At most small law firms, an AI policy is a rule with no record that anyone followed it. The American Bar Association counted 11% of the profession using AI in 2023, and 30% a year later. Clio's 2026 report finds that three quarters of solo and small practices use AI, and 55% with no AI-use policy at all.
The decision has already been made
Picture an associate with a forty-page commercial lease to summarise before a nine o'clock call. The firm has never said anything about AI. So she pastes the lease into a public chatbot and has a summary in four minutes.
She is being efficient. But the lease shows the client's name on every other page and the guarantor's Social Security number in the schedule. All of it is now inside a consumer assistant never built for client data (what "safe" means for a contract). Clio found this is the normal case at small practices, where AI arrives under a deadline rather than by decision.
Faster work, same invoice
60% of large firms told Clio that AI is growing their revenue. Fewer than a third of solo and small practices said the same. Large firms pay someone to reprice work that AI has shortened, and small firms do not. So a three-hour job is billed as one hour: the client keeps the saving and the firm keeps the risk.
A rule with no record
Across the profession, the North Carolina Bar Association's survey round-up finds 9% of firms with a written AI policy that is actually enforced. Once a policy is on the shared drive, it tells nobody what people actually do.
Suppose an associate pasted a client's medical records into a chatbot last Tuesday. How would you find out? At most firms this size, you find out when the client does.
Banning AI fails, because it stops the reporting but not the use. The rule that works is also the quickest option. Make the protected route the faster one, and compliance stops depending on self-discipline.
Working with real files
Lawyers work with files, while most tools that strip personal data want a string of text. The Library takes PDF, Word, Excel and PowerPoint files, plain text and images.
Scanned pages are read automatically when a file opens. Detection then runs four layers in one pass:
- Patterns finds values with a fixed format, such as Social Security numbers and account numbers.
- Entity recognition finds values without one: names, employers, addresses.
- Expressions finds your own terms, such as a deal code name or a subsidiary.
- Deep Analysis is optional and handles edge cases.
By default, a file cannot be downloaded until a person has verified the findings. Pressing Download first explains what is outstanding and takes you to the review. You can turn this gate off in Settings.
Redact removes text permanently, which suits anything leaving the firm. Encode swaps a value for a short code and keeps the mapping in a registry on your own machine. The registry is not encrypted, because you already hold the original on the same disk.
Inside a folder, each value gets one code and keeps it. The model can then tell the tenant in the lease is the tenant in the letters. Other folders get different codes for the same value, unless you pin a code by hand. Decode restores the real names locally, from a file or pasted text.
Where the record comes from
All of that makes the work safe, but not provable. In the browser dashboard, you assign each licence to an email address and tag it by team or matter type. Revoking one ends it at the desktop's next online check. A paid plan includes two device slots per person, and a trial gets one.
The Compliance Monitor shows each seat's documents exported, pages processed, values detected, encoded and redacted, and last activity. It marks each seat as compliant or not against a threshold you set. The view exports as a compliance report, and only the account owner can open it or the audit log.
One limit: the monitor reflects only activity in promptShield and cannot see a document handled outside the tool. A seat activated offline is flagged, because its activity cannot be verified. If a partner pastes a contract into a browser at home, nothing here will tell you.
"We have an AI policy" is a claim. A weekly report showing nine of ten seats protecting documents, and the tenth you followed up with in March, is a record an insurer can read.
Common questions
Does a small law firm need a written AI policy?
If anyone uses AI on client work, yes. State bar rules on competence and confidentiality already reach AI, and clients and insurers now ask about it.
What should a law firm AI policy actually contain?
Keep it to one page: approved tools, what must come out of a document first, and the steps to take it out. Add who to ask, and how the firm will see whether any of it happened.
Do our documents leave the office if we use this?
Not in the shipped configuration: detection, text recognition and redaction run on your machine, and our servers receive only licence checks and counts, never document content. Deep Analysis is optional and local by default. If you point it at a third-party endpoint, it sends page text.
Which plan does a ten-person firm need?
Basic is the desktop app: detection, Redact, Encode, the local registry and Decode. Pro adds the licence manager, the audit log and the Compliance Monitor, so you can show the protection afterwards.
Start this week
Pick the one job that pulls people toward a chat window, define the protected way to do it, and make it quicker. Check the record after a month.
We got this wrong ourselves. Download on an unverified file used to do nothing in our app, so people worked around it. It explains itself now.