TL;DR: Optional AI tools tend to plateau at about half of users, whatever the enablement behind them. What breaks the plateau is accountability. Wire the AI into mandatory work, something with a deadline and someone accountable for it, and it spreads on its own. Start with what the organisation is already required to deliver. Then redesign the work around it, because the AI changes the job it lands in.
In the banks where I deploy AI into live credit and lending workflows, a handful of people get years ahead. They quietly rebuild how they work around it, and the capability tends to stall there, with them. The obvious move is to make them teach the rest.
I have watched that instinct fall short. The gap it is trying to close is real, and even if those people taught generously, the capability still would not permeate. Permeation in a large organisation cannot run on advocacy alone.
Optional tools plateau.
Give a good team an excellent optional AI tool. Run a proper enablement programme behind it, champions, clinics, the lot. In the banks I have deployed in, adoption climbs to about half the licensed users and stops. "You could use this" competes with everything else on a busy professional's desk, and it loses more often than it wins. You can lead a credit analyst to an agent, but you cannot make them route their day through it.
The reason is structural. Among organisations already using or piloting AI, HBR Analytic Services found in 2026 that only 12% have it embedded in the flow of the work across most of their workflows. The largest group still run it as a tool beside the job, and a tool beside the job is one you can skip.
People are making a rational choice to leave their work as it is while the tool stays optional, and a worked example from a colleague changes nothing about that calculation.
What broke the plateau was accountability.
The step-change at one bank I work with came when we wired the AI into a piece of work the organisation was already accountable for. It was a portfolio-level review of the highest-exposure customers, one every team had to produce, to a real deadline, owned by someone who answered for it, feeding a governance process that existed with or without us.
The agent stopped being a faster way to do an optional task and became what the work now ran on.
It drafted each review from the bank's own data, and the team checked it, corrected it and signed it off. Mandatory work does not make the tool mandatory. The team could have carried on by hand, and if the agent's draft had not been worth starting from, they would have. It stuck because the draft was good, and building it from scratch would have cost them hours they did not have.
In the weeks that work landed, usage stepped up. The reviews completed at pace and went through governance. Teams that had ignored the platform for months were in it daily, because the work they owed the business now ran through it.
Start with a commitment the organisation already has. Find the report someone has to file, the review a regulator expects, the committee paper in the diary, and wire the AI into that. The accountability is already in the building, and the named person still answers for the result.
Now, not every commitment is worth it. The ones that are have a real deadline, someone who owns the outcome, a reason it comes round again, and a cost to skipping it that gets noticed.
This is also why enablement programmes take you only so far. Champions and clinics do real work, and they get you to about half. Past that, the lever is accountability.
But that lever comes with a cost, and I live inside it too. I run my own function on a system I call Juno, a team of agents with a memory that keeps what they learn, and it has taken the routine work off my desk. What is left is the reviewing. My day is the judgement calls and catching what Juno gets wrong, which was always the part that needed a human. The more Juno takes on, the more comes back to me to weigh, and the day quietly gets denser. I built it to give myself room, and the room filled with the decisions only I can make.
Where this leaves us
Two things follow, and the second is the price of the first.
If you want AI adoption to take in your organisation, attach the capability to work the organisation is already on the hook for. That is the first move.
The same move that solves your adoption problem changes the person's role, and the redesign it forces is the part I rarely see planned for. Hand the routine work to the agent, and it also starts to hold what used to live in people's heads. That is a genuine gain. The expertise that once walked out of the door with every leaver now sits in a system that can retain it, decode it and pass it on.
What shifts underneath is bigger than a tool. The person sits at a new point in the process, touches different systems, and the knowledge that used to move hand to hand now runs through the agent. The processes, the controls and the lines of accountability were built for the way the job used to work.
Leave them there and the work fails at the joins, a check sitting a step from where the call now gets made, knowledge the agent holds that never reaches the person downstream who needs it.
The move that works is to do both. Wire the AI in, then redesign the work around the role it changes. Do the first and skip the second, and you have swapped an adoption problem for a harder one.
Questions I get asked
Why do optional AI tools stall at around half the users? Because "you could use this" is competing with everything else on a busy professional's desk, and an optional tool usually loses that competition. People make a rational choice to leave their work as it is while using the tool stays voluntary. A great enablement programme, champions and training move the number, and then it settles.
What actually breaks the adoption plateau? Accountability. The step-change comes when the AI is wired into work the organisation is already on the hook for, the report that has to be filed, the review a regulator expects, the committee paper already in the diary, with a deadline and someone accountable for it. Once the mandatory work runs through the agent, usage steps up, because the work carries it across on its own.
Isn't this just mandating the tool? No. Mandating a tool buys compliance, people opening it to tick a box while the real work happens elsewhere. The work here is already mandatory, and the team can still do it by hand. The agent becomes how the work gets done only if its draft is good enough that doing it by hand is the harder path. Accountability puts the AI in front of real work with real stakes, and being good enough to keep there is still on the AI.
Does getting your superusers to share their methods fix it? It does real work and it is worth doing, and on its own it will not move the plateau. Permeation in a large regulated organisation is a structural question. A colleague's worked example changes nothing about the calculation a busy person makes when the tool is still optional. Wire the capability into mandatory work and it spreads without anyone the need for persuasion.
What is the catch with the accountability approach? The same move that fixes adoption changes the person's role, while the process, the systems and the information flow around it were built for the old one. Leave those in place and the work breaks at the joins, the checks and handoffs no longer sitting where the work now happens. The catch is the redesign it forces. Reshaping the work around the new role is the real job.
