What an AI Document Generator Actually Saves: A Proof of Concept with Real Numbers

Analysis stage: Can an AI tool understand what the user needs?

Before AI could write anything, we had to ask an uncomfortable question: could we trust it to understand what a user needs from a guide?

We started the same way we normally would. We fed AI the raw materials we’d use ourselves: functional specifications, user flows, early mockups, everything we would typically analyse before writing. But we knew that wouldn’t be enough. The way we see it, user guides are a filtered version of product reality. They have to be built around the user’s goals and struggles: looking at the product the way a slightly curious, slightly skeptical user would, with one eyebrow raised.

So, we had to push AI to look at the material through that lens:

• Strip away anything that didn’t directly serve the user and belonged in an internal knowledge base instead.
• Flag what a user would reasonably want to know that wasn’t explicitly spelt out in the source material.

That second point is especially tricky. Anyone who has written documentation knows that not all knowledge is documented. A surprising amount lives in stakeholders’ heads: real-world workflow nuances and things that only actually surface in conversation. A guide that ignores that layer may look nice at a glance but turn out to be not very helpful under closer scrutiny.

Asking AI to think critically didn’t absolve us from doing the same. Our next step was to fact-check the AI results against source materials—first using carefully engineered prompts, and then manually—to make sure the guide wasn’t quietly inventing features that weren’t actually there.

And true to our thorough nature, we ran this stage across two different scenarios:

• High-level requirement descriptions for a mobile sales app, outlining only core tasks.
• A highly detailed specification for an internal online CV form.

Result: After testing AI from every angle we could think of, we saw that this stage, analysing inputs, sorting relevant from irrelevant, and identifying blind spots, could be completed two to three times faster with AI. In our experience, this phase typically accounts for 30–40% of documentation effort, which means that using an AI document generator saved 12–18 hours in every 100 hours of documentation work.

Structure stage: Can AI organise information the way a human would?

With a clear picture of what to include in the guide, the next step was structure. Even after we (repeatedly) emphasised user focus in Stage 1, left to its own devices, AI largely mirrored the structure of the source documents. Sections appeared in roughly the same order. Product-centred topics crept back in. Some sections were tiny (one question, one answer). Others tried to cover too much in one breath.

So we had to push AI to:

• Prioritise user workflows over feature descriptions.
• Start with sections about concise concept overviews tied to real user goals. Then add sub-sections about clear, step-by-step procedures that would help achieve a particular goal.
• Avoid overlapping or repetitive sections.
• It took several rounds of revision, merging sections, tightening hierarchy, and removing duplication, before the outline felt genuinely intuitive and logically consistent.

Result: In the end, an AI document generator helped us create a solid structure approximately 1.5–2 times faster. The thing is, outlining isn’t the biggest time sink in documentation: in our experience, it accounts for no more than about 5% of total effort. So while using AI to create an outline was a logical shortcut (which roughly saved us 3–8 hours), the real test was still ahead: writing.

Writing stage: AI document creation with disciplined minimalism

In technical writing, the actual process of writing is one of the most tightly controlled parts of the job.

There are literally volumes of best practices for writing short texts that still include context, steps, results, prerequisites, notes, and edge cases, and for ensuring the user understands everything they need to know without overexplaining what they probably already do. And that’s before we even get to visual formatting rules.

AI was very capable of producing full first drafts, but it used way more words than necessary.

In our minds, writing is like carving a statue. You start with a block of information and meticulously chip away at everything that doesn’t serve the user. Teaching AI to respect that level of stubborn minimalism meant reinforcing a long list of rules and constraints, ensuring none of them was forgotten along the way.

Writing a section with AI: about the corporate CV form

There was another practical limitation. Early mockups and requirement documents rarely capture every detailed step in the product, which meant that some procedural details could only be added once the development was underway.

Result: AI helped us complete the writing part approximately 1.2–1.7 times faster, which translates into roughly 7–18 hours saved in every 100 hours of documentation work. That may not sound dramatic at first−until you consider that writing typically accounts for 40–50% of total documentation effort, so the overall return on effort adds up quickly.

The overall results of AI-generated documents

So, how much effort did all of this AI experimentation save us? Certainly not the mythical 70% some headlines might promise.

With analysis, outlining, and writing taken together, an AI document generator allowed us to work approximately 1.25–1.3 times faster overall. Or, put simply, to save about 20–25 hours for every 100 hours of documentation effort.

To illustrate this in organisational terms, consider 3 product teams, each with a dedicated full-time technical writer delivering a major release every 2 months. Over an 8-week cycle, each writer contributes 320 working hours.

Applying a 20–25% efficiency gain translates into 64–80 hours saved per writer per release cycle. Across three teams, that amounts to 192–240 hours saved every 2 weeks.

Over 6 release cycles per year, that becomes 1,152–1,440 hours of additional documentation capacity annually — roughly 7–9 months of combined writing time, without increasing headcount.

For teams operating under continuous delivery, this directly impacts backlog management, release alignment, and the overall quality of documentation.

Where human judgment remains non-negotiable

Knowing where AI productivity gains end is just as important as knowing where they start. During the proof of concept, four areas always needed human involvement:

• Stakeholder interviews, since AI cannot pull unwritten context out of someone’s head (yet).
• Product-truth verification under the responsible AI principles.
• Screenshots and formatting adjustments.
• Judgment calls in ambiguous scenarios (because you just don’t outsource that decision to a model).

AI accelerated the heavy lifting (analysis and writing), but it did not replace our cognitive muscles.

The AI + Information Developer combo is especially powerful if you have:

• Tight launch timelines.
• Lean budgets.
• Frequent product updates.
• Large documentation sets.

In those environments, especially, a 20–25% save probably means saving your Information Developers from burnout and keeping even the most scrupulous users happy.

Conclusions

A better question for enterprise leaders isn’t if AI saves documentation time — it does — but what those time savings mean for your team setup, tool investments, and quality control.

Our proof of concept shows three key takeaways:

• AI-assisted documentation isn’t about replacing junior staff; it enhances the work of experienced technical writers, who can guide and review the AI’s output. Organisations that think otherwise risk building up quality problems faster than they save time.
• The balance between documentation and product will change. As AI speeds up writing, the bottleneck moves to human tasks like stakeholder interviews, fact-checking, and editing. Documentation teams should focus on these skills instead of just producing more content.
• With faster release cycles under continuous delivery, documentation tools should be chosen not only for speed but also for how well they fit into the development process, using the same source files, updating step-by-step, and spotting gaps before users do.

The 20–25% productivity gain shown here is just the starting point. The companies that will do best are those that treat AI-assisted documentation like an engineering process, using clear prompts, controlled inputs, careful human review, and ongoing improvements to the workflow.