On October 8, 2026, at its Gemini at Work event, Google gave a room of enterprise buyers a run of numbers. Nearly 90% of the Fortune 100 use Gemini Enterprise. Nearly 500 Google Cloud customers have each processed more than one trillion tokens. SOMPO is deploying more than 10,000 custom agents. Futurum's write-up the next day added one more line that I read three times: Commerzbank cut manual document review from 20 hours to one. It's a lovely number, and it's incomplete. It tells you where the work went. It says nothing about where the people went.
Say a reviewer at that bank had a working week built around that review. Nineteen of those hours have just turned into air. Somebody has to decide what fills that air, and in most companies, nobody has been given the job.
A week before the Google event, on October 1, EY published its Global CEO Outlook. It surveyed 1,200 large-company CEOs in 21 countries, with fieldwork between August and September 2026. Half of them named AI as the single largest contributor to material productivity gains. Only 16% said they had strong confidence in measuring AI ROI. Almost a quarter, 23%, said they struggle to turn those gains into measurable financial outcomes. Andrea Guerzoni, EY-Parthenon Global Vice Chair, summed it up in a sentence worth repeating: "AI is creating real productivity gains, but productivity alone is not a strategy."
The gap between those two facts is where companies lose money, and an AI transformation roadmap for executives has to contain a way to close it. My position is simple and a bit unfashionable. Most roadmaps are supply plans. They schedule the arrival of capability: the pilot, the platform, the agent fleet, the training module. They almost never schedule where the hours that capability frees will go. An hour saved with no destination is an hour lost. You don't see the loss because nobody booked the hour in the first place.
Freed time behaves like a liquid
BCG's fourth annual AI at Work survey covered nearly 12,000 workers across more than a dozen markets. PRESS Insider's March 2026 summary reported that 42% of regular AI users save at least eight hours a week, which is a full working day. In the same survey, 66% of frontline employees said they get limited or no guidance on how to use the time they save. BCG warned that the savings could "leak" out of organizations unless companies deliberately measure and reinvest them.
Leak is the right verb. Most companies have no container for time. Money sits in accounts with owners and approval limits. Headcount sits on an org chart with names next to the boxes. An hour freed inside a role sits nowhere, so it runs down to the lowest point of the week. It ends up in the inbox, in the meeting that grew from thirty minutes to fifty, or in a second pass on a deck nobody asked anyone to revise.
The Indeed Hiring Lab data shows where some of it lands. Its survey reached more than 80,000 people in eight countries in May and June 2025, and HR Dive reported the findings on January 8, 2026. More than 8 in 10 employed workers who use AI said they save at least one hour per day. About a third (32%) of U.S. workers who saved time with AI used it for work-life balance. Almost a quarter (23%) put it toward breaks and stress management.
I don't read that as a scandal. People are tired, and an hour of recovery is a reasonable thing for a person to take when the employer hasn't offered anything better. It does tell you the default destination of a freed hour: the worker's own recovery. A company that has promised its board those hours will show up as margin has made a promise its operating system can't keep. Yusuke Aoki, an economist at Hiring Lab, put the adoption half of this plainly: "Simply using AI is not enough." The reallocation half is plainer still. If nobody plans where the hours go, the people holding them will decide, and each of them will decide alone.
The roadmap ends where the vendor's metrics end
So why do so many roadmaps stop at deployment? Because the instruments we inherited measure supply.
Go back to the Gemini at Work numbers. A trillion tokens per customer, ten thousand agents, Fortune 100 penetration. Each one counts something being installed or consumed. These are honest metrics for a company that sells capability, and Google is right to report them. The trouble starts when a buyer copies them into its own steering committee deck. "Agents deployed" turns into a milestone. "Tokens processed" turns into evidence of adoption. The board gets a slide that looks like progress and reads like a utility bill.
The EY data shows the same lopsidedness inside companies. More than half of CEOs (56%) reported at least some reliable visibility into AI costs and returns. Only 16% had clear, real-time visibility of ROI. Almost a quarter (24%) could track some AI costs. The cost side of the ledger is getting instrumented much faster than the return side, and the reason is mechanical. Costs arrive as invoices, and invoices come with numbers on them. Returns arrive as hours, and hours come with nothing attached.
Google's own announcement makes the asymmetry hard to miss. Futurum reports that the new Gemini agent, which Google describes as "a single, universal agent for work," comes with real-time, per-project spend caps that pause an agent when a cap is hit. Persistent coworker agents get their own @agents.company.com email address, storage and Workspace account, and act under that identity. There are per-agent audit trails and an Agent Gateway that Google calls "an AI network firewall."
Lay that list next to the human hour the agent just freed. The agent has an identity, a budget, a ceiling and a log of everything it touched. The hour has no owner, no price, no ceiling and no log. We have built better bookkeeping for the machine than for the time it hands back to us, which is a strange thing to have done by accident.
What an AI transformation roadmap for executives should contain
The fix is a design choice. It doesn't need a new tool. I call it a destination ledger, and it changes the order in which a roadmap gets written. Today the sequence usually runs: choose the use case, run the pilot, scale the winners, measure the savings. The destination ledger moves the last step to the front.
Write the destination before the deployment
Every automation proposal gets one extra required field: where will the freed hours go, and who will notice if they don't arrive? I accept three honest answers.
The first is harvest. The company buys fewer hours. That can mean a hiring freeze in the affected team, natural attrition left unbacked, a contractor budget that shrinks, or overtime that stops. Harvest is the easiest to verify and the hardest to talk about, which is exactly why it gets left out of the deck and then quietly expected anyway.
The second is redeploy. The hours go to named work that's already waiting. A compliance backlog of a specific size, a market the team has wanted to test for two years, a customer segment nobody had capacity to call. The test is whether the new work existed as a written list before the automation shipped. If it only gets invented afterwards, it is leakage that someone has given a name.
The third is compound. The company sells the speed. Faster turnaround becomes a priced service tier, a shorter quote cycle, a guarantee in the contract. Of the three, compounding is the one most likely to show up as revenue, and it's also the rarest, because it needs sales and product to sign up for a promise operations has only just become able to keep.
A proposal that says "capacity will be reinvested" without naming the work goes back to whoever wrote it. I'm strict about this because the vague version is what EY's 23% look like from the inside.
Price the hour before you free it
Suppose a fully loaded analyst costs $90 an hour and an automation frees 19 of her hours a week across 46 working weeks, which comes to 874 hours or about $78,660 of capacity a year. That figure is invented for illustration, and its value lies in being written down before the project starts. Once a number like that sits in the proposal, the business case reads differently. The software license stops being the main cost. The bigger risk is that $78,660 of capacity turns, quietly, into a calmer inbox.
Pricing the hour also changes which projects win. A use case that frees four hours from a senior underwriter can beat one that frees twenty hours from a role with no backlog behind it. Hours are worth whatever the work they can move to is worth. Roadmaps that rank use cases by raw time saved end up favouring the hours that have nowhere to go.
Give someone the job of spending it
In most companies the money and the hours belong to different people. IT or a central AI team pays for the platform. The line manager absorbs the freed time. Finance goes looking for savings at budget season. Each of the three holds a third of the story, and the gap between them is exactly where the leak happens.
The destination ledger needs one owner per function, and that owner answers for both sides: what the AI costs and what happened to the hours it freed. You don't need a new title. You need one line added to an existing executive's quarterly review. Where did the hours go, and can you show me? If the answer is "the team feels less stretched," that is a legitimate answer, and it should be written down as a choice to spend the capacity on resilience. A conscious choice to spend it on rest is fine. Spending it without anyone noticing is how the money disappears.
Shrink something on purpose
EY found that nearly half (48%) of leaders are reinvesting the additional capacity. That sounds healthy until you see the next numbers. Only 15% say overall productivity gains are being reinvested in workforce development. And 69% say their growth strategy is focused more on strengthening the existing business.
"Strengthening the existing business" can absorb any number of hours without leaving a trace, which makes it the best hiding place leakage has. A team freed from twenty hours of manual review will find twenty hours of existing business to strengthen by the following Monday. Some of that will be real. Much of it will be work that grows to fill the room it's given.
Deletion is the only kind of reallocation you can check. Every AI transformation roadmap should include a list of things the company will stop doing because the automation shipped: a report, a meeting, a sign-off step, a vendor contract, a whole layer of review that existed to catch human errors the machine no longer makes. If the roadmap only adds things, the freed hours will spread through the organization in a thin film you can't measure.
The skills shortage and the leak are one problem
The most striking EY number has nothing to do with productivity. 72% of CEOs expect skills shortages to become a greater barrier to growth than access to capital. Nearly half (47%) say their organizations are not developing AI skills quickly enough. The BCG survey shows the same gap from below. 72% of workers believe AI has changed the skills their roles require, and only 36% feel they've received adequate upskilling.
Put those numbers next to the leak and the picture gets almost funny. These companies say skills are scarcer than money. Their employees say they have been freed by up to a full day a week and given no instructions for it. Only 15% of CEOs say productivity gains are being reinvested in workforce development. The training budget they keep saying they can't find is the time their tools already hand them every week.
I'd go further. Freed hours are the cheapest reskilling budget a company will ever get, because the hours are already paid for. The person is already on payroll, already knows the business, and already holds the context that a new hire would take six months to build. A roadmap that frees capacity in one quarter and requests new training headcount in the next is buying the same asset twice.
A practical version is to pre-commit a share of every freed hour to deliberate learning before the automation goes live, so the commitment isn't renegotiated each time a deadline looms. Say a team agrees that a quarter of the time a new agent frees goes to building the next skill the function will need. The number isn't the point. What matters is that the agreement exists before the hours do, because once the hours exist, the inbox has a stronger claim on them than any training plan.
Agents with addresses, hours without them
Google's coworker agents tell you which way enterprise software is heading. They get identities, permissions inherited through OAuth, storage, a budget and an audit trail. They have, in a literal sense, an address. A persistent agent at a large company is now easier to locate inside the org than the capacity it creates.
I would steal that design for the human side. Give freed capacity the same four properties the agent gets. It needs an identity, meaning a named owner. It needs a budget, meaning a price per hour and a total. It needs a ceiling, meaning a limit on how much can go to unplanned work before someone has to sign off. It needs a log that records, at the team level, where it went. None of this needs software. A spreadsheet reviewed monthly by the function's capacity owner is enough to start. The discipline comes from asking the question on a schedule. The tooling matters much less.
SOMPO deploying more than 10,000 custom agents is a remarkable piece of supply. The question I would ask SOMPO's leadership, and any company running at that scale, is how many of those agents have a written destination for the hours they free. If the answer is a small fraction, the company has bought a fleet of very capable machines and pointed them at an unmarked drain.
The objection from the floor
The fair pushback from people who have run operations is that tracking freed hours turns into surveillance. Employees will hide time, managers will invent work to look busy, and the ledger becomes a performance.
I agree with the risk and disagree with the conclusion. The cure is to measure at the destination and keep the measurement away from the individual. Skip the question of how many minutes Priya saved on Tuesday. Ask what this team shipped this quarter that it couldn't ship last quarter, and what it stopped doing. Those questions are about the organization's choices, and they can be answered without a timesheet. Teams that use freed hours to rest can say so, and the company can decide whether that's the trade it wants. The ledger asks the company to spend the hours on purpose. It doesn't require spending them on output.
The second objection is that it's too early. The tools are changing every quarter, so why build accounting around a moving target? Because the leak is happening now. BCG's 42% of regular users saving a full day a week aren't waiting for the tools to settle. The hours are being spent this week whether or not anyone wrote down where. A company that waits for stability before deciding where capacity goes will find, when it finally looks, that the capacity has already been spent, mostly on things it would never have approved.
Why this sits with the CEO
The reason this belongs on the executive agenda, rather than with the AI team, is that the destination of capacity is a strategy decision dressed up as an operations detail. Harvest, redeploy and compound lead to three different companies a few years out. A firm that harvests becomes leaner and more defensive. A firm that redeploys gets wider and takes on adjacent work. A firm that compounds turns speed into a product and charges for it. Each is defensible. Choosing none of them by default is the expensive option, and that default is where most of the surveyed CEOs appear to be sitting.
Guerzoni's line about productivity holds up on a second reading. Productivity is a rate. Strategy is a choice about what the rate is for. A roadmap that ends at deployment hands that choice to thousands of individual employees and lets them make it one quiet hour at a time.
Build the ledger before you build the fleet
The vendors are doing their part. In one morning, Google shipped agent identity, spend caps, audit trails and a firewall for agent traffic. More of that is coming from every major platform, and it will be good. None of it will tell you where your freed hours should go, because no vendor can sell you your own org design. That part has to be designed by people who know your backlog, your margins, your skills gaps and your appetite for shrinking things on purpose.
That's why an AI transformation roadmap built around tool purchases will keep producing EY's numbers: real productivity gains and fuzzy returns. The companies that turn productivity into growth will be the ones that write the destination ledger first. They will price the hour, assign an owner, name what gets deleted and pre-commit capacity to the skills they say they can't hire. This is architecture work, and it has to happen before the next wave of agents arrives with its own email addresses and budgets.
Agor AI Advisory builds that architecture with executive teams. We map where your freed capacity is leaking today and design the destination ledger function by function. We rewrite the roadmap so every deployment arrives with a written answer to the question nobody asked at Gemini at Work. Your hours already exist. The only open question is whether you or your inbox decides where they go.
Schedule a strategic consultation with us today.
Sources
- EY Newsroom, "AI drives productivity, but CEOs struggle to turn gains into growth," October 1, 2026
- Futurum, "Google Collapses Enterprise AI Into a Single Gemini Agent at Gemini at Work 2026," October 9, 2026
- PRESS Insider, "AI saves workers time that companies do not know how to use, study shows," March 2026
- HR Dive, "Workers use time saved by AI to improve their roles, Indeed finds," January 8, 2026