Written by Julien Ricciarelli-Bonnal
10 November 2026
The Essentials
Artificial intelligence makes it possible to prepare reports, presentations and analyses in minutes. Yet the time saved by the person generating a document can become extra work for whoever must read it, check the evidence or make a decision based on it. Research on individual productivity gains and low-quality AI output suggests that faster production does not automatically improve the performance of an entire organisation. The meaningful measure is the complete workflow, including drafting, verification, corrections and final use. A document produced more quickly is not necessarily a job completed more quickly.
A marketing manager prepares a thirty-slide presentation for senior leadership. With generative artificial intelligence, the outline, market analysis, charts and recommendations take minutes rather than hours. The manager sends the file feeling, quite reasonably, that an enormous amount of drafting time has been saved. But the finance director reviewing it discovers figures without verifiable sources, comparisons drawn from different periods and assumptions that nobody approved. What took one employee twenty minutes to produce may take another two hours to turn into something the business can use.
The situation is illustrative, but the mechanism is increasingly familiar. AI can radically reduce the cost of producing a document. It does not automatically reduce the cost of understanding, validating or acting upon that document. Most business material is not created simply to exist. It travels between people who must interpret it, challenge it, complete it or accept responsibility for a decision based upon it.
That raises an economic question that is often missing from conversations about productivity. When a company celebrates the time saved through AI, is it measuring the time saved by the person generating the output, or the time saved by everyone who has to work with it afterwards?
A finished document for its author may be unfinished work for its reader
The production and use of a business document involve different responsibilities. A person writing a report may be seeking approval for a proposal or explaining an opportunity. The recipient must assess whether the information is reliable, consider alternatives and sometimes make a decision that commits money, people or the organisation’s reputation. Elegant formatting may make the task easier, but it cannot establish whether the underlying reasoning is sound.
The problem emerges when a document has every outward sign of completion without supplying the information needed to trust it. A plausible number is not necessarily an accurate one. A confident recommendation does not prove the alternatives were examined. A tidy executive summary may conceal the uncertainty removed to make the conclusion sound decisive. The more professional the document looks, the easier it becomes to mistake presentation quality for evidential quality.
When content becomes cheap to create, attention becomes more expensive
Mistakes are only part of the issue. Even if each AI-generated document were reasonably accurate, production becoming effortless could still increase the workload of people expected to read it. A project leader who previously received two carefully developed proposals may now receive eight. A board that requested a short decision note may be presented with fifty pages and several alternative slide decks.
This abundance changes professional behaviour. The effort once required to develop a serious proposal naturally limited how many were circulated. As that effort falls, it becomes tempting to send every possibility rather than select the few that deserve attention. The recipient’s capacity to read, compare and make decisions, however, has not expanded at the same rate.
The issue becomes especially visible as AI is integrated directly into everyday workplace software. We have previously examined how embedded AI changes the way documents and decisions move through an organisation. Generating more material is not neutral when every additional report creates an expectation of review, comment or approval. Content that costs almost nothing to produce can still be expensive for someone else to assess.
Individual productivity gains are real, but they are not the whole calculation
It would be misleading to suggest that AI merely moves work from one desk to another. Genuine efficiency improvements have been observed. In a field experiment involving more than 7,000 knowledge workers at 66 firms, researchers led by Eleanor Dillon found that employees using the AI tools studied spent approximately two fewer hours each week on email during the latter part of the experiment. The paper, revised in 2025, did not find an equivalent shift in the overall composition of tasks.
Research by BetterUp Labs and Stanford Social Media Lab provides a different perspective. In a 2025 survey of 1,150 US employees, around four in ten respondents reported receiving what the researchers called ‘workslop’: AI-generated material that looks polished but lacks enough substance to advance the work. Respondents affected by it estimated that resolving an incident cost close to two hours. Those figures reflect survey responses, not a universally measured productivity loss.
The studies examine different populations and different aspects of AI adoption. They cannot be combined to calculate a single overall impact. Together, however, they illustrate why productivity at the level of one employee should not be mistaken for productivity across an entire workflow.
Verification can become an invisible assignment for the recipient
The burden is frequently uneven. The person generating a report knows what they asked the system, which information they provided and which revisions they made. The person receiving the final document may know none of those things. To assess it properly, they can be forced to reconstruct assumptions and calculations the author never explained.
Imagine a consultancy supplying a market assessment produced largely with AI. The report includes competitive trends, growth opportunities and a confident set of recommendations. Yet the client still needs to establish whether the figures apply to its geography, whether the competitors are relevant and whether the proposals respect its operational limitations. Some of the diagnostic work has effectively been handed back to the client, even though the consultant can truthfully say that producing the report took much less time.
Verification did not begin with generative AI. Human-written reports have always required scrutiny. What is changing is the gap between the speed at which apparently finished work can be produced and the time required to establish whether it deserves confidence. The wider that gap becomes, the easier it is to move intellectual effort onto someone else without explicitly asking them to perform it.
The hidden cost also appears in trust and professional relationships
A weak document does not merely consume correction time. It changes the relationship between sender and recipient. After receiving several reports containing avoidable errors, a manager may begin treating everything from the same colleague as needing additional scrutiny. They ask for more evidence, spend longer checking routine claims and eventually decide that doing the work themselves would be simpler.
The answer is not to subject every AI-assisted draft to a burdensome audit. The level of checking should reflect what the output could affect. An informal note about an internal meeting does not carry the same consequences as a contractual proposal, investment analysis or public statement. Verification becomes more efficient when reviewers focus on critical figures, sources, assumptions and commitments instead of treating every sentence as equally dangerous.
Measure the complete workflow, not just the moment a file is sent
A meaningful productivity assessment must follow the work from beginning to end. How long did drafting take? How many minutes were spent checking references, correcting unsupported claims, resolving queries and securing approval? How many revisions were necessary before the output could be used? Most importantly, did the tool support a better decision, or merely create the document preceding it more quickly?
That requires a more demanding definition of finished work. A report is not complete simply because it has a cover, headings and a conclusion. It is ready when it answers an identified question, distinguishes facts from assumptions, supports important claims with traceable information and makes unresolved issues visible. The sender retains responsibility for those standards even when a machine generated most of the wording.
A practical approach to AI Business Governance and clearly assigned review responsibilities starts here. Organisations can agree who prepares, checks and approves different types of output, and who must handle rework. They can require sources for decisive figures, limit unnecessary length, label uncertainty and prevent unreviewed first drafts from being presented as completed professional analysis.
The strongest uses of AI will be those that eliminate a real step, make information more reliable or help an organisation decide more effectively. The technology is capable of all three. But producing more material, more quickly, creates value only when the rest of the organisation does not have to pay for that speed with additional effort.
The question is no longer simply how much time artificial intelligence saves us. It is whether that time has actually disappeared from the organisation’s workload, or merely been transferred to someone else’s calendar.
Written by Julien Ricciarelli-Bonnal
10 November 2026
