Sales copy and materials. All of them state the same numbers.

Product sheets, pages, presentations, emails, posts. Written in different years by different people, and for the same product they give three different weights. This piece produces them from a base of facts approved once, and has them approved before they go out.

This is the page for a single piece of the system. The other pieces, and the criterion for choosing which to start from, are on the services page.

The problem isn't writing, it's that materials contradict each other

A company with ten years of history has dozens of materials in circulation, and nobody knows which one is right. The net weight on the 2021 sheet, the pieces per carton on the 2024 price list and the delivery times in yesterday's email are three numbers that should match, and often don't.

The first to notice is always a buyer, and they notice at the worst moment: when they have already placed an order on one number and receive another. At that point the problem stops being a writing mistake and becomes a trust problem.

A wrong material doesn't show until someone uses it to decide. Then it all shows at once.

What changes in practice

The missing material stops being missing. Requests that stall today because nobody has time to prepare a sheet in English or a page for a new product get unblocked, because the draft is there within the day.

The second change matters more and shows later: all the materials say the same thing. Whoever reads you in three different places finds the same number three times, and doesn't have to ask you which one is right.

The third is that you know where the blockage is. The time from draft to publication is the indicator, and in most businesses that time isn't in the writing, it's in the approval sitting on someone's desk.

Readers read a fifth of it

The first sentence carries the fact, and the reason has been measured. Jakob Nielsen analysed 45,237 page views to calculate how much text is actually read: at most 28% of the words, if a person spent all the time they are on the page reading.

The realistic number is lower, and he writes it word for word: «More realistically, users will read about 20% of the text on the average page». A sales material that keeps the important figure in the third paragraph is hiding it from eighty per cent of those who open it.

The practical consequence is a writing rule, and it applies to every material: the complete answer in the first two or three sentences, with the number in it, and the reasoning below. Whoever reads everything still finds everything; whoever reads three lines leaves knowing the thing they needed.

The base of facts is decided once

It is the piece that makes the rest work, and it is built during analysis. Every item has a value, a date of last check and a person who approved it, so that an old number can be recognised at a glance instead of surviving for years inside a PDF.

What goes into the base	What happens when it changes	What you get
Product numbersweights, formats, cartons	The item is changed in the base, and the system lists every published material that carried the old value.	No price list in circulation with last year's weight, and no wrong order following from it.
Termslead times, minimums, delivery terms	The change reaches whoever has to approve it before it ends up in a material, because it commits the business.	The terms you promise in writing are the ones you can really keep.
Certificationswith the expiry date	The expiry date sits next to the document, and the system flags the materials that cite it before it expires.	No sheet claiming a certification that is no longer valid in front of a buyer who checks it.
The voicehow you speak	The materials already written that sound like you are gathered, and the words you never use stay out.	The drafts come out in your language, and the correction takes two minutes instead of a rewrite.

What the system does, step by step

Every material goes through the same five steps, whatever its form, and each one leaves a log entry saying what happened and who decided it.

Step	What happens	What you get
The requestwhat is needed and for whom	The material, the reader and the one thing it must achieve are declared. A material without a declared reader doesn't start.	No pages written for nobody, which are the ones that then go unused.
The draftfrom your facts	The draft is built on the numbers in the base and on the materials you already have, with the answer in the first sentences.	A text that contains the things you know, instead of sentences that would suit anyone.
Checking the numbersby machine	Every figure in the draft is compared with the base, and those that don't match are highlighted.	The kind of error nobody rereads, a plausible but wrong figure, is stopped before it goes out.
Approvala person	The material goes into the queue of whoever has to approve it, and the system flags what is stuck and for how many days.	The blockage becomes visible, and whoever decides knows where it is instead of finding out at the end of the quarter.
Publicationwith the marking	The approved material goes out, with machine-readable marking where the content was generated by the system.	The transparency obligation is met at source, instead of being fixed afterwards on dozens of files.

The review, which is the part that makes the rest publishable

A model that writes always produces a text, even when it doesn't have the fact. It fills the gap with a plausible number, and that number looks just like a correct one: that is why the automatic check of figures against the base comes before the human reading, not after.

The human reading is there for what a machine can't do, which is to understand whether that sentence can be said to that customer at that moment. It is a commercial decision, and it stays with whoever is answerable for the relationship.

The approval queue is also where the service is measured. The agreed number is the time between ready draft and publication, and it is counted on a real batch before starting, so that the later comparison makes sense.

Generated text is declared, and from August 2026 it is a legal rule

Article 50 of Regulation (EU) 2024/1689 applies from 2 August 2026. For artificially generated content it requires providers to ensure «that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated».

For systems already on the market before that date the marking obligation alone is postponed to 2 December 2026, while the article's other obligations apply straight away. What it really means for a small business, with the scope and the four things to put in order, is on the page about Article 50 of the AI Act for Italian SMEs.

The position here is written down and doesn't change with the rule: the declaration is made anyway, because a buyer who discovers on their own that they have read a generated text stops trusting even what was true. The full reasoning is on the page about the principles we build with.

How many are already doing it, and who is falling behind

It is no longer something for a few. According to Eurostat, in 2025 20.0% of European enterprises with at least ten persons employed used artificial intelligence technologies, against 13.5% in 2024. Generating written or spoken language concerns 8.8% of enterprises, text analysis 11.8%.

The gap that matters isn't between countries, it's between sizes. In the same survey enterprises with more than 250 persons employed are at 55.0% and those between ten and forty-nine at 17%. Thirty-eight points apart, on the same technology and in the same year.

That gap closes on jobs like this one, where the expensive part is the time of people who, in a small business, also do other things. The survey counts enterprises with ten or more persons employed, so the smallest stay out of the count and nobody knows where they stand.

What goes out on its own and what waits for a person

Everything that doesn't commit the business goes out on its own: the draft, checking the figures against the base, flagging the materials to update and the reminder about what has been stuck in the queue for too many days.

Nothing that is published or sent to someone goes out on its own. A price, a term, a delivery promise and every line bearing your name wait for a person to read and approve them, and the approval leaves a record of who and when.

The same thing changes name with the trade

The mechanism is identical; the missing material isn't. It is worth looking at your own case, because that is where you see which request is stuck today.

Sector	The missing material	Where you see it
Food and agricultureand export	The technical sheet in English, with nutritional values and allergens, requested by a buyer who has been waiting for eight days.	What a technical sheet that doesn't waste time contains
Restaurantsand bars	The menu updated in three different places, where the dish dropped in March still appears and the price is the old one.	The service of a dining room where the phone rings unanswered
Hospitalityaccommodation and events	The document stating what is included and what isn't, which today gets rewritten by hand inside every quote, always a little different.	The pre-season of an accommodation business

What this piece doesn't do

It doesn't publish on its own and doesn't send anything to anyone. It prepares, checks, queues and flags: the last click stays with a person, and this isn't a limit to remove once the system has matured.

It doesn't translate on its own the materials that have to go out in another language. Translation with the agreed glossary and the reread by someone who knows the destination market is a piece of its own, in multilingual drafts.

It doesn't do search engine positioning. Writing the material and working so that someone finds it are two different jobs: the second is presence on search engines and AI assistants, and without the first it has nothing to show.

Questions and answers