Where the industry actually is

Is it just me?

No. Almost nobody has got AI doing work yet, and the reason is not the technology. Here is what the 2026 data actually says, with every source, sample and field date, so you can check any of it.

Act 1

You have an assistant, not a workforce

Almost nobody has automated anything. The visitor is normal, not behind.

  • 6%of small-business workers who use AI have automated a workflow with minimal human involvement. Sixty-four percent use it for personal productivity.US Chamber of Commerce Foundation, Main Street AI Monitor, Ipsos KnowledgePanel, n=1,070 workers at businesses of 2-499 people · 8-11 May 2026
  • 6 to 1Chat subscriptions take 5.6% of business AI spend. Support bots, voice AI and agent orchestration together take under 1.3%.Ramp AI Index, corporate card and bill-pay data, n=34,183 businesses · July 2026
  • 22%of smaller organizations are scaling AI agents — flat year over year, while large enterprises went from 27% to 40%. The gap is widening.McKinsey State of AI, n=1,719 · May–June 2026

Act 2

And nobody can tell whether it is working

The blind spot shows up in government data, analyst surveys and consulting research independently.

  • 11%of organizations have no idea what their function spent on AI last year.Gartner, n=1,303 organizations above $50M revenue · January–April 2026
  • 18%of professional-services organizations know their organization tracks AI return on investment at all.Thomson Reuters, n=1,500+, 27 countries · February 2026
  • 37%of enterprises report any EBIT impact from AI (unchanged year over year), while AI spending rose roughly 47%.McKinsey State of AI, n=1,719 · May–June 2026

Act 3

The ones who win do it by structure, not tools

Three unrelated methods converge on the same finding, and it is not about the model.

  • 73% vs 25%High performers fundamentally redesign workflows; everybody else does not. Of twenty-five attributes tested, workflow redesign had the single biggest effect on EBIT impact.McKinsey State of AI, n=1,719 · May–June 2026
  • 3.5 vs 6.1Leaders pursue 3.5 AI use cases on average. Everybody else pursues 6.1 — and leaders anticipate 2.1 times the return. Focus beats breadth.BCG AI Radar · 2026
  • 9.3 vs 0.6AI leaders are 6% of companies and beat their industry by 9.3 points of three-year shareholder return. The tier directly below them — companies visibly busy with AI — beats it by 0.6.BCG, 600+ US public companies above $5B market cap, from filings and installation records rather than a survey · July 2026
  • lastTechnology immaturity ranks last among causes of AI failure across every independent source. The top cause is unclear problem definition, cited by 84% of practitioners.RAND Corporation, n=65 practitioner interviews · 2026

Act 4

Control is what makes it go faster

Governance is usually sold as a brake. The data says it is the accelerant.

  • 16xOrganizations with embedded AI controls deploy sixteen times more agents than those without, while spending four times less of their AI budget, holding 18% higher operating margins and seeing 25% fewer incidents.IBM Institute for Business Value, n=2,000 C-level executives, 33 countries · January–April 2026
  • 54AI agent incidents per organization in the past year, on average. Seventeen percent were high severity, taking four hours or more to contain.IBM Institute for Business Value, n=2,000 · January–April 2026
  • 77%say AI adoption is outpacing their organization's ability to govern it.IBM Institute for Business Value, n=2,000 · January–April 2026
  • under 25%State-of-the-art tool-using agents score around 72% on a single attempt at a task, and get the same task right eight times running less than a quarter of the time. Consistency, not capability, is the wall.τ-bench (Sierra), arXiv 2406.12045, and the Princeton HAL leaderboard · 2026
  • 36%chance that a twenty-step workflow completes correctly at 95% per-step reliability. At 99% per step it is 82%.Compounding arithmetic, safe to state as such ·
  • ~70 minof work is what today's best model handles at an 80% success rate. The same model reaches twelve hours at a 50% success rate — and autonomy claims are almost always quoting the second number.METR time-horizon data · retrieved 3 September 2026

And the ones we will not use

The most quoted number in this category is the weakest

These are the figures a page like this normally leans on. Each is here with the reason it is not on the page above, because a statistic you cannot defend costs more than the one you did not use.

Not used

40%+ of agentic AI projects will be canceled by the end of 2027.

A prediction rather than a measurement, derived from a January 2025 poll of 3,412 Gartner webinar attendees — a self-selected audience — with no published methodology showing how the figure was reached. It is also the single most used number in this category, so it does nothing to distinguish us.

Instead: The share of businesses scrapping most of their AI initiatives went from 17% in 2024 to 42% in 2025, and the average organization abandoned 46% of proofs-of-concept before production. Same instrument two years running, n=1,000+. S&P Global Market Intelligence.

Not used

95% of AI pilots fail.

Usable only with its method explained. The study's own funnel shows 80% of companies never piloted anything, and its bar for success was measurable profit improvement inside six months — from pilots that mostly had no pre-deployment baseline. Used bare it marks the user as somebody who repeats headlines.

Instead: Say what it actually measured: most of those pilots never had a baseline, so nobody could have proven success even if it happened. That turns the industry's most-quoted number into an argument for instrumentation.

Not used

94% of enterprises get no return from AI.

A mis-citation of McKinsey, which found 37% report some EBIT impact and 6% are high performers. There is no 94% figure to cite.

Not used

88% of companies use AI.

True of large enterprises in an executive survey, and false of the audience this site is written for. For small businesses the honest figure is 22.4% of US firms — and using the wrong one will feel false to a visitor's lived experience before they can say why.

Not used

any productivity gain sourced to self-report.

In a randomized trial, developers were 19% slower with AI and still believed they had been 20% faster. A thirty-nine-point perception gap sits under every 'our users report X% gains' number in this market — including the ones we will eventually collect ourselves.

One honest note

Every number here is about the market, not about us

They establish that the problem is real. They cannot establish that we solve it, and no page of somebody else’s research ever could. What we can show you is the instrument: draw your business, bind one agent to one seat, and watch whether it reports. The second half of this argument is what our first clients are proving, and we will publish those numbers the same way as these, with the sample and the date on them.

Design your business, free