Building the compute capacity required for AI is the largest infrastructure project ever attempted by humanity."

Most of that spending is vendor and hyperscaler capex. It reaches a company's own AI bill through decisions nobody signs off on individually:

•      GPU capacity reserved for AI workloads and sitting well below full utilization most of the time.

•      Large frontier models used for tasks a smaller model would handle just as well.

•      Duplicate and retried API calls that multiply the bill without adding output.

•      Context re-sent on every call – system prompts, tool schemas, conversation history – billed each time again.

Each is small on its own. Together they arrive before any productivity gain does, which is why the AI bill grows before there is a benefit to offset it. This is the first real AI cost most companies meet, and it needs an owner.

‍

Efficiency and Productivity Impact: Gains vs. Hidden Costs

Zylo's 2026 SaaS Management Index, built on more than 40 million SaaS licenses and $75 billion in spend under management, found that enterprise spending on AI-native applications rose 108% year over year to an average of $1.2 million – and 393% in organizations with more than 10,000 employees. In an accompanying survey of 218 IT leaders, 78% reported unexpected charges tied to consumption-based or AI pricing models, and 61% had cut projects because of unplanned software cost increases.

The same report shows how far purchasing has drifted from central control: business units now control 81% of SaaS spend, while IT directly manages just 15%. A growing share of AI spend is bought outside the team that would normally govern it.

Two hidden-cost patterns follow from this.

•      Shadow AI. Employee-expensed tools and unsanctioned subscriptions that never reach a central budget – and never reach a DLP policy either. Invisible until someone tallies the overlapping subscriptions.

•      Inference cost creep. Usage-based pricing that scales with adoption, with no single decision point where anyone approves the increase. Cursor's mid-2025 move to compute-based billing is a public example: the change produced unexpected charges, a public apology, and refunds.

Efficiency gains from AI are real, but they are uneven and rarely measured against a baseline – and, as the next section shows, sometimes not gains at all. Without cost visibility, they are easy to spend twice.

‍

AI Impact on Industries: Finance, Consulting, and Tech

Adoption looks similar across industries; impact does not. The difference usually comes down to whether a company uses AI to restructure how work gets done, or to speed up the process it already had.

In financial services, JPMorgan Chase offers the most concrete public number available. In a Bloomberg TV interview on October 7, 2025 (as reported here), CEO Jamie Dimon said the bank spends roughly $2 billion a year on AI and sees roughly $2 billion in annual benefit – breakeven, which he called "the tip of the iceberg" as more use cases mature. AI at JPMorgan spans fraud detection, risk management and contract review, plus an internal model used weekly by about 150,000 employees. It is worth sitting with that: one of the most AI-mature banks in the world publicly reports parity, not profit.

In consulting, BCG's analysis makes a point that generalizes. AI creates spare capacity, but that capacity is not automatically worth anything; someone has to decide what to do with it. Firms that layer AI onto existing workflows see small gains, while those seeing real impact redesign how the work is done. BCG's 10-20-70 principle puts the ratio bluntly: roughly 10% of the effort goes to algorithms, 20% to technology and data, and 70% to the people and process work that makes the change stick.

In technology and software companies, the most useful data point is a warning about measurement itself. In METR's 2025 randomized controlled trial, experienced open-source developers were 19% slower with AI tools, yet came away convinced the tools had sped them up by roughly 20%. That gap between perceived and measured performance is exactly what a business case has to catch early, and it is why self-reported time savings cannot be the basis of an AI impact figure. We unpack the study and what to measure instead in our guide to calculating AI ROI.

‍

AI Adoption Challenges: What Happens When Nobody Measures

The cost of not measuring shows up as failed initiatives. In November and December 2025, Gartner surveyed 782 infrastructure and operations leaders about AI use cases within I&O – auto-remediation, self-healing infrastructure, agent-led management of workflows between systems. Only 28% of those use cases fully succeed and meet ROI expectations, and one in five fails outright. Fifty-seven percent of I&O leaders have experienced at least one AI failure; among those, 38% cited persistent skill gaps and 38% cited poor data quality or limited data availability.

Model sophistication didn't drive success. Among the 77% of leaders who delivered at least one working use case, Gartner attributes success primarily to integrating AI into the workflows and systems people already use, and to securing full support from business executives. The most mature areas produced the most wins: 53% of I&O leaders reported that their AI wins came in IT service management.

Note the scope: these are AI use cases applied to the IT operations function, not the build-out of AI infrastructure itself. Secondary coverage often compresses this into "AI infrastructure projects", which reverses the meaning.

McKinsey points at the same root cause from another angle. Nearly three-quarters of its AI high performers report fundamentally redesigning workflows because of their AI use, up from 55% a year earlier — against just one quarter of everyone else. High performers are also 3.3 times more likely than other organizations to intend to use AI to fundamentally transform their business within three years, and twice as likely to say their organization has defined processes for measuring the impact of AI initiatives. Ambition, redesign, and measurement go hand in hand. Most organizations are still running a faster version of the old process.

Without a measurement framework in place, all of this surfaces for the first time at a budget review six months in.

‍

From Impact to a Number: What Has to Be True First

Everything above describes where AI impact appears. Turning it into a single figure – the formula, the full cost inputs, the KPI set, the payback period – is a separate exercise, and we walk through how to calculate AI ROI step by step in a dedicated guide.

What matters here is what comes before the arithmetic: whether the inputs are trustworthy, and whether they keep arriving. A number calculated once, from figures people supplied about themselves, is not a measurement – it is a survey with a decimal point. Costs and outcomes both keep moving, so the reporting has to keep moving with them.

Continuous measurement is also what separates transformation from tinkering. Deloitte's survey also found that only 34% of organizations are using AI to deeply transform – creating new products and services or reinventing core processes – while 30% are redesigning key processes and 37% are applying AI at the surface, with little or no change to how the work happens. Transformation and measurement tend to arrive together, because you cannot redesign a process you have never measured.

This is the gap OptScale AI's AI ROI & Impact pillar is built to close. Rather than reporting spend on its own, it joins gateway telemetry – tokens, model, provider, cost, actor – to your systems of record on a single row, pulling merges, review rounds, reverts, and ticket transitions read-only from GitHub and Jira.

Three properties keep those numbers credible in a budget review:

•      Outcomes come from systems of record. Merges, reverts, and ticket transitions are read from GitHub and Jira, so nobody has to fill in a survey or remember to instrument an SDK call.

•      Comparisons are cohort-based. AI-assisted work is measured against unassisted work from the same period and against the team's own pre-adoption baseline, so every productivity claim has a reference point.

•      Measurement, not surveillance. Individual analytics are off by default, require an org-level decision to enable, suppress any group smaller than five, and log every view. Team and organization analysis works fully without them.

Join confidence, heuristic-match labeling, and gateway coverage reporting sit behind those three properties, so a figure never looks more certain than the data behind it.

On the cost side, the same gateway applies smart routing, cache alignment, system-prompt trimming, and token compression, and reports savings by lever, net of provider prompt caching, rather than as one blended number, because each lever behaves differently across workload mixes. Combined, the levers can cut AI spend by up to 60%; how much of that applies to you depends on your workload.

‍