Usage of AI inside large companies is now growing and easy to prove. The harder question, and the one showing up in more board packs each quarter, is what it returned.
Most boards do not have an answer. McKinsey's State of AI survey found that only 39% of organisations could point to any EBIT impact from AI at all, and most of those put it below 5% of EBIT.
The Adoption Plateau is the phase where everyone has access, usage is measurable, and business impact is invisible. It happens because the company bought licences and left the work exactly as it was. It is a normal stage rather than a failure, and it usually arrives twelve to eighteen months after a broad rollout.
We see it early because of where we sit. Between April and July 2026 our teams logged 447 meetings with senior leaders across our partners. 42% described the Adoption Plateau in their own words. This is more than those who cited budget pressure (29%), unclear AI strategy or ownership (27%), or resistant middle management (21%).
A VP of IT at a European medical device maker described it as the gap between people using Chat GPT and the company getting value from AI. A transformation lead at a European energy group said usage was stuck at individual experimentation, with no shared way of working.
The cost is easy to size and rarely sized. Microsoft 365 Copilot has a published list price of $30 per user per month. At 20,000 seats that is $7.2 million a year, before implementation (even with a volume discount, it is still a significant expense). It is usually recorded as operating expenditure, so it hits the run-rate in full, this year. The CFO sees that line every month. The return is visible to nobody.
Three things, in this order: redesigning how the work gets done, assigning accountability to a leader who owns both the process and the team, and measuring the impact in financial terms.
Phase one focused on deploying the technology. Phase two is about changing how work gets done.
Phase one covers procurement and enablement: negotiating the enterprise agreement, completing data-protection assessments and works council consultations, establishing a centre of excellence, training employees, and building prompt libraries. Many large companies handled this well. But this phase ends once employees have access to the tools.
At that point, the main barrier to value becomes the cross-functional nature of AI implementation. Delivering impact requires changes across technology, workflows, policies, and teams, but responsibility for these sits in different parts of the organization.
The centre of excellence can encourage adoption but cannot redesign business processes. IT manages the technology but does not own the work. Business units own the workflows but were often not involved in selecting or implementing the tools. As a result, no single function has the authority to make all the changes required.
Put simply: a licence gives people permission to use AI. Creating value requires changing how the work gets done.
Try this: Pull the last 90 days of license data. Then, for each function head, require them to identify a specific decision, policy, or process step that has been concretely altered or improved as a direct result of these AI tools. That specific number of documented changes constitutes your true adoption rate, moving beyond simple usage statistics
In our sample, 19% of accounts said outright that they cannot measure business impact, and that sits underneath much of the 29% reporting budget pressure. A CHRO at a European insurer asked us how to measure impact beyond prompt counts and token usage.
Most companies have built only the first of three levels.
AI technology vendors will build level one (Activity) for you. Levels two and three need a baseline taken before deployment, and that is the step almost everyone skips.
There is a harder problem underneath. An operations director in Germany gave us the clearest version of it: you cannot fire half an hour. Twenty minutes given back to 400 people is 133 hours a week of real capacity, and it shows up nowhere in the accounts. It becomes money only when someone decides where it goes: into a headcount plan, into lower contractor or BPO spend, into absorbing growth without hiring, or into client-facing time that earns revenue. Without that decision, the capacity is real and the saving is imaginary.
Beyond the financial impact, the true engine of adoption is the human experience. If employees feel enabled rather than replaced, they become more engaged, which in turn drives organizational velocity. Driving AI transformation when employees resist is slow, painful, and rarely successful.
Tracking metrics like "champion coverage per department," "behavior adoption," and key employee sentiment elements (e.g. autonomy, confidence in AI, AI fluency) can help boost your AI transformation efforts. These are the leading indicators of the process and financial gains that follow: higher engagement means teams take greater ownership of their workflows, which is the necessary precursor to systemic change.
Try this: Pick one process. Write down the three financial numbers your CFO already tracks, and pair them with one leading indicator of human engagement (like champion coverage or sentiment). Commit to moving the financial metric by a stated amount and the human indicator by a meaningful margin within two quarters, name the executive accountable, and say in advance where the released time will go.
They move beyond the plateau by building six specific structural and human capabilities:
Our partners are direct about wanting mechanics rather than philosophy. The most useful feedback we had all year came from an executive at a German engineering group, who said our mastery of the mechanics of change was on a different level, and that they needed more of it. That is the bar.
Try this: Write one workflow, one owner, one metric with a baseline, one target and one date on a single page. Take it to your next executive committee. If any of the five is missing, you are still in phase one.
The Adoption Plateau says very little about the technology. Most companies sitting on it ran phase one well. Unfortunately, doing phase one well earns nothing on its own, and the licence bill arrives either way.
The choice is simple: continue funding tool access with no further upside, or fund the redesign of work where the real return sits. Delaying this decision costs roughly $360 per seat every year in stagnant license fees.
The Human-AI Collaboration Board Diagnostic. Ninety minutes, four to eight board or executive committee members, live-facilitated. You leave with a scored position against the three gaps above, a measurement chain drafted for one process, and one workflow to move first with the owner named in the room.
Managing Partner & Co-Founder
