Ryan Roslansky described a particularly unhelpful office exchange: a colleague sent an AI-written document, he asked AI to summarize it, and the result added no thinking to the idea. His August 21 LinkedIn post called this a “doom loop.” The document had moved between people and machines without moving the work forward.
That distinction sits underneath Futurism’s October 18 article about enterprise AI spending and low-quality output. An AI bill records the cost of generating something. The time colleagues spend checking, correcting and interpreting it belongs in the calculation too.
For anyone deciding whether an AI workflow deserves another month of funding, the useful question is how much work reaches an acceptable finish. Counting drafts, tokens or active users can miss that finish entirely.
The work after the draft
Imagine a team preparing a weekly customer briefing. These are illustrative numbers, not measurements of a particular company or AI product.
Before introducing AI, an analyst spends 60 minutes writing the briefing and a manager spends 15 minutes reviewing it. The combined effort is 75 minutes. An AI-assisted version takes the analyst 20 minutes, which looks like a 40-minute saving on that person’s calendar.
But suppose the manager now spends 35 minutes checking the sources, another colleague spends 15 minutes resolving an ambiguous recommendation, and the analyst needs 10 more minutes to repair the document. The team has spent 80 minutes altogether. Drafting became faster while completion became slower.
Change the checking requirement and the result changes. If review takes 15 minutes and corrections take five, the same 20-minute draft produces a 40-minute workflow. That would save 35 minutes against the original process, assuming the final briefing meets the same standard.
The example explains why two teams using the same tool could reach different budget decisions. The labor productivity question is what the team produces for its combined effort. The writing step matters, but so do the quality of the inputs, the difficulty of verifying the output and the number of people who inherit it. A polished paragraph can be cheap to produce and expensive to trust.
Microsoft’s own lesson starts with the process
In a September 17 account of Microsoft’s internal AI work , strategy and transformation executive Kathleen Hogan described an early approach built around distributing tools, training employees and encouraging usage. Microsoft learned that access alone did not change how work happened.
Her account also describes individual tasks getting faster without improving the whole workflow. A quicker step can simply send more work into the next queue. Microsoft says redesigning processes, including data access and human review, produced better results in selected internal projects.
Those are Microsoft’s own observations about its operations. They make a practical point for a smaller team: map the handoff before celebrating the faster draft. Who checks the numbers? Who decides whether a recommendation follows from them? Where does an uncertain answer go? Those responsibilities are part of the workflow being funded.
The incident statistic has a specific denominator
EY’s September 15 AI Risk and Governance Survey release reports that 36 percent of respondents said their organizations had experienced an AI incident or failure with a materially negative impact. Its examples include data loss, financial damage, operational disruption and brand damage.
The online survey covered 202 US senior AI decision-makers at publicly traded companies with at least $1 billion in annual revenue. Fieldwork ran from May 28 through June 15, 2026. EY reports a margin of error of plus or minus seven percentage points at the 95 percent confidence level.
That is a self-reported finding from a defined group of large businesses. It is not a measured failure rate for AI-generated documents, and it does not tell us how often a particular assistant makes an error. It does show why a business evaluating AI costs may need to look beyond subscription and usage charges.
A spending limit answers only one question
The Guardian’s October 13 reporting describes employers restricting AI use and investing in training. A restriction can contain spending without showing whether the remaining work has become more useful.
A team can set a lower usage allowance and still spend hours untangling weak output. It can also spend more on a carefully designed workflow that leaves people with less checking to do. To compare those choices, keep the task and acceptance standard stable.
For the hypothetical briefing, that means recording preparation, drafting, review and repair time for comparable briefings. Record which ones passed review, which needed substantial rework and which missed their deadline. Keep the software charge alongside those observations. A shorter draft time earns its place in the calculation, but it cannot stand in for all of it.
The same approach makes a failed pilot useful. If source checking consumes the saving, the next experiment could narrow the assistant’s job to organizing verified notes. If nobody can tell whether a recommendation is sound, the task may need better source material or a clearer human decision point before more generation helps.
Newsroom’s earlier look at AI revenue accounting asks what a headline number actually counts. The workplace version of that question is just as concrete: does the apparent saving include everyone who touches the output?
Roslansky’s exchange offers a good stopping point for an experiment. If a document needs another machine to compress it and still leaves its reader without a useful idea, producing more of it will not solve the problem. Give the workflow a finish line, then count what it takes to get there.
