Your associate drafts a client follow-up with an assistant on her own login, at home, on a Sunday. On Monday the email goes out, and it reads well. A month later the client asks why a figure in it was wrong. Nobody at the firm can see what she asked, what came back, or what she changed.
Whose record is it?
It is hers, not the firm's. A personal account is tied to a person and their own settings.
The firm has no admin seat on it. It cannot export it, search it, or keep it when she leaves.
So the work was done for your client, and the record of how it was done sits outside your walls. That is the whole problem in one line: the firm signed the work, but it does not hold the record.
Why does this matter more than which tool she used?
Because the procedure matters more than the product. The plan choice is smaller than people think.
The larger question is this: could anyone at the firm find, a month from now, what your AI wrote for a client? If the answer depends on one person's private login, the answer is no.
Even business plans differ in what they keep and who can reach it. The OpenAI help center says that in ChatGPT Business, admins cannot see members' private chats by default, and self-service export is not available. If a firm workspace can be that closed, a personal account is further out still.
I will not rewrite the plan details here. They are in which plan, and what it records. The point for this note is simpler: a firm account is where a record can start. A personal one is where it ends.
What does supervision need that a personal account cannot give?
It needs four things written down, the same four you would want for a capable new hire.
Scope: what the assistant may touch, named as specific data, apps and APIs. A personal account has no scope you set. It has whatever she pasted in.
Briefing: what it needs to know, written rather than carried in someone's head. In my experience, a bad output is rarely the model and rarely the prompt. It is almost always what you did not tell it. A briefing that lives in one private chat history cannot be checked or shared.
Review: a named person reads the output before anyone outside the firm sees it. That only works if the reviewer can see what was asked as well as what came back.
Record: what was asked, what came back, where it lives. A personal account answers the last part with "somewhere you cannot reach."
The models make a capable person's mistakes. They misread what you meant, fill a gap with a guess, and sound surer than they should. It writes like it is sure, and that is not the same as being right. You cannot catch that pattern in work you cannot see.
Does keeping it in firm accounts mean using AI less?
No. It usually means you can hand off more.
AI earns its keep in the work in front of a judgment, never in the judgment itself. That covers two kinds of work. Work that repeats and stays inside the firm. And client work that someone at the firm can check, because they already know what right looks like.
When the line is drawn clearly, people stop guessing what is allowed. They stop drifting to their own logins to get something done quietly. A clear line around what stays human lets a firm delegate more, not less.
What I see most often is not bad intent. It is a capable person who was never told where the work should live, so they used the tool already open on their phone. The fix is a written rule and a firm account that is easy to reach, not a lecture.
One practical step: ask each person which AI tools they used for firm work last month, and on which login. Write the answers down. That list is the start of your scope.
A personal account is a private notebook. Client work belongs in the firm's filing cabinet, where someone else can open the drawer.
If you want a second pair of eyes on where your firm's AI work actually lives, start with the free Second Look.
