← Strategy & Moats
Foundational

Enterprise AI Adoption in 2026: Why the Numbers Disagree and How to Measure Real Use

Published adoption figures for 2026 range from about a fifth of firms to two-thirds of enterprises, because they count different things. Learn what each source measures, why licences overstate use, the adoption funnel from seat to outcome, and how a PM drives and proves adoption inside a company.

CONCEPT / STRATEGY & MOATS

TL;DR: "How many companies use AI?" has several honest answers in 2026, because sources count different things. A survey of about 400 US enterprise decision-makers by Menlo Ventures reported adoption rising from 55% to 65% and spending at a $176.8 billion annualised run rate (company-reported, from an investor in Anthropic). The US Census Bureau's survey of businesses of every size found 17% to 20% using AI, and 37% of firms with 250 or more employees. Inside a company the same gap appears between licences bought and work changed. Measure adoption as a funnel from seat to outcome, and judge success by the workflows that changed, not by seats.

Why the headline numbers disagree

SourceWho is countedWhat it reports
Menlo Ventures, 3 November 2026About 400 US enterprise decision-makers, surveyed September 2026Generative AI adoption among enterprises from 55% to 65%; spend at a $176.8B run rate (company-reported)
US Census Bureau, Business Trends and Outlook Survey, 26 May 2026US businesses of all sizes17% to 20% using AI between December 2025 and May 2026; 37% of firms with 250 or more employees
Microsoft earnings calls, 2026Microsoft 365 Copilot customersOver 30 million paid Copilot seats (July 2026), against a commercial base Microsoft put above 450 million paid seats in January 2026 (company-reported, different dates)

Three differences explain most of the gap between rows. The denominator: large enterprises against every business, including the smallest. The respondent: a decision-maker reporting company use against a business reporting its own operations. And the question: the Census Bureau changed its question wording in November 2025, so earlier figures are not directly comparable. Quote any adoption number with its source, its denominator and its date.

The funnel inside one company

Buying licences is the first step of adoption, not the measure of it. Follow the funnel.

DIAGRAM
From licence to outcome (illustrative company) Licensed 10,000 procurement's number Activated 7,000 fix: access, onboarding, single sign-on Weekly active 2,500 fix: use cases people recognise, champions In a core workflow 900 fix: integrate where the work happens, data access Outcome measured 3 workflows fix: baselines before rollout
Read diagram description

An illustrative adoption funnel for 10,000 licences. 7,000 people activated the tool, 2,500 use it weekly, 900 use it inside a core workflow, and the company can measure an outcome change in 3 workflows. Each stage needs a different intervention.

Each stage stalls for a different reason, so each needs a different fix. Activation stalls on access and setup. Weekly use stalls when people cannot see what the tool is for in their job. Workflow use stalls when the tool cannot reach the data or the system where the work happens. And outcomes go unmeasured when nobody recorded a baseline before rollout.

What the productivity evidence says

The evidence on value is mixed, which is itself the lesson: results depend on the task, the people and how use is measured.

  • In METR's February 2026 update, returning experienced developers took 18% less time with AI tools, with an interval wide enough to include no effect, and METR itself cautioned about selection effects. Its 2025 study had measured a 19% slowdown.
  • Atlassian's October 2026 survey found the share of developers saving ten or more hours a week fell from 68% in 2025 to 25% in 2026, while 65% regularly correct AI output (company-reported survey).
  • DORA's 2025 report linked AI adoption to higher delivery throughput and, still, to more instability.

Plan for value to arrive in specific workflows, not evenly across a workforce, and design the measurement before the rollout.

What a PM does about it

Pick workflows, not audiences. Choose three to five workflows where the tool demonstrably changes the work, such as drafting responses to support tickets, reconciling invoices or summarising account history before a sales call, and roll out to the teams that own them first.

Record baselines. Cycle time, volume handled per person, error or rework rate, measured for weeks before the tool arrives. Without them, every later claim is anecdote.

Instrument depth. Track weekly active users, but also actions per active user inside the target workflow and the share of that workflow's volume touched by AI.

Clear the blockers early. Security review, data access and identity integration decide adoption more than prompt training does. Treat the CISO's concerns as requirements; see delegated access for agents for one of the recurring ones.

Match the price to the value. Vendors are moving from pure per-seat pricing towards seat plus consumption; Microsoft described its Copilot model as "per-seat-plus-consumption" in July 2026. Buyers can then pay for use rather than for licences nobody opens, and sellers must show usage to justify renewal. See pricing AI products.

In the interview

Expect "we rolled out an assistant to 10,000 employees and weekly use settled at 12%; what do you do?" or "the CFO asks whether our AI add-on is failing". A weak answer runs a training campaign. A Hire answer segments users. A Strong Hire answer quotes adoption numbers with their denominators, draws the funnel from licence to outcome, finds the stage where the company stalls, fixes that stage for a few named workflows with baselines, and changes the success metric from seats to measured outcomes.

Practise with drive adoption of an AI tool inside an enterprise and only four percent bought the AI add-on.

Next: from pilot to production and pricing AI products.

RELATED CONCEPTS
PRACTISE THIS CONCEPT