> ## Content Index
> Fetch the complete content index at: https://bytevyte.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# McKinsey: Enterprise AI Budget Overruns Tied to Coding Agent Sprawl
- URL: https://bytevyte.com/mckinsey-enterprise-ai-budget-overruns-tied-to-coding-agent-sprawl/
- Published: 2026-09-25T19:24:33.000Z
- Updated: 2026-09-25T19:24:33.000Z
- Description: McKinsey reports enterprise AI budget overruns as coding agents scale: 93% exceed budgets and 60% of agentic costs go to refining output.
- Author: Bytevyte Editorial
- Tags: ai-beats

**McKinsey** has found that most enterprises are spending beyond their AI budgets, with the push into automated coding as the main driver. The consultancy's research, reported on Sept. 22, 2026, shows software teams leaning harder on coding agents even as measured returns lag. Falling per-token prices have not produced smaller bills, because agentic workloads multiply the number of model calls a single task needs. That gap sits at the center of the enterprise AI budget overruns now visible across large organizations.

Agentic systems rarely make one model call and stop. They call a model repeatedly, invoke external tools, retrieve data and retry steps that fail. Each operation carries a cost, and the costs compound across a single workflow. McKinsey's July 2026 analysis attributed 60% of agentic AI costs to response refinement alone, the work of checking, correcting and regenerating output once a first draft exists. Refining an AI-generated answer costs more than producing the original response.

Cost variability follows from that structure. McKinsey senior partner Lari Hämäläinen has said the same task can cost up to 30 times more depending on which agent performs it, a spread that widens as companies assign agents more complex work. Expenses track the reasoning path an agent takes, the tools it selects and the number of steps it needs, none of which a procurement plan captures in advance.

That unpredictability is its own budget problem. McKinsey has cautioned that AI spending may rise further and become less predictable as agent adoption spreads, because the meter runs on how much reasoning an agent performs rather than on a fixed licence. Annual planning cycles built around seat counts and reserved compute have no natural way to absorb that variance, which pushes the shortfall into mid-year overruns.

## Why Enterprise AI Budget Overruns Keep Growing

McKinsey's 2026 survey work puts numbers on the pattern.

| McKinsey finding                                                                | Figure | Survey basis                              |
| ------------------------------------------------------------------------------- | ------ | ----------------------------------------- |
| Enterprise AI teams exceeding budget                                            | 93%    | July 2026 agentic cost report             |
| Agentic AI cost tied to response refinement                                     | 60%    | July 2026 agentic cost report             |
| Organizations past experimentation into active deployment                       | 62%    | May 2026 Enterprise AI FinOps survey      |
| Respondents whose AI operating costs constrained usage                          | 20%    | 2026 State of AI survey                   |
| Companies that dropped a software purchase because coding agents could build it | 32%    | 2026 State of AI survey, 1,719 executives |
| Organizations qualifying as AI high performers                                  | 6%     | 2026 State of AI survey, 97 countries     |

The Enterprise AI FinOps survey, conducted in May 2026 with 120 enterprise participants and 75 qualified respondents, found AI spending rising nearly fourfold as organizations move from isolated use cases to enterprise-wide adoption. Twenty percent of respondents said AI-related operating costs, token spend included, had already limited how much of the technology they use. The fourfold jump is the mechanism underneath the enterprise AI budget overruns documented across both surveys.

The headline figures come from different instruments, and the distinction matters. The 93% overrun rate and the 60% refinement share come from McKinsey's July 2026 agentic cost analysis, while the build-versus-buy finding and the 6% high-performer rate come from the State of AI survey of 1,719 executives in 97 countries. Read together, they describe a market where deployment is broad, cost control is weak and value capture is concentrated in a small minority.

## The Buy-Side Illusion Behind Coding Agent Savings

Coding agents have changed the build-versus-buy calculation. In the 2026 State of AI survey, 32% of organizations said they declined at least one software purchase because their agents could build the equivalent internally. Roughly 31% of large enterprises are scaling agentic tools, and about a third of companies report building software in-house instead of buying it.

A canceled purchase order is not a saving on its own. McKinsey's Tara Balakrishnan has argued that money avoided up front may not survive the maintenance bill that arrives later. The survey data supports that caution: the share of respondents reporting EBIT impact from AI held flat at 37% even as reported adoption reached 80%.

That 80% versus 37% split is the clearest single indicator in the dataset. Widespread individual use of AI has not converted into enterprise-level earnings at the same rate, and the budget data explains part of why. The cost of running agents lands in the same operating expense line that the productivity gains are supposed to reduce.

Only 6% of organizations meet McKinsey's definition of AI high performers, attributing at least 5% of EBIT to AI use and describing its impact as significant. Adoption has become close to universal while enterprise value has not. Expectations of AI-driven job cuts keep outrunning the cuts themselves as well: 39% of respondents expect headcount reductions from AI, up from 32% a year earlier, while actual 2025 workforce reductions landed well below what the prior year's survey predicted.

The pressure lands unevenly. Software-as-a-service vendors face a buyer base that can now justify internal builds, and the 32% figure gives that shift a number. Platform vendors selling agent tooling collect more as agent step counts rise, since usage is metered. Enterprise buyers carry the residual risk, because a task that costs 30 times more on one agent than another still produces the same deliverable.

## Managing Run Cost Instead of Token Price

The cost line enterprises watch most closely, token price, is rarely the one that moves the budget. Vendor price cuts are genuine, but they are outpaced by the volume of tokens a single task consumes. The number of steps an agent takes is set by workflow design rather than by pricing, and controlling the bill means controlling steps and the human review wrapped around them.

McKinsey's technology budget guidance separates run spending from change spending. Agentic AI sits on the change side, so it competes for funding that has to come from somewhere else. Organizations that simplify application portfolios and design services for reuse hold down run costs and free money for agentic work; those that keep stacking new systems on legacy ones carry both bills at once.

The trade-off runs through capability as much as price. The highest-reasoning models consume the most tokens per step, and cheaper agents fail more often on complex work. The defensible position is selective routing: reserve the expensive agent for tasks where errors are costly, and push routine steps to cheaper models with tighter step limits.

There is also a sequencing problem in most deployments. Agents get rolled out to prove capability, and cost governance arrives later, once the workflows are already embedded. Retrofitting budget controls onto an agent fleet running thousands of steps a day is harder than setting per-task limits before the rollout begins.

Ownership of the bill is another open question. When a coding agent's output ships through a product team, the tokens are consumed by developers but paid for from a central AI budget, and nobody reconciles the two. Chargeback models that assign agent spend to the workflow that generated it tend to surface waste faster than quarterly reviews do.

Three measurements would give CIOs a clearer signal than total AI spend. Cost per completed task, tracked by workflow, shows which agents earn their keep. The share of steps ending in retries or human correction locates where the money goes, since refinement is the largest single cost category. The maintenance burden of internally built tools determines whether a skipped licence fee becomes a saving or a deferred cost.

## Why this matters

The McKinsey findings move AI cost control from a procurement exercise to an operating discipline. The 6% of organizations capturing EBIT gains are distinguished less by how much they spend than by whether they can price a completed task rather than a token. With most enterprises still planning to increase AI investment, and with agent spend inherently variable, the gap between high performers and everyone else will be settled on run cost rather than on model pricing.

## Sources

[The cost of intelligence: How CIOs can manage AI demand at scale](https://mckinsey.com/capabilities/quantumblack/our-insights/the-cost-of-intelligence-how-cios-can-manage-ai-demand-at-scale?ref=bytevyte.com)

[Burning through the AI budget | McKinsey & Company](https://www.mckinsey.com/featured-insights/charts/burning-through-the-ai-budget?ref=bytevyte.com)

[Recalibrating CIO technology budgets for the AI era | McKinsey](https://www.mckinsey.com/capabilities/mckinsey-technology/our-insights/recalibrating-technology-budgets-for-the-ai-era?ref=bytevyte.com)

Photo by [Igor Shalyminov](https://unsplash.com/@ishalyminov?utm%5Fsource=bytevyte&utm%5Fmedium=referral) on [Unsplash](https://unsplash.com/?utm%5Fsource=bytevyte&utm%5Fmedium=referral)

## Related Articles

- [Rising Enterprise AI Spending Forces Strategy Shift at Microsoft and Uber](https://bytevyte.com/rising-enterprise-ai-spending-forces-strategy-shift-at-microsoft-and-uber/)
- [Deloitte's 89% AI Agent Pilot Failure Rate Exposes the Enterprise Production Gap](https://bytevyte.com/deloittes-89-ai-agent-pilot-failure-rate-exposes-the-enterprise-production-gap/)
- [Agent Cost Governance Is the New Bottleneck as Enterprises Spend $1M+ a Year on Agents](https://bytevyte.com/agent-cost-governance-is-the-new-bottleneck-as-enterprises-spend-1m-a-year-on-agents/)

✔Human Verified

---

*Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.*