An AI needs fewer than 10 pages to write in your company's voice. My own writing style fits on about three. When a technology team in Dubai asked how much of its archive the AI would have to learn from, I put their voice at under 10 pages. The harder question is what stays out.
The question came up during an AI audit with a company whose commercial team works out of Dubai, the same city I run my business from. They write white papers, blog posts, LinkedIn posts and editorials for their leadership, and their first idea was the one most teams have. Upload everything into one big company brain and let the AI learn from all of it. That idea mixes up three separate jobs: teaching the AI a voice, giving it facts when a task needs them, and protecting the knowledge that should never leave the company. Each of the three wants a different answer, and getting them apart is most of the work.
A whole writing style fits on three pages
I know the size because I built my own. I did not want my AI to sound like AI, so I built a writing skill that makes it speak the way I speak. It is a Markdown file, and converted to Word it comes to maybe three pages. The result, as I put it to the team on that call, sounds scarily like me, and nothing in it carries an em dash.
That is the point most teams miss when they plan a big upload. As I told them, "The AI needs less text than you think to know what you want." A voice is a set of habits: the words you reach for, how long your sentences run, the phrases you would never use. Habits fit on a few pages. Feeding the AI a thousand pages of old content does not make those habits clearer, it buries them under everything else you once wrote.
Under 10 pages of voice for a whole company
The team in Dubai had a full content library, with white papers as the source for almost everything else they publish. They assumed that matching their voice meant giving the AI all of it. My estimate for them was less than 10 pages, and I said I could build that with full confidence.
The form matters as much as the size. When someone on the call asked whether there was anything to build at all, or whether they could simply upload the files, my answer was short: "It's a skill. It's not an automation." A skill is a set of instructions the AI loads when it writes for you. I offered one for LinkedIn, one for the website, one for white papers and one for press pitches, all built on the same structure and shared across the team. Nothing runs on a timer in the background. The voice comes along whenever someone on the team sits down to write.
The archive belongs in a skill that loads it when needed
Voice is one job. Facts are another, and here the urge to upload everything costs the most. The white papers were the company's core material, so the obvious plan was to make the AI know them at all times.
How much context you give the AI is one of the most important things in working with it, so I asked the question that settles this: "And does it need to know the white paper with every chat? Most probably not." The white papers therefore get their own skill. You start it with something like slash white paper, it goes to the folder by itself, reads every white paper and comes back ready to write. The rest of the time they stay out of the way.
I asked how many there were, and I had guessed 300. They had about 10, each around 20 pages long. That is easily possible, and it is a very different project from the one they had pictured.
Keep the last 5 percent of real IP outside the system
The third job is protection, and the question came from a company in a serious technical field. Some of what they know is published, some of it sits behind a download form, and a small part is knowledge they would not want anyone else to have. My advice was direct: "if it's really crazy IP I would not put it in."
You can draw that line inside the company, with one side that works with AI and one side that stays manual. I expect the manual side to be small, and I put it at around 5 percent of what they know. That 5 percent stays outside the system, with the people who own it.
When they asked whether Claude's own policy covers this, I made clear that it was my position and nothing more. "This is me saying." A rule you set yourself is one you can actually keep.
Fewer than 10 pages, and a short list of what stays out
So how many pages does an AI need to write in your company's voice? Fewer than 10, and in my own case about three. The voice goes into one short skill, the reference material goes into skills that load it only when a task needs it, and the small share of real IP never goes in at all.
What remains is a set of decisions, and they take less time than a big upload. That matters for a team like the one in Dubai, which produces a lot and has more important things to do than maintain an AI system by hand. The goal I set for them was a setup that keeps getting smarter without the team doing much at all.
One next step: before you upload anything, write down which 5 percent of your knowledge stays outside, and only then collect the three pages that describe how you write.