The usual story is not a partner announcing an AI strategy. It's a candidate attorney at 7pm, with a letter from the other side, pasting it into a free chatbot "just to get the gist." Nobody malicious. Nobody trained. The file has now been in a system the firm doesn't control, and the summary that comes back is what they'll half-remember in the morning.
I'm not your attorney, and this is not a legal opinion on privilege or POPIA. It is the line I use when a firm asks where to start. Admin can move. The matter cannot go wandering.
The letter and the diary are different objects
Firms lump "AI" into one decision, then either ban it or ignore the ban. Split the work by what a stranger would learn if they saw the text.
| This can be drafted from a template | This stays with a person, on the firm's own systems |
|---|---|
| We've received your instruction. Your file number is 4418. Please send the ID copy and the rates account. | Any summary of what the client did, or what the other side alleges |
| A checklist for opening a matter: conflict search, FICA, engagement letter, file number | The advice itself, including a "quick view" on prospects |
| A Friday list of matters with no time posted and no activity in 14 days | A draft that quotes the pleadings, the medicals, or the WhatsApp thread |
| Diary holds that name the meeting, not the facts | Anything you'd be unhappy to read back in a discovery affidavit |
The left column is how a practice stops leaking evenings. The right column is why a blanket "the firm uses AI now" email is a bad idea. People hear permission. They paste the interesting document, because the interesting document is the one they don't want to read twice.
A new instruction at 16:40
Say a conveyancing client phones as the receptionist is packing up. The useful automation is boring on purpose. A matter shell opens with the client's name, the property description they already said on the phone, and an empty list: ID, rates figures, existing bond, HOA rules if it's a complex. A conflict search is queued for a person, not waved through by a model that has never heard of the cousin on the other side. An engagement letter goes out from a template the director approved in March, with the fee in it, not a fee a model guessed.
Nothing in that sequence needs the model to know why the sale is tense. If the template starts asking for "a short background so we can personalise the letter," delete that sentence. Personal in a law firm is the person's name and the property. It is not a retelling of the dispute.
The operational version of this, once a firm knows the boundary, is what we describe on the law firms page. This article is the boundary. That page is the install.
If the prompt needs the facts of the matter to be useful, it is not an admin prompt. Close it.
Where people paste anyway
Three habits show up in almost every firm that "hasn't really started with AI" and already has:
- A personal ChatGPT login, paid for on a director's card, shared in a WhatsApp group
- A junior using the free tier because the work password is a fuss
- Someone dropping a PDF of pleadings in to "see if it can find the date"
The fix is dull. One firm account, on a plan where you can switch training off and see a log. A written list of file types that are allowed (blank precedents, your own letterhead templates, a diary export with the notes stripped). A list of file types that are not. If a tool can't show you the log, it doesn't get client work, even the boring kind.
POPIA is the statute your clients have heard of. Legal professional privilege is the older problem, and they are not the same problem. A message can be lawful to store and still be a terrible thing to have handed to a vendor. Ask your own professional indemnity broker what they want to see before you connect a model to the document store. Don't take a blog post's word for it, including this one.
The first job I'd actually switch on
Not research. Not drafting particulars. A document chase that does not describe the case.
"We still need the rates clearance figures and a copy of your ID. Reply to this email with both." That can run on a timer for matters in a named stage, and stop the moment the client writes anything other than the attachment. The reply that says "also, the buyer is threatening to walk" is a person's email. The system should not answer it.
Second, the Friday list. Matters with no activity. Fee earners hate it and then use it. It doesn't need to be clever. It needs to be the same list every Friday, from the practice system, not from someone's memory of who they meant to phone.
If the firm isn't a law firm at all, the general version of picking one job is how to integrate AI when you don't know where to start. Different industries leak time in different places. An estate agency's Sunday seller update is not this problem, and a store's "where is my order" mail isn't either.
If you want the checklist held up against the way your matters actually open, that's a discovery call. Bring a blank precedent, not a live file.
Questions people ask
Can we paste a client letter into ChatGPT to summarise it?
I wouldn't. A consumer chat window is a third party, and you often don't control what it keeps. Summarise from inside a tool the firm pays for, on an account the firm owns, with a written rule about what may be uploaded. If you don't have that yet, a person still does the summary.
Is diary management safe?
Yes, if the entry says 'consult, 10:00, van der Merwe' and not the facts of the dispute. The danger is the note someone adds because they're being thorough.
Will this replace candidate attorneys?
Not the thinking. It can take the chase emails and the opening checklist off their evening. The advice, the draft that quotes the facts, and the call with the client stay with people who carry a practising number.
Do we need a formal AI policy first?
A one-page rule is enough to start: what may be uploaded, to which system, who can see the log, and what never leaves the file. A 20-page policy nobody reads is how the paste still happens on a Sunday.