Consider a hypothetical industrial service quotation. An account manager used to spend forty-five minutes assembling the first draft. With AI, it takes fifteen.
The customer still waits four days.
Engineering checks the configuration on Tuesday. Finance reviews the discount on Wednesday. Someone notices an old delivery date on Thursday. The draft got faster; the decision did not.
Now multiply the thirty minutes saved by two hundred quotations a month. A hundred hours. Enough to look impressive in an adoption dashboard, and quite possibly enough to change the business. But the dashboard cannot tell you who gets those hours, what they do with them, or whether a single customer receives a quotation sooner.
That is the question behind my film REBUILT: why does AI make people more productive without reliably moving company profits? The film makes the case for redesigning work around agents. This essay follows the saved hour to the place where that redesign either pays off or disappears.
The decisive moment comes after the model has done its job.
A saved hour is capacity with no destination
If the employee is salaried, saving thirty minutes does not reduce the next payroll. If engineering is the constraint, generating more drafts does not increase the number of approved quotations. If there is no additional demand, clearing the queue faster does not create more orders.
None of that makes the improvement worthless. It tells us what kind of improvement we have bought.
There are several useful destinations for freed capacity. The team can serve more demand without adding staff. It can reduce overtime or external spend. It can improve quality, spend more time with customers, or return some breathing room to people who have been working beyond a reasonable load.
Each requires a different management decision. Each also requires different evidence. The trouble starts when all of them are reported as cash savings.
My rule would be to name the destination before approving the business case. If the gain will support growth, identify the backlog or demand it can serve. If it will lower cost, identify the expense that changes. If it will improve quality, identify the failure that should become less common. If it will improve working conditions, say so and measure that honestly.
An hour cannot be counted twice because two departments found different ways to describe it.
| What you expect to gain | The decision somebody must make | Evidence to watch |
|---|---|---|
| More capacity | Assign the freed time to a real backlog or additional demand | Completed work, service level, and quality |
| Lower cost | Reduce overtime, contractors, or another avoidable expense | Actual spend after the change |
| More revenue | Use the capacity to improve a selling or service constraint | Incremental contribution, with competing explanations tested |
| Better work | Give time back to quality, learning, or a sustainable workload | Rework, customer experience, or employee outcomes |
This is a decision table I propose, not a research benchmark. Its purpose is to stop a claim from changing category on the way to the board.
The gap is real. The headline needs updating.
McKinsey's August 2026 survey makes the distinction visible. Eighty percent of respondents said AI had improved their individual productivity. Thirty-seven percent attributed some enterprise EBIT impact to AI, essentially unchanged from the previous year. Nearly three-quarters of its AI high performers reported fundamentally redesigning workflows, compared with one-quarter of other respondents.
That is a survey association, not proof that redesign caused the financial result. Still, it is a useful place to look for the missing link.
The story has also moved since I published the film. On 30 September, BCG's 2026 Applied AI Index reported that almost half the companies represented in its sample were obtaining value from AI: 7.5 percent classified as future-built and 41 percent as scaling. The research surveyed more than 1,300 executives across over twenty sectors.
BCG's definition of value and McKinsey's enterprise EBIT question are different measures. Putting the percentages on the same chart would suggest a comparison the research does not support. But the newer finding matters: an article claiming that almost nobody gets value from AI would already be behind the evidence.
The argument I would defend is narrower. Useful gains are spreading. Turning those gains into a dependable enterprise result still requires decisions about the work around them.
Microsoft's 2026 Work Trend Index offers another clue. In its analysis of self-reported outcomes, organizational factors such as culture, manager support, and talent practices carried substantially more predictive importance than individual factors. Its widely quoted 67-to-32 split describes statistical association. It is not a finding that two-thirds of every company's financial return comes from organization design.
The practical implication is straightforward: teaching someone a better prompt cannot, on its own, change who is allowed to decide.
Follow the case through the handoffs
Return to the quotation. The drafting assistant has removed half an hour of work. The next question is where the case spends the rest of its life.
I would trace a small sample of recent quotations from request to customer receipt. For each, record time spent doing the work, time spent waiting, and time spent correcting it. Include the awkward cases. Averages can make a process look healthy while the orders that matter most sit untouched.
The map may reveal that AI has improved the cheapest step. It may reveal that the real constraint is missing customer information, a configuration exception, or a commercial decision only one person can make.
Those lead to different interventions. An agent might gather the required information, check the configuration against approved rules, prepare the evidence for an exception, and route the decision to the right owner. A standard quotation might follow a faster path. A consequential promise still needs the person authorized to make it.
Removing that signature would not necessarily improve the process. Giving the signer a complete, checkable case might.
There is a second possibility, and it deserves equal attention: the process needs fewer steps. If two teams inspect the same information because a handoff used to lose it, fix the handoff before automating both inspections. Otherwise the agent becomes an efficient clerk for an arrangement nobody should have preserved.
The unit to redesign is the completed case. Measure from the customer's request to a usable answer, including the work that comes back.
What the examples actually show
Two cases from the film make this point more clearly than a catalogue of agent products.
In August 2024, AWS reported that Amazon used Q Developer to help migrate tens of thousands of production applications to Java 17. It estimated more than 4,500 developer-years of work avoided compared with manual upgrades, and $260 million in annual savings from performance improvements. These are company-reported results. The second figure should not be casually rewritten as payroll saved by AI.
The important operating choice was a bounded class of work with a verifiable finish line: applications upgraded and working. The return had a destination.
Allianz's Project Nemo is similarly specific. Launched in Australia in July 2025, it uses seven agents for low-complexity food-spoilage claims after severe weather. Allianz reports an 80 percent reduction in processing and settlement time. A person makes the final payout decision. The claim concerns a defined process, not an 80 percent improvement across all insurance work.
For me, the human decision is part of why this example is useful. It shows that a workflow can change substantially while keeping accountability at the point where money leaves the business. Autonomy is a design choice inside the process, rather than a prize for removing the last person.
Neither case supplies a guaranteed return for your company. Both offer a better question to ask of your own: what complete class of work could we handle differently, and how would we know it worked?
Who has the right to spend the saved hour?
The uncomfortable part is managerial.
If the employee saves time but has no authority to change the next step, the gain stops at the boundary of their role. If a process owner can change the workflow but cannot reassign capacity, the result may stop there too. A programme can have an executive sponsor and still have nobody empowered to make the decision that converts capacity into an outcome.
I would require one named owner for the whole case, one agreed destination for the freed time, and one person in finance willing to examine the result. That is a proposed operating discipline, not a claim that every improvement needs a steering committee.
The strongest objection is fair: should a useful assistant really have to prove a P&L effect before anyone can use it? Better writing, less drudgery, faster learning, and a calmer working day have value even when they never appear as a separate line in the accounts.
I agree. Cheap, reversible assistance should not carry the same burden as an expensive process transformation. Let people use it, with proportionate controls. The burden belongs to the claim being made. A productivity purchase can justify itself on productivity. A transformation programme promising millions needs a credible explanation of where those millions come from.
Giving people breathing room can be an excellent decision. Calling it a cost reduction while leaving every expense unchanged is a poor one.
One quarter, one complete case
The film proposes a ninety-day cycle. I would use that as a review cadence, not a promise that any workflow can reach production in three months.
Start with a process whose output you can check, whose delays matter, and whose owner can change it. Establish the current position before introducing the agent: volume, elapsed time, rework, cost per completed case, and the customer outcome. Write down the destination for any capacity released.
Then run the redesigned process in shadow mode. Compare its decisions with real cases. Move to supervised execution only when the evidence supports it, with limited access, logs, and a working way to stop. Expand the permitted actions separately from expanding the volume.
At the review, count the cost of checking and repairing the agent's work as well as the work it avoided. Include integration, support, and the people keeping the process reliable. Where possible, compare similar cases handled through the existing process, rather than attributing every improvement to AI.
Scale if the completed case gets better and the promised destination receives the capacity. Change the design if the next queue gets longer. Stop if the quality or economics do not hold.
On Monday, take ten recent cases from your most important AI initiative and trace where each waited. Beside the minutes the model saved, write the name of the person who can change the next handoff. Beside that, write what the business intends to do with the time.
If either space is blank, you have found useful work for the leadership team.
The hour is already saved. Somebody still has to decide what it is for.