McKinsey published "The state of AI in 2026: On the road to ROI" on August 25, 2026 — an online survey of 1,719 participants across 97 nations, fielded May 4 through June 8, 2026. The numbers that matter for builders are not the macro spend charts. They are the gap between 80% of respondents reporting individual productivity gains and 37% reporting any enterprise EBIT impact — flat versus last year's survey — plus 32% who decided not to buy software because agentic coding tools let them build in-house instead.
If you ship with Claude Code, Codex, or an internal harness, this is your receipt: the market McKinsey measures is feeling faster at the keyboard long before finance can tag a line item. That is not hype denial. It is the same pattern explainx.ai mapped in the build-vs-buy ROI framework — attribution lags leverage when you add tools instead of redesigning workflows.
TL;DR — what practitioners should take away
| Question | Direct answer (McKinsey Aug 25, 2026) |
|---|---|
| Do people feel faster? | 80% report individual productivity gains |
| Does finance feel it? | Only 37% report enterprise EBIT impact — flat YoY |
| Is build beating buy? | 32% skipped a software purchase because they built with agentic coding tools |
| Are tokens a brake? | 20% say AI operating costs constrain deployment |
| Are agents scaling? | 40% of large enterprises scaling agents, up from 27% |
| Who actually wins? | ~6% "AI high performers" — workflow redesign, not bolt-on |
| Headcount? | 39% expect declines next year vs 14% who saw them last year |
The productivity–EBIT gap is the whole story
McKinsey's central tension is not "AI doesn't work." It is "AI works in the inbox before it works in the earnings call."
Four in five respondents say AI makes them individually more productive. Just over one in three say their organization sees EBIT impact — and that share did not meaningfully move from the prior year's survey. High performers — roughly 6% of the sample — are the ones who can tie AI to operating profit and report significant value. McKinsey again ties that group to fundamental workflow redesign: fewer steps, different handoffs, humans reviewing exceptions instead of pushing paper.
That matches what explainx.ai sees in loop engineering deployments: the teams that win do not paste a copilot into step 7 of a ten-step process. They ask whether the process should exist, then encode the new path in CLAUDE.md, skills, and scheduled agent loops. Bolt-on chat is the 94% playbook. Rewiring is the 6% playbook.
For executives framing the next budget conversation, the AI readiness checklist scores whether you can actually build and govern in-house — which matters now that a third of McKinsey respondents are skipping vendor SKUs entirely.
32% built instead of bought — agentic coding changed the procurement math
The stat that should wake up SaaS GTM teams: 32% of McKinsey respondents decided not to purchase software because agentic coding tools made an in-house build feasible.
This is not "developers hobbying on weekends." It is procurement logic. When a product manager plus one engineer with Claude Code commands can ship an internal workflow app in a sprint, the vendor has to justify ongoing seat cost against:
- Context the vendor cannot see (your data model, approval chain, weird edge cases)
- Iteration speed — your team ships the Tuesday fix without a QBR
- Token burn you already pay for engineering harnesses
explainx.ai covered the consumer-side version in Software for one. McKinsey's number is the enterprise mirror: the build path crossed the buy path for nearly a third of surveyed orgs. Pair that with Is software dying or changing? — thin SaaS with no moat is the first category on the cancel list.
Practical check before your next renewal: if the vendor feature is "AI wrapper over forms you already have," run a one-week in-house spike with your existing harness. McKinsey says a meaningful share of your peers already did and kept the budget.
40% of large enterprises are scaling agents — up from 27%
Agent adoption is no longer confined to innovation labs. Among organizations with $1B+ revenue, 40% report scaling AI agents somewhere in the enterprise, up from 27% in McKinsey's previous survey.
"Scaling" still usually means one or two functions, not company-wide autonomy. McKinsey's prior reports consistently show agent deployment concentrated in IT, customer operations, and software engineering — rarely more than 10% scaling agents in any single function. The direction is clear anyway: agents graduated from "we ran a pilot" to "we have a line item and a runbook."
For builders, the implication is operational, not visionary:
- Design for failure — agents that execute multi-step workflows need escalation paths, not demo-happy happy paths. See human-in-the-loop guidance before you wire write access.
- Measure tokens per task, not tokens per seat — Databricks' four cost levers apply the moment an agent loops.
- Use loops, not one-shot prompts — loop engineering is how teams babysit PRs, cron migrations, and long fixes without a human clicking "continue."
20% are token-constrained — operating cost is a real ceiling
One in five respondents say AI operating costs — inference, API usage, seat tiers — constrain how far they can push AI. That is consistent with August's Codex quota incident: agentic sessions burn on compactions, background tools, and multi-step plans, not on a single chat turn.
If you are the engineer getting blamed for "AI spend," bring finance these levers (summarized from Databricks and McKinsey's cost theme):
| Lever | What to do |
|---|---|
| Model routing | Default cheap; escalate only when evals fail |
| Session hygiene | New thread for image-heavy work; avoid infinite compaction |
| Build vs buy | In-house only when repeat runs amortize token cost |
| Spend gates | Warn at 80%, downshift model before hard cutoff |
McKinsey's 20% constraint figure is a survey answer, not an invoice — but it explains why some teams have harness access on paper and throttle on practice.
Headcount expectations are running ahead of reality
McKinsey reports 39% of respondents expect headcount declines in the next year from AI — versus 14% who report they already saw declines in the past year. Expectations doubled; realized cuts did not.
That gap matters for builders in two ways:
- Your job security argument inside the company should cite workflow outcomes (cycle time, error rate, shipped internal tools), not "AI will shrink the org." Leadership expects shrinkage; operations mostly sees reshuffling and hiring freezes.
- The 32% build-in-house stat suggests headcount pressure may show up as fewer vendor seats before it shows up as fewer engineers — especially where agentic coding keeps internal teams productive.
What Hacker News keeps getting wrong — and right — about McKinsey ROI
McKinsey surveys reliably rank high on Hacker News when they drop, and the comment threads repeat three themes worth separating:
1. The Access database parallel. In the 1990s, Microsoft Access put a database on every knowledge worker's PC. Adoption looked universal; enterprise value did not. HN commentators argue today's copilot seats rhyme with that history — individual empowerment without structural change. McKinsey's 80% vs 37% split is the quantitative version of that anecdote.
2. "No ROI yet" is compatible with real speed. Practitioners report shipping internal tools, closing tickets faster, and skipping SaaS — while finance still lacks attribution models. McKinsey measures EBIT impact, not "did your team merge more PRs." Both can be true.
3. McKinsey credibility. Consultants selling transformation will emphasize transformation levers — workflow redesign, five-to-one people-vs-tech spend — that are directionally right but hard to audit from outside. Treat the survey as directional evidence from 1,719 self-reported executives, not as ground truth about your sprint board.
The honest synthesis for builders: you are probably in the 80% who feel faster. Whether your company joins the 6% who prove EBIT impact depends on whether anyone rewires the process your agent touches — not whether you upgraded to the newest model.
What to do this week (builder checklist)
- Pick one workflow with measurable before/after (triage time, manual copy-paste hours, vendor seat cost) — not "use AI more."
- Run a build spike before renewing a thin SaaS SKU; McKinsey says 32% of peers already chose build.
- Add guardrails if you scale agents: iteration caps, token budget, human approval on writes — see loop engineering guardrails.
- Track cost per merged PR or closed ticket, not total AI invoice — finance speaks in attribution; give them a ratio.
- Score org readiness with the four-dimension checklist before asking for enterprise-wide agent rollout.
Related on explainx.ai
- Loop engineering: design coding agent loops
- Claude Code commands — complete reference
- Why Codex quota drains fast (Aug 2026)
- AI ROI framework: build vs buy for executives
- AI readiness checklist for business leaders
- Software for one — personal apps with agents
- Is software dying or changing?
- Databricks — managing AI coding costs at scale
Sources
- McKinsey & Company — The state of AI in 2026: On the road to ROI (published August 25, 2026; survey n=1,719, May 4–June 8, 2026)
- McKinsey & Company — Prior State of AI survey — workflow redesign and high performers (2025 edition, for YoY comparison context)
Survey figures reflect McKinsey's August 25, 2026 publication and respondent self-reporting — verify against the linked primary report before citing in board decks or renewal negotiations. Agent tooling, token pricing, and harness behavior change quickly; re-check your own meters and runbooks after major vendor patches.