What does the myth say?
That once connected, an AI employee watches how your company works and absorbs it: your tone, your policies, your customers, your exceptions, without anyone telling it. The belief is attractive, because it promises no setup, and alarming, because it implies the thing knows more than you told it and you cannot see what. Both the attraction and the alarm are worth addressing, because the mechanism is neither as magical nor as opaque as the belief suggests.
The examination is about what learning means for a Digital FTE, and where you can see it.
Where does the belief come from?
From the way consumer AI products describe themselves, from the general sense that models learn from data, and from the fact that a Digital FTE does get better over its first weeks in a way that looks like absorption from the outside. It also comes from a real concern: that a vendor trains its model on your data, so what your FTE learned about you ends up somewhere else. That concern is legitimate and has a specific answer.
The belief conflates three things: model training, memory, and improvement from feedback. Only the last two happen here, and both are visible.
What is true in it?
That a Digital FTE does get better in its first month: the send-back reasons tune its drafts, corrected facts change its next proposal, and by the third week the send-backs are usually about tone rather than substance. That the platform pre-fills the Brand Brain from your site and from what other FTEs already know, so the first draft is not naive. And that a manager who never reads the memory will find, months later, that it contains facts nobody confirmed. Learning happens; it is just not silent, and it is not unbounded.
The myth's error is in the words on its own.
What does the mechanism actually do?
The Brand Brain is the FTE's memory of your business: voice, facts, policies, products, preferences. It is pre-filled at hiring from your website and from what the platform already knows, and every item carries a provenance, from your site, from onboarding, from a send-back on a named date, added by a named person. The manager reads and corrects it at hiring and confirms it quarterly. Send-back reasons feed proposed changes to it, which the manager approves. Facts a draft needs but the memory lacks are marked as gaps, not invented. The FTE reads from the Brand Brain when it drafts; it does not learn from anything it is not shown, and nothing it holds is used to train a model.
The Memory tab is the whole of it. If it is not there, the FTE does not know it.
- Pre-filled with provenance; confirmed by the manager
- Improved by send-back reasons and approved facts
- Gaps marked, not filled by guessing
- No model training on your data
Why does provenance matter more than learning?
Because a fact with no source cannot be corrected, only overwritten, and a memory that learns silently accumulates facts nobody can trace. When a draft contains a wrong claim about your refund policy, the useful question is where the FTE got it, and the provenance answers: from a page on your site that is out of date, or from a send-back that was misread. The correction then fixes the source, not just the draft. A memory that learned on its own could only be reset, which is why the platform does not build one.
Provenance is also the answer to the auditor's question, what does this system know about us and how, in a form you can export.
How can you check for yourself?
Open the Memory tab of any role. Pick an item and read its provenance. Correct one fact and approve the change. Send back the next draft with a reason that references a fact, and read the proposed memory change that follows. Ask in writing whether the vendor trains models on your data, and read the answer. Then ask the FTE for something your memory does not contain and confirm it marks a gap rather than inventing an answer. Fifteen minutes, and you know exactly what it learns, from where, and what it cannot learn.
The library's guide to Brand Brain memory management describes the quarterly confirmation in detail.