Contents4
A client learned this the hard way. They were reworking a sales proposal with an AI assistant: one pass, two passes, about forty in total. Each time, they reread it, fixed a sentence and sent the whole thing back. Somewhere along the way, a price gained an extra zero. They did not see it. Their prospect did. The deal went no further.
This is not a beginner’s mistake. It is an effect of the system itself, and it affects everyone, whatever model they use.
In short
When you give an AI its own answer back, it rewrites the previous version, never the original. Every small drift becomes the new reference.
The more you iterate, the more likely precise details (prices, dates, figures, clauses) are to move without anyone noticing.
You do not remove drift, you build a system that makes it visible: a reference original, targeted changes, a comparison of versions, a project memory and a human review of figures.
Telephone, the AI version
A language model does not “correct” a text the way a person who keeps the original in mind does. It produces a new version from what it is given. Give it version 12 and it knows nothing about version 1. A softened phrase, a rounded figure, a forgotten condition: on the next pass, it has become the truth.
Its working memory is also limited. Over a long conversation, the first exchanges, the ones where the price or the scope was set, end up weighing less than the latest ones, or fall out of view.
The most misleading part is that each version, taken alone, looks clean. The text flows and is well built. The drift only shows when you compare.
The same effect, at company scale
The extra zero is a striking case, but the pattern is everywhere once an entire chain runs through AI: a generated specification, generated answers to it, a generated analysis of those answers, then an AI that approves the lot. The last step naturally recognises a style that looks like its own, and nothing in the chain goes back to what the company wanted in the first place.
An invoice whose total changes on the fourth copy, code that loses a security check from being “simplified” once too often, a clause that vanishes from a contract: same mechanisms.
What actually prevents it
We do not ask AI to be perfect. We organise the work so that drift cannot go unnoticed.
- Start again from the original. The reference stays the approved document, not the latest answer. Every change applies to it.
- Ask for a precise change. “Shorten the second paragraph” rather than “improve everything”. The less the AI touches, the less it can damage.
- Compare versions. A gap between two versions shows up in seconds side by side, never by rereading a text that looks right.
- Keep a project memory. A decisions file next to the work: what has been settled (price, scope, tone, source figures), read before each pass and updated after. The AI has a fixed place to read what must not move, instead of relying on the flow of the conversation.
- Have a human check the figures. Prices, amounts, dates, commitments: whatever is costly to get wrong goes past human eyes before it leaves.
How we work
This is how we run our projects, and what we set up for our clients. Every project has its written memory, which agents read before acting and update after. Deliverables go through automatic checks that compare what changed and reject what does not match the source. And whatever commits the company, a price, a published figure, a clause, is reviewed by a person before it goes out.
AI saves a considerable amount of time. This system makes sure that time is not paid back later, in front of a client.
Jim · founder of Brain Plus AI
Also in business ai
Stepping back10 min
The human role alongside AI: knowing how to judge
AI excels at some tasks and is confidently wrong on neighbouring ones. The human's value shifts: knowing where the line runs, keeping your judgement, answering for the result.
Stepping back6 min
The 70/20/10 rule of AI: quoted everywhere, measured nowhere
70% of the effort on people and processes, 20% on technology, 10% on algorithms: the rule circulates like a survey finding. It is an in-house BCG rule, worded six ways in three years. How it became a statistic, and why the idea still holds.
Understand7 min
Who enforces the AI Act in France? The law is still pending
Adopted by the French Senate on 18 February 2026, the bill naming the AI Act enforcement authorities is still waiting for the National Assembly. Who would supervise what under the Senate text, what already applies without it, and what a business can prepare in the meantime.