By BIJ2 AI · October 11, 2026
AI adoption becomes meaningful when people keep using a tool, its work passes review, and the process holds up after the pilot ends.
An organization can buy hundreds of AI licenses and run a convincing demonstration while little changes in its daily work. Elsewhere, employees may use AI regularly without a formal company rollout. Calling both situations “adoption” hides the difference between buying access and changing how work gets done.
Managers need to know which tasks use AI, how much review the results require, and whether employees continue using it once pilot support ends. A count of licensed users cannot answer those questions.
Our earlier articles examined AI costs and economic value. Here, we examine how use develops over time and how a team can forecast it. The Work Behind a Larger AI Agent Rollout covers the operating decisions involved in expanding an agent’s responsibilities.
What adoption figures actually measure
The U.S. Census Bureau’s May 26, 2026 BTOS analysis reported that business AI use ranged between 17% and 20% during December 2025–May 2026. The question changed in November 2025: it broadened from AI used in producing goods or services to AI used in any business function. A trend line that crosses that change can mix actual diffusion with a change in what respondents were asked.
The April 2026 Census working paper on AI diffusion separates firm use, business functions, and worker tasks. For November 2025–January 2026, it reports 18% firm adoption, versus 32% when weighted by employment. Among adopters, 57% used AI in three or fewer functions. A technology can therefore reach a substantial share of workers while remaining limited to a few functions inside adopting firms.
Different samples also produce different answers. Firm Data on AI, a 2026 working paper surveying nearly 6,000 senior executives across the United States, United Kingdom, Germany, and Australia, reports about 70% of firms using AI. Yet more than 80% reported no employment or productivity effect over the previous three years. These are reported outcomes, not experimentally identified effects. The headline adoption percentage cannot be directly compared with Census’s differently defined and sampled measure.
A July 2026 preprint examining S&P 500 company filings classifies 11% of firms as having deeply integrated AI in 2025 and another 10% as using it in production or service delivery. The authors classify what large listed companies disclose. Their measure cannot capture every employee’s use of an AI tool.
Before comparing adoption estimates, check the technology definition, population, reference period, denominator, weighting, and question wording. “AI” and “generative AI” are not interchangeable survey categories.
How use develops after a pilot
The table below sets out six stages a team can use to organize its rollout evidence. BIJ2 AI proposes these stages as a planning aid; they have not been validated as a maturity scale. A team may return to an earlier stage or stop a project if the results do not justify further spending.
| Stage | Evidence to collect |
|---|---|
| Access | People can use the tool Measure eligible users and approved access. Record actual use separately from access. |
| Experimentation | A real task is tested Record task type, baseline performance, costs, and failure cases. |
| Repeated use | Employees continue using it Track whether the initial users continue using it on recurring eligible tasks. |
| Workflow integration | The workflow has clear responsibilities Define inputs, handoffs, approvals, records, and fallback procedures. |
| Dependable operation | Results remain acceptable Measure quality, exceptions, review effort, and incidents over time. |
| Expansion | Results hold up in a new setting Test it with other teams, task types, and working conditions before expanding. |
The St. Louis Fed’s November 2025 research update reports that 37.4% of workers used generative AI for work in August 2025. Its use accounted for 5.7% of total work hours. Those figures describe different aspects of adoption: how many people use it and how much of their work involves it. The authors also revised earlier estimates after testing the order of survey questions, another reason to check methods before comparing results.
The work required to make AI useful
A tool that drafts a routine email well may struggle with a decision that depends on undocumented company knowledge. Before selecting a model, define the task, the errors that matter, and the standard for a satisfactory result. A finance team might check reconciliation explanations against reviewer decisions; a support team might track resolved cases and repeat contacts.
Conflicting records, outdated information, and unclear access permissions can undermine a rollout. The OECD/BCG/INSEAD 2025 report identifies skills and data maturity as important barriers. It draws principally on 2022–23 surveys of 840 G7 enterprises and 167 Brazilian enterprises that had adopted AI. Because the sample focuses on adopters and predates the recent spread of generative AI, it cannot establish current adoption across all businesses. It does help explain why buying software leaves substantial implementation work unfinished.
Training, evaluation datasets, process changes, and documentation all cost time and money. The productivity J-curve research by Brynjolfsson, Rock, and Syverson explains why investment in those supporting capabilities can precede measured gains from a general-purpose technology. That mechanism does not establish that a struggling AI project will eventually pay off. Managers still need to check whether training and redesign improve its results.
Assign someone to set the quality standard, resolve exceptions, and decide when employees should return to the earlier process. Include the handoff to whoever acts on the AI output. A fast draft can still leave a long review queue or create repair work later.
The OECD’s September 16, 2026 practitioner study examines organizations developing or deploying agentic systems. Its companion account describes interviews with 25 organizations in 11 countries. None reported unrestricted autonomy; participants used checkpoints and layered controls. These selected interviews describe implementation practices, rather than estimating how common they are. Teams starting a deployment can use limited permissions, reversible actions, and clear escalation rules while they gather evidence on performance.
Track use, quality, and review effort
Before measuring adoption, define which tasks qualify for AI assistance. Keep that definition consistent across reporting periods, and record any changes. The measures below give a team a starting point for tracking use, quality, effort, and cost.
| Measure | Suggested definition | What it helps detect |
|---|---|---|
| Reach | Approved users with access ÷ eligible users | Availability gaps; it does not establish actual use. |
| Task coverage | Eligible task attempts using AI ÷ all eligible task attempts | Whether use is part of routine work. |
| Continued use | Initial adopters still using the workflow after a defined period ÷ initial adopters | Whether employees keep using the workflow after pilot support ends. |
| Acceptance | AI-assisted outputs meeting the specified standard ÷ outputs evaluated | Whether activity becomes usable work. |
| Total effort | Preparation + generation + review + correction + downstream rework | Whether a faster step reduces end-to-end effort. |
| Exceptions and incidents | Escalations or failures by type and severity, with rates and counts | Operating limits and harms hidden by average performance. |
| Cost per accepted result | Allocated operating and review costs ÷ accepted results | Whether expansion remains economical. |
Record task complexity and worker experience alongside the dashboard measures. If employees give AI the easier cases and handle the harder ones themselves, a simple comparison may overstate its gains. Collect only the information needed for evaluation and protect employee and customer privacy.
The METR 2025 randomized study involved 16 experienced developers and 246 tasks in familiar open-source repositories. Tasks on which AI was allowed took 19% longer in that setting. Its February 24, 2026 follow-up reported problems with selection: some developers and tasks were excluded because participants did not want to work without AI. Such exclusions make later estimates harder to interpret. A team evaluating its own rollout should document who participated, which tasks were included, and which tools were used.
NIST’s 2024 Generative AI Profile provides voluntary guidance on testing, governance, and risk management. Monitoring should continue after release, because a model update, new data source, or expanded permission changes the system being evaluated. Compliance with a framework is not itself evidence of productivity or safety.
Forecast how much work AI will support
For an internal rollout, forecast use on recurring tasks. Start with the number of eligible cases, then estimate how many the team can support with an AI workflow, how often employees will use it, and how often the result will pass review.
A simple planning calculation
Accepted AI-assisted results = eligible cases × share supported by an AI workflow × use rate within those workflows × acceptance rate.
Each percentage must refer to the cases counted in the preceding step. The calculation estimates the volume of acceptable assisted work. It does not establish that AI caused an improvement or completed the work without human help.
Suppose a service team handles 1,000 comparable eligible cases per month. The table shows three possible outcomes six months after launch. The percentages are hypothetical assumptions used to demonstrate the calculation.
| Scenario | Cases supported by an AI workflow | AI-use rate | Acceptance | Accepted assisted cases per month |
|---|---|---|---|---|
| Slow | 40% | 40% | 80% | 128 |
| Moderate | 70% | 60% | 90% | 378 |
| Rapid | 90% | 80% | 95% | 684 |
The moderate case produces 1,000 × 0.70 × 0.60 × 0.90 = 378 accepted assisted cases per month. A manager can examine each assumption separately: whether the team can prepare data for 70% of cases, whether employees will use AI on 60% of that work, and whether 90% of the assisted results will pass review.
Estimate each month’s supported case volume, use rate, and acceptance rate, then update those estimates with actual results. Include employees who stop using the tool and workflows the team discontinues. Forecast routine and complex cases separately if their use or review requirements differ, and combine them using the number of cases in each group.
An S-shaped curve may describe adoption over a longer period, but a few months of pilot data rarely establish where use will level off. Compare the curve with a simpler forecast and test both against later periods or teams excluded from model fitting. Present the range of scenarios as planning alternatives. Calling it a statistical confidence interval requires a model that supports that interpretation.
Decide when to expand
Decide before the pilot what would justify expanding, revising, narrowing, or stopping it. The criteria should cover quality, risk, continued use, review effort, and cost per accepted result. Apply them to ordinary working conditions as well as the cases selected for a demonstration.
Where feasible, randomize the rollout across comparable teams or tasks. A staggered rollout can also support comparison, though managers must account for differences in demand, staffing, task mix, and prior performance. Record the training and process changes that accompany AI access so readers can see what the evaluation tested.
This work creates a role for domain experts who can design evaluation cases, check accuracy, test handoffs, and analyze patterns of use. An economist, accountant, or operations specialist can help explain whether a result is useful and what remains unresolved. Readers considering related assignments can browse the BIJ2 AI directory and check each provider’s current requirements. BIJ2 AI may receive a fee from qualifying referrals; a listing does not guarantee work.
A rollout earns a place in daily operations when employees keep using it and its results meet the team’s standards. Following the cases that pass review, the effort they require, and the users who stay gives managers a firmer basis for deciding what to expand.