AI adoption is becoming a terrible measure of AI progress

The dashboard can be green while the business result is missing

Imagine the quarterly AI update. Approved tool access has expanded. Monthly active use is up. Training completion is high. A growing share of employees say they use AI every week. Every indicator on the adoption dashboard is moving in the right direction.

Then someone asks the uncomfortable question: what changed?

Did customer response times improve? Did a sales process become faster without lowering quality? Did a compliance workload fall? Did a product team shorten its cycle time? Did a service become easier to deliver? Did the organisation release capacity that was actually redeployed somewhere valuable?

If the answers are vague, the dashboard may be measuring contact with AI rather than progress from it.

That distinction is becoming more important as access and routine use spread. Early in a technology cycle, adoption can be a useful proxy because simply getting the tool into people's hands is a meaningful hurdle. Once use becomes common, the same metric becomes less informative. The organisation needs to know not only who is using AI, but whether use is crossing the boundary into changed work and measurable outcomes.

Section's 85 per cent claim is a useful signal, not a neutral fact

Section AI supplied the immediate trigger for this article. Its January 2026 AI Proficiency Report surveyed 5,000 knowledge workers in organisations with more than 1,000 employees across the United States, United Kingdom and Canada. The accompanying blog post argued that 85 per cent of reported AI use cases were unlikely to generate business value. ¹

The important word is judged. Section classified respondents' use cases against its own view of what is likely to produce return on investment through significant time savings, revenue impact or other business measures. The research did not follow thousands of workers into their organisations and measure the realised financial value of each use case.

That does not make the result useless. It makes it a commercially interested signal that needs to be handled carefully. Section sells AI transformation planning, strategy support, coaching and proficiency services. Its definition of useful progress naturally points towards deeper training, workflow redesign and organisational enablement.

The more useful finding is therefore not that 85 per cent of workers create no value. It is that weekly use and business value can diverge sharply, depending on what is being counted as value. Section itself sharpened this point in June. Weekly AI use in its survey had risen from 55 per cent to 67 per cent, yet 73.5 per cent of workers were still classified as experimenters using AI mainly for one-off tasks. Again, that is Section's proficiency model, not an independent market standard. But the direction is worth testing against other evidence. ²

Independent evidence shows the same gap from a different angle

Gallup provides a useful comparison because it asks workers about both personal productivity and organisational change. As of May 2026, 65 per cent of employees in organisations that had implemented AI said it had a positive effect on their productivity and efficiency. Yet only 14 per cent strongly agreed that AI had transformed how work gets done in their organisation. ³

Those two findings can both be true. An employee can draft faster, search faster, summarise faster or prepare a first pass more easily while the surrounding approvals, queues, meetings, handovers, systems and responsibilities remain exactly as they were.

Firm-level evidence points in the same direction. An October 2026 revision of the NBER working paper Firm Data on AI surveys nearly 9,000 senior executives across seven countries. It reports that 78 per cent of firms use AI, while executives estimate that adoption raised annual labour-productivity growth by only 0.2 percentage points over the preceding three years. This is a revised, later survey wave; the earlier four-country figures in the original article are not treated as current results. ⁴

None of this proves that AI creates little value. Measurement lags, adoption is recent, and many benefits may be local, qualitative or still too small to appear in firm-level productivity. It does show why usage is a poor substitute for an outcome measure. The more AI becomes normal, the less a usage count tells management about what the organisation has actually gained.

The missing step is not more adoption. It is conversion

The central management problem is conversion from one kind of evidence to another.

Access is converted into activity when employees actually use approved tools. Activity is converted into useful work when AI improves a defined task. Useful work is converted into organisational capability when the improvement becomes repeatable, supported and owned as part of a workflow. Capability is converted into value only when the changed workflow affects an outcome the organisation cares about.

Each step can fail independently.

A licence can sit unused. An active user can spend time on trivial tasks. A good personal use case can remain trapped in one person's prompt history. A repeatable workflow can save time without improving a business or service outcome. A genuine time saving can disappear because no one decides what the released capacity is meant to do.

This is why simply pushing adoption harder can miss the problem. If the blockage sits between task improvement and workflow redesign, another training drive may increase activity without improving conversion. If the blockage sits between workflow improvement and organisational outcome, the missing decision may be about process ownership, staffing, priorities or how saved capacity is redeployed.

Geoffrey Moore described a related adoption problem in Crossing the Chasm. Pragmatic mainstream customers do not adopt a novel technology simply because its core capability works. They need what Moore called the "whole product": the integrations, support, processes and other surrounding elements required to use it successfully in their own environment.

AI creates a closely related organisational gap. An individual employee can improvise the missing pieces around a chatbot, but an organisation cannot turn that improvisation into dependable capability until it builds the workflow, ownership, controls, measures and support around the technology. The core tool can therefore create personal utility before the organisation has assembled the whole product required to capture value consistently. In a 2026 discussion of agentic AI, Moore sharpened the point by arguing that chasm crossing becomes visible when an entire function re-engineers itself, rather than when a tool merely accumulates users. ⁷

Workflow change is where the evidence becomes more interesting

McKinsey's July 2026 readiness survey offers supporting evidence, with the usual caveat that it is consultancy research based on self-reported responses rather than a neutral census. It surveyed 750 employees and leaders across regions and found that leaders were 5.3 times more likely to report enterprise value capture when workflows had been redesigned than when they remained unchanged, 32 per cent versus 6 per cent. ⁵

The useful part of that result is not the precision of the multiple. The sample was deliberately constructed to include organisations at different maturity levels, and the enterprise-value questions were answered by leaders. The useful signal is that respondents reporting value also report changes in how work is organised, not simply wider access to tools.

That matches the difference visible in the Gallup data. Individual productivity can improve inside an unchanged system. Enterprise value usually requires the system to react.

The same issue appears after the saving has been created. Gartner reported from a July 2025 survey of 114 HR leaders that only 7 per cent of organisations provided guidance on how employees should use time saved by AI. That is a small survey and should not be treated as a universal prevalence estimate. It does, however, identify a very practical management failure: if an hour disappears from a task but the organisation has made no decision about the hour, the benefit may never become an organisational outcome. ⁶

What a better AI progress dashboard should measure

A useful dashboard does not need to abandon adoption metrics. It needs to put them in their proper place.

Access, active use and training are leading indicators. They tell management whether the organisation has created the conditions for AI to be used. They do not tell management whether the use is consequential.

The next layer should measure conversion. What share of active users have at least one defined use case with a baseline? What share of useful use cases have become repeatable workflows? What share of those workflows have named owners, quality measures, exception handling and a business or service outcome? What happens to time or capacity released?

This creates a much more demanding conversation, but also a more useful one. A low conversion rate from access to activity suggests an enablement or relevance problem. A low conversion rate from activity to useful task suggests poor use-case discovery. A low conversion rate from useful task to repeatable workflow suggests an operating or integration problem. A low conversion rate from workflow to outcome suggests that the organisation has improved work locally without connecting the improvement to value.

The AI Value Conversion Test

Use this as a management dashboard, not a maturity score. The question is whether value survives each conversion step from access to outcome.

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

Do not replace one vanity metric with another

There is an obvious danger in responding to weak adoption metrics by inventing a new universal AI score. Different functions create value differently. A faster legal review, better marketing research, lower contact-centre handling time and improved clinical administration do not belong on one simplistic scale.

Nor should every individual productivity gain be forced into a financial calculation. Some benefits matter because they improve quality, resilience, employee experience, service access or decision speed. The point is not to reduce AI to a single return-on-investment number.

The point is to preserve the chain of evidence. Management should be able to say what changed at the task level, what changed at the workflow level and what outcome the organisation expected to move. Where the benefit is qualitative, name it. Where it is speculative, say so. Where no outcome has been defined, call the initiative an experiment rather than a success.

The question for leaders is changing

A year ago, asking how many people were using AI could reveal whether the organisation had moved beyond discussion. That question still has value where access remains constrained or use is negligible.

But in organisations where AI use is already common, adoption is becoming a weak management KPI. It rewards motion even when the operating model is unchanged.

A better question is: how much of our AI activity is converting into repeatable work with a measurable organisational consequence?

That formulation is less flattering to dashboards because it exposes the gaps between tool availability, employee behaviour, workflow ownership and management decisions. It is also much closer to the reason organisations bought the technology in the first place.

The next phase of AI management is therefore not about making adoption disappear as a metric. It is about demoting it. Access and usage belong at the top of the funnel. The real management work begins when leaders ask what survives the journey to the bottom.

Executive view

• Adoption metrics remain useful lead indicators, but they are increasingly weak proxies for organisational progress.
• Section AI’s 85 per cent figure is its judgement of reported use cases, not a neutral measurement of realised value.
• Personal productivity gains are more common than organisation-wide changes in how work gets done.
• Track conversion from access to useful tasks, repeatable workflows and outcomes; decide what happens to released capacity.
• A better dashboard measures where that chain breaks, rather than treating usage as the destination.