The AI chasm has moved
inside the organisation

The chasm is now inside the organisation

The new adoption divide sits between having AI and changing how work is organised around it.

The old adoption question was whether an organisation had crossed from experimentation into use. That question is becoming less useful.

AI is now embedded in office software, search tools, customer platforms and specialist applications. Employees can access it without a formal programme. Business units can buy it without redesigning a process. A company can therefore be an “AI user” while changing very little about how work is organised, governed or measured.

The harder divide now sits inside the organisation. It separates access from individual use, individual use from repeatable workflow, and repeatable workflow from operating-model change.

Geoffrey Moore’s original “chasm” described the difficult move from early adopters to a mainstream market. The term is used here as an extension, not as an attempt to revise that model. For many employees, AI is already an ordinary feature inside familiar software. Yet many organisations remain on the near side of an internal chasm: the gap between having the technology and building a dependable way of working around it. ¹

This distinction explains an apparent contradiction in the evidence. Surveys can report widespread AI use at the same time as businesses report little effect on productivity, employment or financial performance. They are often measuring different layers of adoption.

“AI adoption” is answering several different questions

The October 2026 revision of a United States Census Bureau working paper provides one of the clearest views of this problem because it measures AI diffusion at three levels: the firm, the business function and the worker task. The paper uses the Census Bureau’s second AI supplement to the Business Trends and Outlook Survey, a nationally representative weighted survey of US private non-agricultural employer firms. Responses were collected across six biweekly panels from 17 November 2025 to 8 February 2026.

The revised paper estimates that 17.9 per cent of firms reported using AI in a business function during the previous two weeks; the employment-weighted share was 31.2 per cent. In a separate question covering the previous six months, 22.6 per cent of firms reported worker use of AI in tasks supporting business functions, or 40.5 per cent on an employment-weighted basis. ²

Those figures should not be arranged into a simple funnel. Firm adoption and worker use are not sequential stages measured on a common denominator. Instead, they reveal two different gaps.

In the revised paper’s firm-weighted cross-layer analysis, 64.0 per cent of firms reporting worker-task use also reported firm-level AI use. The remaining 36.0 per cent did not report the other layer. Conversely, 80.8 per cent of firms reporting firm-level AI use also reported worker-task use, leaving 19.2 per cent without reported use at the worker-task layer. ²

The authors interpret this asymmetry as evidence that both bottom-up and top-down diffusion matter. It is consistent with what this article calls visibility and implementation gaps, but it is not a direct measure of either. The two questions use different reference windows—six months for worker-task use and two weeks for firm-level use—and both answers come from the firm’s respondent rather than from matched employee reports.

Other surveys reinforce the distinction between breadth and depth. In round 37 of the European Central Bank’s Survey on the Access to Finance of Enterprises, fielded from October to December 2025 among more than 5,000 euro-area firms, more than 70 per cent of firms, survey-weighted, reported some AI use, including infrequent use. Only 7 per cent of all firms reported “significant use”, which the ECB analysis defines as intensive use. ³

Federal Reserve Bank of Atlanta Working Paper 2026-3 combines parallel questions from stratified firm surveys fielded between November 2025 and January 2026 in the United States, United Kingdom, Germany and Australia, with almost 6,000 executive respondents. Its sample-size-weighted cross-country estimate is that 69 per cent of firms currently used at least one of the listed AI technologies. Across all surveyed executives, including those who reported no use, the authors estimated average personal use at about 1.5 hours per week from categorical responses; 28 per cent reported no personal AI use in a typical week. Looking back over three years, more than 90 per cent of firms on average estimated no AI effect on employment, while 89 per cent estimated no effect on labour productivity, defined as sales volume per employee. ⁴

These results are not directly comparable: their questions, samples, respondents, weighting and reference periods differ. The Atlanta Fed results are executives’ retrospective estimates rather than audited or causally identified effects. That disagreement is itself useful. A firm can have broad access, frequent experimentation and little operating impact.

“Uses AI” has become too broad a category for serious management decisions.

The five-level AI Adoption Stack

The difficult transition is no longer simply from no access to access. It is from individual use to repeatable team workflow, function redesign, and operating-model integration and value.

A more useful discussion begins by identifying the level at which adoption has actually occurred. The stack is not a universal maturity score. Different functions in the same organisation may sit at different levels. Its purpose is diagnostic: to show where the hand-off from technology to organisational value has stalled.

On a narrow screen, each level is shown as a separate card.

1. Access

The organisation has licences, approved tools, embedded features or API access. Access removes a technical barrier. It says little about whether employees use the capability, whether it addresses important work or whether its output can be trusted.

2. Individual use

Employees use AI for tasks such as drafting, summarising, searching, coding or document analysis. Value at this level can be real, but it is often private, uneven and difficult to observe. One employee may save time while another creates review work for a colleague. An apparent productivity gain may be absorbed rather than converted into higher output, lower cost or better service.

3. Repeatable team workflow

AI is incorporated into a defined sequence of work with agreed inputs, quality checks, ownership and exception handling. This is the first level at which an organisation can test a repeatable operational proposition. The relevant question shifts from “Are people using it?” to “Can this workflow produce a dependable result under normal conditions?”

4. Function redesign

A business function changes roles, controls, systems, performance measures or service levels around the new capability. The organisation is no longer adding AI to an unchanged process. It is deciding which steps should disappear, which should remain human, where judgement belongs and how capacity should be redeployed.

5. Operating-model integration and value

AI-enabled workflows are connected across functions and reflected in decision rights, resource allocation, risk ownership and financial outcomes. This is where local improvements become organisational performance. It is also where benefits become harder to claim casually: a faster upstream activity may simply move the bottleneck downstream.

Most adoption remains narrow

The Census evidence suggests that formal use is still concentrated. Among firms reporting AI use in at least one of the 15 surveyed business functions, 57 per cent, firm-weighted, used it in only one to three functions.

A weighted latent-class analysis, restricted to firms reporting AI use in at least one business function, classified 37 per cent as Minimalist Adopters, 31 per cent as Marketing Specialists, 15 per cent as Institutional-Administrative Integrators, 12 per cent as Technical Strategists and 4 per cent as Comprehensive Adopters. The final figure is 4 per cent of business-function AI users—about 1 per cent of all firms—not 4 per cent of all firms. The pattern matters: the typical adopter is not yet an AI-native operating model. It is a firm with selected points of use. ²

Reported task effects were primarily augmentative. Among AI-using firms that reported at least one of the three surveyed task effects, 66 per cent, firm-weighted, reported augmentation alone; the employment-weighted share was 70 per cent. Over the previous six months, 2.0 per cent of AI-using firms, firm-weighted, reported an AI-attributed net employment decrease; the employment-weighted share was 2.4 per cent. The survey did not measure the magnitude of those changes or gross worker flows such as hiring and separations. ²

None of this proves that deeper integration will reduce employment or generate large productivity gains. CES 26-25R is a working paper whose results are descriptive and correlational; it has not undergone the review accorded Census Bureau publications, and no Census endorsement should be inferred. Worker use is reported by firm respondents rather than workers. Reported task and employment effects are respondents’ subjective assessments of what AI changed relative to an unobserved no-AI counterfactual, so attribution error is possible.

Workflow redesign is the difficult part

A concrete example from the ECB shows what movement up the stack looks like. Each quarter, the Corporate Telephone Survey conducts around 70 video interviews, each lasting about 30 minutes, with senior representatives of large companies active across the euro-area economy. Traditionally, one economist usually conducts the interview while another takes handwritten notes; the team then prepares one-page meeting summaries, scores responses and synthesises an overall report.

The ECB did not obtain substantial gains merely by giving economists access to a general chatbot. Early experiments with off-the-shelf tools produced only modest improvements. It then built a more specific workflow: authorised transcription, confidentiality-preserving redaction, a bespoke language-model tool, systematic comparison with human summaries, prompt and parameter testing, human review and continued evaluation of scoring accuracy.

The ECB authors report that early, internally assessed results found the AI-generated summaries to be of similar quality to human-written versions. While the workflow was still in development, they estimated that LLM-generated summaries saved around 25 person-hours in each of the first two survey rounds and at least 31 person-hours in the January 2026 iteration. The more important point is not the number. It is the work required to make the estimate credible. ⁶

The gain came from redesigning a workflow, not from distributing access. The process had an owner, a defined dataset, quality measures, safeguards, an exception path for participants who declined AI transcription and a clear role for human judgement.

This is an ECB blog describing an internal case, not a controlled productivity study. Summary quality was assessed by Corporate Telephone Survey staff, and the time saving is an internal estimate for a survey iteration rather than a per-interview or independently audited measure. The scoring work was also still under evaluation: the reported scoring test used 136 interview transcripts from October 2024 to October 2025, with each transcript scored five times. Participants can opt out of AI transcription, in which case the traditional process remains available.

That is the internal chasm in miniature.

Strategy and trust are operational inputs

The current April 2026 version of CESifo Working Paper 12373 analyses the Gallup Workforce Panel from 2023 to the first quarter of 2026. Its weighted analytic panel, restricted to full- and part-time US workers aged 20 to 65 with non-missing AI-use data, contains 33,743 unique workers and at least 78,939 person-wave observations. In the paper’s most saturated cross-sectional specification, based on 19,699 person-wave observations because workplace items were routed to subsets of respondents, employees who reported that their organisation had communicated a clear AI strategy were 26.8 percentage points more likely to report frequent AI use—daily or a few times a week—conditional on demographics, survey-wave effects and other workplace-practice measures. ⁵

The relationship between AI use and employee outcomes also varied with perceived strategic clarity and trust in leadership, but the results are not uniform across specifications. In within-person models, the strategy interactions remained clearest for engagement; the corresponding burnout and job-satisfaction interactions attenuated and were not statistically distinguishable from zero. The paper’s trust results concern perceived productivity gains and reported patterns of use, not independently measured productivity.

These are observational associations and should not be treated as proof that communicating a strategy causes productive use or better outcomes. AI use, strategy clarity, trust and outcomes are employee-reported; the paper does not have a direct employer identifier; and its job-switcher design is explicitly not a claim of causal identification from randomised exposure. The findings do, however, challenge the idea that adoption is mainly a matter of individual enthusiasm. Employees make decisions about whether and how to use AI from managerial signals: what is encouraged, what is safe, what will be recognised and who will carry the consequences of failure.

Governance therefore cannot remain a policy document attached to Level 1. It must be designed into the workflow. Each serious use case needs a named owner, permitted inputs, evaluation criteria, human sign-off rules, an escalation route and a way to stop or reverse the process.

The budget must follow the adoption level

Many AI programmes are funded as software purchases and judged as transformations. That mismatch almost guarantees disappointment.

Access can be purchased from a technology budget. Repeatable workflows require product and process ownership. Function redesign requires operational leadership, integration capacity and change funding. Operating-model integration requires decisions about incentives, staffing, service levels and the allocation of benefits.

The ECB’s survey of intensive AI users makes this distinction visible. More than 84 per cent of firms reporting intensive AI use said they had invested in the technology, compared with 33 per cent of moderate users. Intensive users also more frequently cited growth and innovation, including research and development and expansion of products and services, while firms at an earlier stage often cited cost reduction and operational efficiency. The ECB authors argue that moving beyond licences for general AI tools into customised solutions or upgraded digital infrastructure generally requires larger and longer-term investment. ³

The practical implication is not that every organisation should spend more. It is that leaders should stop funding one level while expecting results from another.

Six questions for the next AI portfolio meeting

1. At which level of the Adoption Stack does each priority use case actually sit: access, individual use, repeatable workflow, function redesign or operating-model integration?
2. What evidence would justify moving that use case up one level?
3. Who owns the performance of the workflow, rather than the procurement or availability of the tool?
4. Which budget covers integration, evaluation, data preparation and process change after the licence has been purchased?
5. Which controls, review points and escalation paths must change as the use case becomes more embedded?
6. If the promised capacity appears, will the organisation use it for greater output, better quality, new work or lower headcount—and who has authority to make that choice?

Move one workflow, not the whole organisation

The immediate management task is not to declare that the organisation has crossed into AI adoption. It is to identify where adoption has stopped.

Select one consequential workflow. Establish its baseline. Define the evidence needed for dependable operation. Assign an owner. Fund the redesign around it. Decide in advance how any released capacity will be used.

AI is becoming an ordinary software feature. That does not make AI-enabled management ordinary.

The next chasm is not between sceptics and users. It is between using AI and changing how work gets done.

Evidence note

The sources measure different layers, populations and periods: firm adoption, employment-weighted adoption, worker-task use, intensive use, executive time, reported organisational effects and one designed workflow. They should not be combined into a single funnel or maturity score. The Adoption Stack and the visibility and implementation gap labels are editorial synthesis used to make those differences manageable.