A Canadian financial institution recently put generative AI in front of more than 5,000 employees, on the back of a pilot in which participants reported average savings of up to two hours a week, and is still working out how to convert that into anything the organization can see. That outcome is not unusual. It is the default, and it is a design problem rather than a technology one.
I wrote about ATB Financial's unusual candour on this in ATB Gave 5,000 Employees AI. The Hard Part Came Next. That piece diagnoses the gap. This one is about the other half: what actually has to change inside an organization for saved time to become value the business can book.
Saved time does not accumulate. It evaporates.
Start with the mechanism, because it explains everything downstream. When a task that took forty minutes now takes twenty-five, the difference does not collect in a tray somewhere waiting to be spent. It produces a slightly easier day. The work expands gently to fill the space, the pace relaxes a little, the person picks up something that was already queued. That is not laziness. It is how work behaves in the absence of a decision.
Time becomes value only when a constraint moves. If nothing downstream of the task changes, nothing downstream of the task improves.
Redesign the cycle, do not absorb the time
The first move is to pick one process and change its shape rather than its speed. Suppose a review took five days, three of which were document preparation, and preparation now takes half a day. The transformation is not "reviews are faster." It is that the cycle is formally redefined as two days, the queue is re-sized against the new number, the service level published to clients changes, and capacity is planned accordingly.
Someone has to make that call and own it. Without it you do not have a two-day process. You have a five-day process with more slack in it.
Decide who banks the saving
Productivity accrues to individuals. Budgets, targets and service levels belong to business units. Nothing carries the first into the second automatically, and "the organization" is not an owner.
Before rollout, name the unit that will absorb the gain and state what it will do with it. There are only four honest answers: take on more volume without adding headcount, shorten a cycle time, redeploy hours toward work that has been deferred for years, or reduce spend. Any of them is legitimate. Choosing none of them is what produces the result ATB has described.
Baseline before, not after
A baseline cannot be reconstructed. Once the tool is in everyone's hands the pre-state is gone, and every subsequent claim about impact becomes an argument rather than a measurement.
This matters more than it sounds, because the numbers that justify these programmes usually come from a pilot: a few hundred volunteers, reporting their own impressions, in the most favourable conditions the organization will ever create. That is a fine reason to proceed. It is not a measurement of what happens at five thousand people, and it cannot be turned into one after the fact.
Measure two things beforehand. First, intended usage: which team, doing which tasks, how often. Second, the business outcome you expect to move: cycle time, throughput, error rate, cost per unit. ATB's chief people officer has said she would set clear usage benchmarks from the start if she were doing it again. That is the most transferable sentence in the entire case.
Rebalance the role, not just the toolkit
If a role's work is now materially faster, the role has changed. The job description, the performance expectations and often the shape of the team have not caught up. ATB's chief people officer has made the same point from the people side: once the repetitive work is automated away, what remains is more cognitively demanding and more creative, and the organization owes people the skills to meet it. Acting on that is an organizational design decision rather than a technology one.
This is where transformation programmes either become real or quietly revert to being tool rollouts with a communications plan attached.
One process beats everyone
Deploying to an entire workforce maximises adoption and diffuses accountability perfectly. Deploying into a single workflow concentrates both: one owner, one baseline, one outcome, one story you can take to the board and repeat somewhere else.
Enterprise-wide access is a good thing to give people. It is not a transformation programme, and the trouble starts when it is funded and reported as though it were.
The uncomfortable part
Every decision above is a management decision. None of them is a technology decision, and none of them can be made by the team that ran the deployment. That team did its job when the tool arrived and worked.
This is why AI transformation stalls inside capable organizations with good technology and willing staff. The tooling decision has a clear owner, a budget line and a completion date. The operating-model decision has none of those until somebody assigns them, and it rarely gets assigned because it does not look like a project.
Execution, not technology, is where transformation succeeds or fails. The tool arriving is the beginning of the work, not the end of it.