What AI Changes After the Demo Ends
Who gets the time AI saves, and who inherits the work it creates?
November 8, 2026Read on Substack
I have spent enough years building and selling technology to businesses to know what a good product demonstration can do. The data is ready, the task is carefully chosen with the client in mind, and a result that used to require considerable effort appears in a fraction of the usual time. People are right to be impressed. Some of these capabilities are genuinely remarkable.
But a demonstration cannot show how the organization will eventually have to live with the product. Once the software becomes ordinary, other questions emerge. Who checks what it produces? Who manages exceptions? Who changes the surrounding process? And what happens to the time everyone was promised they would save?
On Monday, Web Summit opens in Lisbon. The agenda includes a session called “The orchestration gap: Fixing the enterprise’s AI problem.” That is a revealing shift in emphasis. We have spent years watching what AI can produce. Now the consequences of putting that capability to work are becoming harder to ignore.
A Gallup analysis published in October makes the tension concrete. In a survey of 7,845 American employees who had used AI at work at least a few times in the preceding year, 63 percent said it helped them complete tasks faster. At the same time, 31 percent said their AI use had led their employer to ask them to take on more responsibilities. Among respondents who said AI made them faster, that second figure rose to 37 percent.
These are employees’ reports, not measurements of actual time saved or proof that AI caused a company-wide increase in workload. Many respondents reported benefits beyond speed. Still, the answers reveal a question that rarely appears in a demonstration: when technology creates additional capacity, who decides what that capacity is for?
There is another number in the Gallup data worth reading alongside the productivity results: 52 percent of all employees surveyed said they had less influence over the adoption of new technology than they wanted. That does not show that AI made them less autonomous. It points to the decision-making arrangements beneath the speed gains.
People can become more productive within a workflow while having little say in what that workflow asks them to produce next. The allocation of the gain is an organizational choice, often made by someone other than the person who generated it.
Where The Saved Time Goes
Consider a task that previously took four hours and can now be completed in two. The immediate gain is easy to describe. But those two hours do not arrive with instructions for how to use them.
They might become time to understand a difficult problem, check the work more carefully, serve another customer, or finish the day earlier. They might also become an expectation to produce twice as much. None of these outcomes follows automatically from the technology. They depend on what the organization values, measures, and asks of the people using it.
A field experiment by Eleanor Dillon and her colleagues gives this distinction some empirical weight. Across 66 firms and 7,137 knowledge workers, the researchers randomly assigned access to a generative AI tool integrated into familiar workplace applications. In the latter half of a six-month experiment, workers who actively used the tool spent around two fewer hours per week on email and reduced some work outside regular hours. Yet the study did not detect corresponding changes in the quantity or composition of work tasks.
That is an important positive result. A useful tool can genuinely reduce a recurring demand without simply creating another one in its place. It is also a reminder that the changes people can make individually do not necessarily reorganize the work they share with others. The study found little evidence that giving individual workers access to AI changed their meeting patterns.
One plausible reason is the difference between individual tasks and shared coordination. Email can be drafted or processed by one person; a meeting depends on the availability and expectations of several. Finishing correspondence earlier does not, by itself, change the standing meeting in the calendar or a decision coming from another department. Individual efficiency is a real gain, but shared capacity depends on arrangements that an individual tool cannot redesign on its own.
The results do not tell us what every company should do with saved time. They demonstrate why it is worth separating two questions we often collapse into one: did the tool improve the task, and did the workplace become better as a result?
I use AI extensively in my own research, business, and editorial work. I would not want to return to producing everything without it. But an increase in capacity can quickly become a new baseline. Once a faster turnaround is expected, the original time saving may no longer be experienced as available time. It has been absorbed into the next commitment.
There can be good reasons to use the gain for more output. The problem begins when that choice is never made consciously, and the people doing the work have little agency.
The Work That Changes Hands
A finished document can be generated faster than a reader can determine whether it is correct. An AI system can prepare a customer summary or marketing plan in moments, but someone still needs to establish whether it reflects the actual customer, the right evidence, the latest commitments, and the consequences of a mistake.
As production becomes cheaper, verification may become more important. Some checks can be automated. Others require someone who understands the context and has both the time and authority to challenge what the system produced. When that work is omitted from the implementation plan, the efficiency calculation is incomplete.
Marketing technologist recently reviewed four AI-led marketing automation platforms after attending vendor demonstrations. One showed how a business objective could become a proposed customer onboarding journey using synthetic data. It was a striking change in what the software could assemble, but she raised a practical question: who would audit a 42-step workflow built by AI? Someone still has to determine whether the customer segments, permissions, business rules, integrations and exceptions make sense before such a journey reaches real people.
That question sits alongside another decision she was making: would the new software justify leaving a platform that teams had spent years learning, integrating and maintaining? There is a cost to migrating data, rebuilding workflows, retraining colleagues, and discovering unfamiliar failure modes.
A convincing demonstration shows a possibility under selected conditions. It does not necessarily establish that adoption will reduce the organization’s total work.
This is not evidence that AI deployments generally fail, or that new platforms are not worth adopting. It is a reminder that what looks like a technical improvement can create work elsewhere in the organization. Someone must manage the migration, reconcile the data, retrain the team, review the outputs, and decide what to do when a workflow does not behave as expected.
A company can count the hours no longer spent writing a report while overlooking the time now spent interpreting, checking, and defending it. It can measure a greater volume of completed tasks without noticing that each additional task produces another decision someone must own.
The work has changed form. Whether the total burden is lower depends on the entire arrangement, not on the speed of the first step.
Imagine a sales team that can draft five proposals in the time it once spent producing one. That looks like an enormous productivity gain until those proposals reach pricing, legal review, and the account teams responsible for making the promises credible. If the later stages still operate at their previous pace, the organization has increased the volume entering the queue, not necessarily its capacity to make sound commitments.
The bottleneck has moved. AI could help with those other steps too, but that takes deliberate workflow design and an agreement about responsibility, not simply another prompt. Counting faster drafting while ignoring the queue would mistake a local improvement for a system-wide one.
A Human In The Loop Still Needs Authority
This leads to a problem with one of the most reassuring phrases in AI deployment: human in the loop.
It sounds as though responsibility has been preserved because a person still reviews or approves the output. But what exactly can that person do? Do they have the source material needed to verify the answer? Enough time to inspect it? A practical way to stop or correct the process? And will the organization support them when they disagree with a recommendation that the system has produced confidently?
These questions extend beyond generative AI. An OECD survey of more than 6,000 mid-level managers in six countries examined algorithmic management tools used to instruct, monitor, and evaluate work. Such software is not necessarily AI-powered, and the findings should not be presented as a study of generative AI alone. Nevertheless, the organizational concerns are relevant: many managers reported benefits from better information and faster decisions, while nearly two-thirds of those using algorithmic management tools expressed concerns about effects on workers. Unclear accountability for wrong decisions was the most frequently reported issue.
There is a difference between assigning someone responsibility and giving them the practical conditions to exercise it. A manager asked to approve a stream of AI-generated material may be formally accountable for every decision while lacking the attention, evidence, or discretion required to review it properly.
That does not make human oversight futile. In consequential settings, slowing down, escalating an exception, or rejecting a recommendation may be exactly what good oversight looks like.
The point is that oversight has a cost and a purpose. Treating the human as a final approval button can obscure both.
The Workplace That Emerges Afterward
I have written before about how new technology can expand our capabilities without necessarily making our lives more coherent. This is one place where that distinction becomes practical.
A company may adopt AI to reduce repetitive work, then use the resulting capacity to improve service, make decisions more carefully, or give employees greater control over their time. Another may use the same technology to raise output targets while leaving review capacity, decision rights, and recovery opportunities unchanged. Both can report productivity gains. But they are building different working environments.
Neither outcome can be read from the interface. A polished demonstration tells us what the system does under selected conditions. The daily workflow tells us how people are expected to operate around it.
As organizations integrate AI into more tasks, they will make decisions about pace, responsibility, and authority whether they acknowledge those decisions or not. Every additional output has to be read or acted upon by someone, somewhere. Every error has to be noticed by somebody or allowed to travel. Every saved hour can be protected, invested, or quietly filled.
This is why I am interested in what happens after the demo. I want to see the cases where technology actually reduces unnecessary demands, as well as the ones where it moves those demands somewhere less visible. More importantly, I want to understand who had the power to choose the difference.
There will be plenty of impressive demonstrations in Lisbon this week. I expect to appreciate many of them. But the question I will carry into those conversations is one that the demonstration itself cannot answer:
When this becomes the ordinary way of working, what changes for the people who have to live with it?
Notes and Sources
- Jones, J. M. (2026). “AI Benefits at Work Unevenly Distributed.” Gallup / American Job Quality Study, October 6. U.S. survey; AI effect figures are self-reported. The technology-adoption influence statistic refers to the full employee sample.
- Dillon, E. W., Jaffe, S., Immorlica, N., and Stanton, C. T. (2025). “Shifting Work Patterns with Generative AI.” NBER Working Paper 33795, revised November 2025. Randomized field experiment across 66 firms and 7,137 knowledge workers.
- OECD (2025). “How widespread is algorithmic management in workplaces?” OECD Policy Brief, December 19. Survey of more than 6,000 managers in six countries; algorithmic management is not limited to generative AI.
- Oloyede, M. (2026). “I Reviewed 4 New AI-Led Marketing Automation Platforms. Would I Actually Switch?” Marketing Accountability Council, October 28. Practitioner evaluation of vendor demonstrations, not independent production testing.
- Web Summit (2026). “2026 program.” November 9–12, Lisbon; November 10 session “The orchestration gap: Fixing the enterprise’s AI problem.”