Dima Bolkunov measured something most productivity arguments skip: he reads AI-generated text at least a third slower than human-written text of the same length. Add the verbosity most models default to, and total reading time climbs further. In a manufacturing or quality organisation, that cost multiplies. A deviation report, a work instruction, a CAPA summary gets read by twelve people, sometimes forty. The ten minutes your engineer saved drafting it becomes two hours spread across the people who have to interpret it, plus the follow-up questions and the review cycles.

So the question stops being “does AI writing save time” and becomes “whose time, and how much of it”. Below, where AI writing genuinely pays off in quality and operations documentation, where it quietly costs you, and how to tell the difference before you roll it out.

The Time You Save Writing Is Borrowed From Everyone Who Reads It

Bolkunov puts the accounting plainly:

You do not need many readers for the writer’s time savings to evaporate on the readers’ side.

That arithmetic is brutal in a quality organisation. One author saves twenty minutes on an SOP revision. Thirty readers each spend an extra ten decoding it. You have converted half an hour of drafting into five hours of interpretation, and you cannot see it on any dashboard because reading time is never logged anywhere.

So stop arguing about whether the output is good enough. That framing assumes the cost sits with the writer, where it is visible and easy to defend. It does not. Every document with a wide distribution list moves the cost downstream to auditors, line supervisors, and the engineers who have to act on it.

What Actually Disappears: Style, Context, and the Author Behind the Text

Bolkunov’s bachelor’s thesis set out to identify an author from a short paragraph using word-usage statistics alone. It worked. And human readers go further than statistics: they read sentence structure, the order someone chooses to tell things in, which detail gets the most words.

That signal is the first casualty. As he puts it, AI averages style, so it’s hard to see the author behind it. The text arrives clean and anonymous, and the reader loses a layer of information they were using whether they noticed it or not.

Why losing the author makes a document harder to trust

When a plant manager reads a supplier’s explanation of a non-standard deviation, the writing tells them things the writing does not say. Whether the engineer is confident or hedging. Which failure mode they actually believe in. How much of the investigation they did themselves.

Strip that out and the manager has only the claims, with no way to weight them. Bolkunov’s point lands hard here: with AI, context is being lost. Worse, when you know the writer and the text does not sound like her, you stop reading for meaning and start reading for what is missing.

Where style is actually load-bearing in operational documents

Style carries weight wherever judgement is being transmitted, not just facts. Root cause narratives, audit responses, supplier corrective action plans, escalation emails, risk assessments. In all of these, the reader is partly assessing the author.

Style matters far less in documents where the content is already fixed. Parameter tables, calibration logs, structured test records, bill-of-material changes. Nobody reads a torque spec for the voice behind it. That distinction is where your policy should start.

Why AI Text Loses Precision Even With a Detailed Prompt

Bolkunov’s second mechanism is mechanical, not aesthetic. A model stitching a document together has to fill gaps your prompt never covered, and the safest way to fill a gap is to hedge. Ambiguous wording avoids logical mistakes. It also avoids saying anything a technician can act on.

The ambiguity blind spot: why authors cannot see it in their own AI drafts

When you read back a draft, the exact meaning is already sitting in your head. It attaches itself to the vague word and the sentence reads fine. Bolkunov names this directly: the writer has a clear idea in their mind, which gives an ambiguous word an exact meaning.

He adds a second loss. Writing something out is how you find the holes in your own reasoning. Jump straight to a generated result and you skip that check entirely, so the gaps survive into the document and get discovered by someone else, later, usually at the worst moment.

How imprecise wording becomes a downstream quality cost

“Ensure adequate cleaning between batches” is not a style problem. It is a defect source. Acceptance criteria, CAPA effectiveness checks, and work instructions only function when every reader converges on the same interpretation, and hedged phrasing guarantees they will not.

Then the second-order effects land. More clarifying questions back to the author. More iteration cycles on anything written collaboratively. And the one that compounds quietly: less trust between readers and the writer, because people learn that your documents need decoding.

Run the test before release. Hand the draft to someone with no context and ask them to describe the required action. If they hesitate, the precision is gone.

Writing as Thinking: The Step AI Lets You Skip

Bolkunov’s third mechanism is the one operations leaders should sit with longest. Writing forces you to confront your own reasoning in a way that thinking never does. Gaps show up on the page. They stay invisible in your head.

Putting things on paper usually helps us see inconsistencies and gaps in our judgement, which are easy to miss when we are just thinking about them. AI writing takes away this opportunity, as you jump directly to a result.

That jump is the whole problem. A business case for a new line, a root-cause narrative, a three-year transformation plan: in each of these, the struggle to make the argument hold together is the work. The document is a by-product. Hand the drafting to a model and you get the by-product without the work, and nobody, including you, can tell the difference until the plan meets reality.

Which documents exist to produce clarity, not just information

Sort your written output into two buckets. Some documents transmit facts that are already settled: shift handovers, inspection logs, status updates, meeting notes. Nothing is being figured out, so drafting friction adds no value, and AI tools genuinely help.

The other bucket contains documents where the reasoning is not finished when you start typing. 8D reports. Capital requests. Risk assessments. Anything where you are arguing from evidence to a conclusion that could reasonably have gone another way.

Draft those yourself, badly if necessary, then use AI to tighten structure or catch gaps. The order matters. Thinking first, then tooling, keeps the confrontation with your own logic intact. Reverse it and you have a polished document defending a position you never actually tested.

Where AI Writing Earns Its Keep in a Manufacturing Organisation

A blanket ban is lazy policy and your teams will ignore it anyway. The useful split is by readership and consequence, not by whether the task feels like writing. Internal meeting summaries, translating a shift handover into English for a sister plant, reformatting machine data into readable prose, restructuring notes you already took: one or two readers, low stakes, nothing that outlives the week. That is where drafting with AI tools costs almost nothing.

SOPs, CAPA narratives, audit responses, customer commitments, supplier corrective action requests. Dozens of readers, long shelf life, and every ambiguous phrase becomes someone else’s interpretation problem months later.

A reader-count test for deciding when to use AI drafting

Before you draft, do the arithmetic. Estimate reading time, multiply by the number of people who will actually read it, and compare that against the drafting time you expect to save. Bolkunov’s own benchmark gives you the multiplier: he reads AI-generated text at least a third slower. If the document goes to more than a handful of people, the maths stops working.

Apply a second filter on top. If a sentence would need to survive a regulator’s question or a customer dispute, a human writes it. No exceptions, no “I’ll clean it up after.”

Using AI to interrogate drafts instead of produce them

The stronger use is on the other side of the draft. Write it yourself, then ask the model where a technician could read a sentence two ways, which claims have no evidence behind them, what a notified body would challenge first.

You keep the thinking and the style. You get a cheap adversarial reader who never gets tired of reading your CAPA.

The 2026 Document Standard: Human Precision, AI Leverage Elsewhere

Volume stopped being a competitive advantage the moment everyone could generate it. What separates organisations now is whether a document still carries an author and a decision. A reader should be able to tell who wrote it, what they concluded, and what they want done. If none of that survives the draft, you have produced paper, not communication.

Bolkunov draws a usable line in his own disclaimer: AI writing means “text with at least 50% of the words generated by AI.” That threshold is worth borrowing, because policy needs a definition people can apply without a debate about tools.

A three-tier document policy you can write this quarter

Classify every recurring document type by reader count and consequence, then assign a rule to each tier.

• Tier 1, human-authored: SOPs, CAPA narratives, audit responses, validation rationale, anything a regulator or customer reads. AI may check grammar and structure. It does not write the sentences.
• Tier 2, AI-drafted, human-signed: internal reports, status updates, training material. One named person rewrites the conclusion in their own words and puts their name on it.
• Tier 3, unrestricted: notes, translations, reformatting, summaries for one or two readers.

Then take the harder step. Walk your document inventory and ask which items exist because someone needed the information and which exist because a procedure demands a file. Most quality systems carry both. The second category is where AI pays off properly: structured fields, dashboards, and automated status pulls that replace prose nobody was reading carefully anyway.

The best contribution AI makes to a quality organisation is deleting documents, not drafting them. Fewer things to read, written by people who thought it through.

Source: bolkunov.com