AI can turn one unresolved customer problem into several polished answers, complete with research and working prototypes. A product team can produce these options quickly, and each one looks convincing enough to pursue.
Producing in days what once took a month feels like progress.
Then the review meeting begins.
The prototypes answer different versions of the problem. The research supports several incompatible conclusions. Nobody agreed on the evidence required to choose among them. One option needs approval from security, another changes the support workflow, and a third has no obvious owner after launch. The meeting ends with a request for one more comparison.
Nothing is wrong with the output. There is simply much more of it than the organization knows how to finish.
This pattern is showing up across knowledge work. AI reduces the cost of producing a plausible first version: a draft, analysis, design, recommendation, or prototype that gives people something to react to. But businesses do not create value by accumulating plausible beginnings. They create value by deciding which work deserves commitment, carrying it through the constraints that make it real, and stopping the rest.
AI is making it cheap to start work. The advantage is moving to organizations that know how to finish it.
Cheap production creates expensive unfinished work
An AI-assisted start can feel close to free. An idea becomes a research brief, several designs, and a prototype before anyone has made a difficult decision.
The costs appear later.
Every new initiative must be compared with existing priorities. Someone has to check the evidence, resolve dependencies, and decide what happens next. A prototype may need security review; a campaign may need legal approval and production capacity. Whether the work reaches customers or not, the organization pays for the attention it absorbs.
A generated document is unfinished inventory. Until it changes a decision or reaches a completed outcome, it competes for finite organizational attention.
Research on AI at work already shows part of this divide. In a six-month randomized field experiment across 6,000 knowledge workers, access to generative AI changed activities people could alter independently: users spent less time on email and appeared to complete documents faster. It did not significantly change time spent in meetings, where progress depended on other people changing together. The tool accelerated individual production more readily than organizational coordination. Microsoft Research
The individual gains are real, but they do not automatically travel through the rest of the business.
When starts become cheaper, organizations permit more of them. Reviewers face more material, leaders divide their attention across more bets, and teams inherit maintenance work from experiments that were easy to approve.
Low production cost can hide high opportunity cost. Alongside “How quickly can we make this?” teams need to ask, “What will this displace if we decide to keep it alive?”
A prototype no longer earns the benefit of the doubt
A working prototype used to imply that somebody had secured a budget and invested serious time. That did not prove the product was good, but the expense acted as a filter. Few ideas reached that stage accidentally.
AI weakens the filter.
A small team can now produce something persuasive before answering basic questions about the problem. The interface works, the sample data looks credible, and the demonstration resembles a finished product.
That visual completeness changes the conversation. Stakeholders stop asking whether the problem deserves investment and start discussing colors, features, and launch dates. The existence of the prototype quietly becomes evidence for continuing it.
But a prototype proves only that this version can be made. It says nothing about whether the problem matters, customers will use it, or the organization can operate it safely.
Cheap prototypes reverse the burden of proof. Building can happen before commitment, so the artifact itself can no longer justify commitment. The team must earn that separately.
A useful prototype should be attached to a decision. It might test whether support agents understand a recommendation or whether customers complete a new onboarding flow. If nobody can say what the prototype should make easier to decide, it is a demonstration looking for a reason to continue.
Speed makes unclear direction more dangerous
Routine approvals and duplicate data entry rarely improve the work. AI and automation should remove that kind of delay.
Other friction was carrying information.
Writing a brief forced somebody to explain the problem. Turning a request into a specification exposed missing decisions. Slow production also gave different functions time to notice when they were using the same words for different goals.
AI can produce the artifact without requiring those conversations.
Ask for a customer-retention plan and AI will create one even if finance and customer success disagree about which customers matter. Ask for a product prototype and it will fill gaps with reasonable assumptions that nobody consciously chose.
The result can be polished enough to make the underlying ambiguity difficult to see. A vague direction no longer looks vague after it has been turned into headings, screens, calculations, and confident prose.
Clarification needs to happen earlier. Before generation begins, name the decision, the intended user, the constraints, and the disagreement the artifact must help resolve. Otherwise, faster execution lets a weak premise travel further before it meets resistance.
AI does not eliminate the need for clarity. It eliminates many of the moments when a lack of clarity used to reveal itself.
Selection is a capability, not the final step of generation
Most creative and analytical workflows treat selection as the short final phase. Produce several candidates, review them, choose the best one, and move on. When producing the candidates took most of the time, this division felt reasonable.
Generation is getting cheaper. Selection is not.
Choosing among plausible options requires a different kind of work. Someone must decide which evidence counts, notice when two alternatives solve different problems, recognize the exception that changes the recommendation, and distinguish a reversible experiment from a commitment the organization will have to support for years.
More intelligence does not settle these questions automatically. A model can compare a growth strategy with a risk-reduction strategy. It cannot decide which outcome the business values more unless the business gives that tradeoff an owner. It can forecast the likely effect of a policy. It cannot accept the consequences on the company’s behalf.
A field experiment involving 791 professionals at Procter & Gamble offers a useful distinction. AI improved the quality of generated ideas and helped individuals perform at a level comparable to two-person teams on product-innovation work. But when the researchers separated generation from evaluation, human judgment retained value in selecting among the ideas. Organization Science
That finding points toward a change in expertise, not its disappearance. An experienced employee may spend less time creating the baseline answer and more time recognizing whether an answer is acceptable. They know which source is missing, which exception is not routine, which promise the organization cannot make, and which apparently small detail will become expensive in production.
Companies should not treat this as generic review work that can be added after generation. Selection needs its own criteria, time, authority, and skill development. If the organization multiplies output without strengthening its ability to reject, generated abundance becomes a queue for its most experienced people.
Once production is cheap, the quality of selection matters more.
A stopping rule belongs at the beginning
Generative systems are exceptionally good at continuation. There is always another query to run, another source to summarize, another version to produce, another edge case to accommodate, and another prototype direction that might be slightly better.
Generative systems will happily continue until someone tells them to stop. Organizations need to set that limit.
A stopping rule defines the conditions under which exploration will continue or end. It answers four practical questions: what work is being tested, what evidence is sufficient, who owns the decision, and when the evidence must arrive.
The order matters. If the rule is invented after the prototype looks impressive, the criteria will bend around the work already produced. If it is invented after a team has spent weeks researching, stopping will feel like wasting the investment. A rule written before generation begins protects the decision from the persuasiveness of its own output.
Time must be part of the rule too. Analyses become stale. Market windows close. Customer needs change. A possible initiative should not remain in the backlog indefinitely simply because generating one more revision is cheap. Some work should expire.
The rule should still allow for learning. New evidence may justify an extension, but the extension should be a decision rather than the default state of unfinished work.
Build a stopping rule
Give unfinished work a condition for continuing and a date for stopping.
What are you currently exploring?
Your rule
We will continue [name the work] only if [name the evidence required] by [set a date]. [name one owner] will decide whether it continues.
A stopping rule cannot guarantee a good decision, but it makes a decision possible.
Automation leaves people with a harder remainder
Moving from generation to automation changes the finishing problem rather than solving it.
Imagine a service team handling 100 cases. Most are routine. Some require interpretation. A small number are unusual, consequential, or emotionally difficult. Automating the routine cases reduces the number that reaches people, which is usually a real improvement.
It also changes what a normal human case looks like.
The employee no longer moves between easy and difficult work. Their queue contains a higher concentration of exceptions: missing documents, conflicting policies, unusual customer circumstances, cases the system has already failed to resolve. Total volume falls while average difficulty rises.
OECD case studies across workplaces have documented this pattern. In an Austrian insurance operation, automation removed much of the manual indexing work and shifted employees toward intermediate and complex cases. In a UK financial-services example, a chatbot absorbed basic queries while human agents handled a greater share of difficult issues. The work did not simply shrink. Its case mix changed. OECD
That matters for business cases built around volume. If an automated system removes half the queue, it does not follow that the organization needs half the human capacity. The remaining work may take longer, require more experienced staff, create more stress, and carry more risk when handled poorly.
It matters for expertise too. If people see only exceptions, they have fewer routine cases through which to learn the domain. The organization may depend more heavily on expert judgment while weakening the path by which new experts used to develop it.
The right measure is whether the combined system (automation plus people) can close the full distribution of work reliably, not merely how many cases the system completes.
The automation remainder
Watch the human queue change as the easiest cases leave it first.
Illustrative workload—not an industry benchmark
Cases reaching people
100 of 100
Handled automatically
0 of 100
No cases have left the human queue yet.
Cases reaching people
100
Total human time
30h 30m
Average per human case
18.3 min
Complex-case share
40%
The human queue still contains the full mix of work.
Automation can still be worth doing. The point is to finish the reasoning as carefully as the system finishes the routine cases.
Finish fewer things accidentally
The easy response to generated abundance is stricter control over who may use the tools or how many experiments a team may run. That may reduce visible activity. It does not build the capability the organization actually needs.
The better response is to make completion explicit.
Before work starts, name the decision. Before a prototype earns investment, state what it proves and what it does not. Before generating alternatives, decide who will select among them. Before automating routine work, examine the cases that will remain. Give unfinished work a condition for continuing and a date after which it expires.
These practices may sound slower than asking AI to produce the next version. They are not. They prevent the organization from spending weeks reviewing, coordinating, and maintaining work that was never capable of reaching a decision.
The abundance is real. Teams can explore more possibilities, communicate ideas more clearly, and test assumptions with a speed that would have been impractical only a few years ago. The opportunity is not to return to a world where every beginning was expensive.
It is to stop confusing the cost of beginning with the value of finishing.
The next time an impressive draft, analysis, or prototype appears, do not ask only how quickly it was made. Ask what decision it is meant to close, what evidence would allow that decision, who owns the consequence, and when the work will stop if the evidence never arrives.
Starting will keep getting cheaper. Finishing is where the advantage will remain.