The AI Finished the Work. Now Prove It Did Not Break Anything.

Scope note: This essay considers autonomous knowledge work, long document editing, product-management delegation, spreadsheet agents, plan review, and expert validation as one verification problem. It does not estimate economy-wide job effects or claim that every agent and task creates the same review burden.

The AI says the work is done.

The document is polished. The spreadsheet calculates. The research packet has citations. The plan contains six orderly steps and the final box is checked.

Now comes the expensive question: what changed that should not have changed?

Recent research on autonomous knowledge work contains both sides of the argument. Agents can compress hours into minutes and let people attempt larger tasks. They can also introduce small, severe errors that survive because the output looks complete. The human labor does not disappear. It moves from making the artifact to proving the artifact deserves to exist.

Generation became cheap. Verification is becoming the job.

Long delegation accumulates quiet damage

LLMs Corrupt Your Documents When You Delegate, by Philippe Laban, Tobias Schnabel, and Jennifer Neville, tests 19 language models across long editing workflows in 52 professional domains.

The benchmark does not ask a model to answer one question. It asks the model to keep working inside a document while following a sequence of instructions. By the end of the workflows, even the tested frontier models had corrupted an average of roughly 25 percent of the document content.

The damage grew with larger documents, longer interactions, and distracting files. Giving the model agent tools did not solve the problem.

The errors were sparse enough to hide and severe enough to matter. That combination is worse than obvious failure. A broken document invites inspection. A mostly correct document asks for trust.

Long-running agents create a cumulative risk. Every accepted change becomes the starting state for the next one. A small wrong edit can survive several correct edits and emerge inside a finished artifact with no visible alarm.

Autonomy can still create enormous value

The negative result should not flatten the rest of the evidence.

In How AI Agents Reshape Knowledge Work, Jeremy Yang and his coauthors analyze production data from Perplexity’s Search and Computer products. For closely matched tasks, the agentic product completed work in 36 minutes compared with an estimated 269 minutes for people using search. The paper reports 55 percent lower per-query dissatisfaction, along with large estimated reductions in time and cost.

The agent also changed what people attempted. Tasks crossed more occupational boundaries and combined more connected steps. Follow-up work shifted toward verification and extension.

The data comes from the company operating the products, and the comparison relies on matched sessions and estimated human work rather than a randomized workplace trial. It should not be treated as a neutral measurement of every agent.

It still supplies the necessary counterweight. Autonomy is not merely a new source of errors. It can expand the amount and scope of work a person can direct.

The real question is whether oversight expands with it.

Professionals delegate tasks, not accountability

Product Manager Practices for Delegating Work to Generative AI, by Mara Ulloa and her coauthors, studies 885 Microsoft product managers, examines telemetry for 731 of them, and interviews 15.

The title carries the central lesson: accountability must not be delegated to a non-human actor.

Product managers choose work partly by whether they can evaluate the result. Drafting, synthesis, and structured preparation may be easier to delegate than decisions whose quality depends on hidden context or whose consequences belong to the employee.

That distinction is more useful than a list of “AI use cases.” A task is not ready for delegation because the model can produce an output. It is ready when a qualified person can inspect the evidence, identify failure, and remain responsible for the decision.

Delegation begins with the acceptance test.

Review at the end is too late

Many interfaces let an agent run and then present the finished result. That design places the human at the end of a long chain of invisible decisions.

Auditing and Controlling AI Agent Actions in Spreadsheets, by Sadra Sabouri and his coauthors, tests a different approach. Their Pista system breaks spreadsheet work into visible actions that a person can inspect and redirect while execution is still happening.

A formative study with eight participants and a comparison study with sixteen found that active participation changed more than the final result. People understood the task differently, saw their own intent in the agent’s actions, detected errors that later review would have missed, and reported more ownership of the output.

The samples are small. The finding is still sharp: oversight works differently before the artifact hardens.

Post-hoc review asks a person to reverse-engineer the route from the final cells. Active review lets the person stop the wrong turn while the reason remains visible.

The plan also needs inspection

Some systems try to solve the oversight problem by showing the agent’s plan before execution.

A fluent plan can produce the same overtrust as a fluent answer.

In An Experimental Comparison of Cognitive Forcing Functions for Execution Plans, Ahana Ghosh and her coauthors tested small interventions that require people to examine assumptions or imagine alternative conditions before approving an AI-generated writing plan.

Asking participants to inspect assumptions reduced overreliance without increasing measured cognitive load. Participants described the alternative-condition prompt as especially helpful.

The important detail is that the interface did not merely display more information. It required a specific act of judgment.

Transparency without a decision can become wallpaper. Oversight needs a lever.

Expert control must be part of the system

The Expert Validation Framework, by Lucas Gren and Felix Dobslaw, proposes a broader enterprise method. Domain experts define expected behavior, test the AI components, validate results, and continue monitoring the system after deployment.

The framework is a methodology rather than a comparative trial. It makes the authority boundary explicit. Model builders do not become the final judges of domain correctness merely because they built the system. Domain experts remain responsible for the rules by which the system is accepted.

That principle should reach the artifact level. Every delegated workflow needs:

• a bounded task and named owner;
• an inspectable plan with assumptions;
• checkpoints before irreversible or wide changes;
• source links and a visible change history;
• automated tests that can actually fail;
• a final reviewer qualified to accept the work;
• a record of what the reviewer changed or rejected.

The model may do the middle. Responsibility still has a name.

Verification is now a design problem

My objective read of these papers is not that AI delegation has failed. The performance evidence is too strong for that easy verdict.

The trend is more consequential. As agents take longer tasks, quality depends less on the beauty of the final response and more on the structure around execution. The winning system will not merely generate better. It will make checking cheaper, earlier, and more exact.

That changes how organizations should measure productivity. Minutes saved during generation are not the final number. Add review time, correction time, hidden damage, rejected output, and the cost of reconstructing how the agent reached a decision.

The AI finished the work. Fine.

Now prove the work survived.