Version control for legal documents in AI workflows
By the ChamberOne team 4 minute read
Legal teams have always struggled with versions: final_v3, final_v3_revised, final_FINAL. AI adds a new layer of confusion. A tool summarises the contract as it stood on Monday. By Wednesday, the other side has sent changes. The summary is still on the file and still looks authoritative. Or a draft produced by AI is saved beside the lawyer's accepted text with no label to tell them apart. Good version control in an AI workflow comes down to three disciplines: know which version the tool saw, keep drafts distinct from accepted text, and stop stale material reaching clients and other parties.
Key points
- An AI output describes only the version it was given; record which version that was.
- Mark earlier analyses as superseded when a new version arrives and re-run them.
- Use status labels that separate AI drafts, reviewed drafts, approved text and sent versions.
- Keep one master location per document and avoid the word 'final' in file names.
- Run a short pre-send check for the right version, tracked changes, comments and metadata.
Why AI makes versions riskier
An AI output is only as current as its input. If the tool analysed version two and version four is now live, the analysis describes a document that no longer exists. People forget this, because the output looks complete and up to date.
Outputs also multiply. A single request can produce several drafts, and each may be saved, pasted into an email or edited. Without discipline, the file collects many near-identical documents, and nobody can say which one the partner approved.
Record which version the tool saw
Every AI output that relates to a document should carry a pointer to the version it analysed: the file name, version number or date and time. If your system records this automatically, use it. If not, add a line at the top of the output or in the matter notes.
When a new version of the document arrives, mark earlier analyses as superseded. Better still, re-run the analysis on the new version and compare, since changes between versions are often where the risk lies.
- 1Record which version the tool analysed
- 2A new version of the document arrives
- 3Mark the earlier analysis as superseded
- 4Re-run the analysis on the new version
- 5Compare the two results
- Then back to step 1.
Separate drafts from accepted text
The most important line in an AI workflow is the one between what the tool produced and what a lawyer has adopted. Use clear status labels and apply them consistently.
- Draft (AI): produced by a tool, not reviewed.
- Draft (reviewed): reviewed and edited by a lawyer, not yet approved.
- Approved: signed off by the responsible lawyer for use.
- Sent: the exact version that went to the client or another party, with the date.
- Superseded: replaced by a later version, kept for the record.
Naming and storage habits
Choose a simple naming convention and make it a habit: matter reference, document name, date or version number and status. Avoid words such as 'final' in file names, since there is always another one.
Keep one master location for each document and resist saving copies elsewhere. If people must work offline, agree how changes come back, and who merges them. A practice manager can check a few files each month for stray copies and tidy them up.
- Use one simple naming convention: matter reference, name, date or version, status
- Keep one master location for each document
- Agree how offline changes come back and who merges them
- Use the word 'final' in file names
- Save copies elsewhere
Avoiding stale attachments
Sending the wrong file is a familiar mistake with serious consequences. An outdated draft that includes comments, tracked changes or AI-generated notes might expose confidential reasoning. Before any document leaves the firm, run a short pre-send check.
- Open the file you are about to attach, not the one you remember.
- Check the version label and date against the matter record.
- Look for tracked changes, comments and hidden text and remove what should not go out.
- Check document properties for author names and earlier metadata.
- Confirm that the version has been approved by the responsible lawyer.
- After sending, save the sent copy and label it as sent.
A hypothetical example
Suppose a four-person firm is negotiating a services agreement. On Monday, a paralegal asks a tool to list the termination and liability clauses of the counterparty's draft, and saves the list to the matter with a note: analysed version two, Monday. On Wednesday, version three arrives. The paralegal marks Monday's list as superseded, re-runs the analysis and compares the lists. A change to the liability cap, which was buried in a reformatted paragraph, shows up.
Without the note, the Monday list might have been forwarded to the client as current.
Handling documents that several people edit
Negotiated documents pass through many hands: the client, a colleague, the other side and their advisers. Decide who holds the master and say so. Others send comments, and the holder merges them. A shared document edited live by everyone can be convenient, but it makes it hard to say later who changed what and when.
When an AI tool is used on a shared draft, make sure it works on a copy or a labelled snapshot. Otherwise its suggestions may be applied to text that has moved on, and the changes pile up unnoticed.
At the end of a negotiation, save a clean execution version and a record of the final agreed changes, and mark everything else as superseded. That tidy-up is easier at the time than a year later.
Retention of old versions
Superseded versions are part of the history of a matter, and they may matter in a dispute about what was agreed and when. Decide how long drafts, including AI drafts, are kept and where, and apply the decision consistently. Your file retention policy should say so.
Comparing versions
Comparison tools, whether built into word processors or offered by AI features, help to spot changes. Use them, but do not rely on them alone, especially after conversion from PDF or where formatting has changed. A formatting change can generate noise, and a tool may miss a subtle change in a defined term.
For key documents, read the changed clauses in full and ask the other side to confirm that no changes were made other than those marked.
Limits
Version control is a discipline, not a feature. The best system fails if people save to the desktop. It helps to explain why the rules exist, using a real near miss, and to keep the rules few. Review them once a year with those who use them, since the rules that survive are the ones that fit how people actually work.
This article is general information about the law, legal practice and technology, current when it was published. It is not legal advice. Laws and guidance change: check the sources linked above and take advice on your own circumstances from a lawyer qualified where you are.
Read next
Onboarding a new client matter: an AI-assisted step-by-step checklist
Walk through every step of opening a new matter — from conflict check to task assignment — showing where AI saves time and where only a lawyer can decide.
6 minute readUsing AI matter briefs to pick up an unfamiliar file quickly
How a generated brief speeds up handovers and cover, and how to check it against the file before acting.
5 minute readAI task suggestions and matter phases: standardising how work starts
How matter-type templates and suggested tasks help new files start consistently, and why each suggestion needs review.
5 minute read