Uptake Is Uneven and the Workforce Is Not the Reason: What AI Adoption Actually Depends On

14 September 2026 · 7 min read

AI adoption in operations now follows a pattern that would have seemed contradictory two years ago. Access to AI is everywhere. People use agents in their own work by choice, and embedded AI sits inside the platforms the operation already runs. Yet when a capability is deployed into a workflow, with the same licence, the same onboarding and the same internal training for everyone, daily uptake still lands unevenly. Deloitte’s 2026 State of AI in the Enterprise research captures the shape of it: worker access to AI tools expanded by 50% in a year, yet fewer than 60% of workers with access use AI in their daily workflow.

A common response is to read this as a people problem. So what follows is predictable: a renewed training push, a nominated champion, or a round of internal communications explaining why AI adoption matters. When uptake stays low after all that, the conclusion tends to be: the workforce is resistant, change-averse or simply not ready.

In 2026 that reading fails on its own evidence. The people leaving the deployed capability untouched are often the same people using agents in their own work every day, unprompted. A workforce that adopts AI by choice in one place and declines it in another is not resistant. It is responding to something about the work. The question worth asking is what is actually different about the consistent users, and the answer points to who has redesigned their own work. Enterprise AI adoption is a design problem before it is a capability problem.

The consistent users: what their work looks like

Look closely at those who use a deployed capability every day and something consistent shows up. It is not simply enthusiasm or technical confidence, although both help. It is that they have, in effect, redesigned their own work around it. They have found the point in their process where the capability belongs, decided what output they expect from it and built their own routine for checking and using that output. The embedding is real, but they built it themselves, privately, one person at a time.

That is why uptake is uneven. Where using a capability depends on each person designing it into their own work, usage follows personal disposition, role fit and workload pressure, and those vary across any team. The operation has, in effect, delegated workflow design to individual choice.

For the rest of the team, the picture is different. The existing process is still intact, still works and still produces acceptable results. Using the deployed capability potentially means running a parallel step alongside the original workflow, producing an output that is not visibly different from the output of skipping it, at least not in any way the current measures capture. That is optional extra effort, and most people, reasonably, do not take it. Their own agents, meanwhile, get used precisely because they serve the work each person already has to do.

What the evidence shows

None of this needs a resistance theory, and the people involved are not obstructing the organisation’s AI goals. Where the old process still produces an acceptable result and the new capability adds a step, most people carry on with the process the job actually requires. The more useful question is what happens where the work itself changes.

BCG’s June 2026 AI at Work survey of 11,749 workers puts numbers on that difference: companies pursuing workflow redesign alongside deployment are 24 percentage points more likely to see measurable business improvement and 22 percentage points more likely to save employees at least a full day a week. The return sits where the work was redesigned, and uptake travels with it.

Mostly, though, the work has not changed. Deloitte’s 2026 research finds 84% of organisations have not redesigned jobs or workflows around AI, which is why the uneven-uptake pattern is the norm rather than the exception.

The period immediately after a capability is introduced is usually an addition, not a transition. The old process continues because the role still requires it, the measures still reflect it and nothing has defined what changes. Until the existing process is redesigned, the new capability sits beside it as a supplementary option.

Training and champions treat the symptom, not the condition

The people-problem diagnosis is appealing because it points to something fixable. More training, a committed champion, a structured change-management campaign: these are known quantities with established budgets and timelines. They produce activity and they are straightforward to measure. What they do not do is change uptake when the work itself has not been redesigned, and that is the condition that matters most. Change-management research, including Prosci’s work on why AI transformations stall, reaches the same place: behaviour is shaped by the operating environment, the incentives, workflows and measures around the person, and when those remain unchanged, old behaviours survive even where people are willing.

When Envisago reads an operation’s AI adoption through AIVOM™, the AI Value Operating Model, uneven uptake commonly surfaces inside the Design and Capability dimensions at the same time. The capability gap is real, but in our experience it is rarely the primary cause. More often the work has not been redesigned, and no amount of training changes that until the design condition is addressed.

Adoption follows redesign: when uptake stops being a choice

When the task sequence requires the capability, adoption stops being a behavioural outcome and becomes a structural one. Operation-level redesign does deliberately, and for everyone, what the consistent users did privately for themselves: a defined point where the capability is picked up, a clear output it is expected to produce and a handoff that depends on that output being there. The person using it is completing the job as it is now designed. This is the shift that separates teams with minimal uptake from teams with near-universal usage.

Role clarity matters as much as workflow integration. If the person does not know what judgement they are expected to apply to the output, the capability produces uncertainty rather than speed. The handoff that follows depends on someone knowing what the AI has done and what human review is required. Where that is undefined, ambiguity appears at exactly the point speed should.

The measures must also move with the work. If success is still tracked against metrics built for the old process, the new capability produces nothing the organisation can see, and invisible value does not sustain adoption. The same design conditions shape AI readiness across the organisation and the governance around what AI is permitted to do.

Finding the workflows where redesign has the most to gain

The first question is not which teams need more training. It is where the work has already been redesigned and where it has not. That reading distinguishes teams that are structurally blocked from teams that are genuinely capability-constrained.

The highest-impact candidates share a profile: high volume of AI-relevant work, an old process still intact and nobody yet owning the redesign task, with measurable friction and a direct link to performance outcomes the organisation already tracks. How to prioritise those candidates at the workflow level is its own discipline, and the practical test in each case is whether AI changes only how fast a step is done, or whether it changes which steps exist, who does them and how work moves through the operation. The first is deployment without process change. The second is the redesign that sustains adoption.

Start with understanding

Uneven AI adoption in operations is the predictable outcome of deploying capability without redesigning the work around it. The pattern repeats because the operating condition repeats, and it repeats now in operations full of willing, self-taught AI users. Training improves what people can do, and the job continues to require what it always required until the design changes.

Start with understanding which parts of your operation are structurally blocking adoption and which workflows are ready for redesign. Without that understanding, investment in capability runs ahead of the design conditions that would make it useful, and uptake stays uneven regardless of the effort applied.

Envisago’s free AI Operating Impact Briefing is a structured reading of where your operation currently stands across the four AIVOM™ dimensions, Value, Design, Capability and Performance, including the operating conditions shaping uptake. It names your priority areas and one clear place to begin.

Start your free Briefing at aivom.envisago.com.