How to Audit AI Use Across Your Law Firm With a Tamper-Evident Record

Finding the AI tools already in use at your firm

Find shadow AI through four sources, not a survey alone. Pull single sign-on and OAuth grants to see which AI apps staff connected to firm accounts, review expense and card statements for AI subscriptions, inventory browser extensions on firm devices, and then run an anonymous survey to catch consumer tools used on personal accounts. The technical signals find what people will not self-report ^[3].

Start with the systems that already know. Your identity provider logs every application staff have signed into with a firm account, so an export of single sign-on and OAuth grants surfaces the AI tools connected to firm data. Expense reports and corporate card statements reveal paid subscriptions that never went through procurement. Managed-device inventories show which AI browser extensions are installed, which matter because an extension can read page contents including client documents.

Then close the gap those signals miss. Consumer tools used on personal accounts and personal devices leave no trace in firm systems, so an anonymous survey, framed as amnesty rather than enforcement, is the only way to see them. Ask what tools people use, for what tasks, and whether any client information has gone into them.

The output is a real inventory: tool, who uses it, on what account, for what work, and whether client data is involved. That inventory is the thing every later control depends on, and it is what our guide on finding shadow AI accounts goes into in more depth.

What to log about each AI interaction

Log six fields per AI interaction: the tool used, the person, the matter or client, the data classification involved, the purpose, and whether a required verification step was completed. This is enough to answer the two questions that matter later: did privileged data reach a third-party tool, and did a lawyer verify any output that went into a filing. Anything less cannot answer either.

The log is not surveillance; it is the evidence layer for a duty you already hold. The goal is to be able to reconstruct, for any matter, what AI touched it and whether the firm's controls were followed.

Why a tamper-evident record matters to insurers and courts

A record only helps if no one can quietly change it after an incident. A tamper-evident log is append-only and cryptographically chained, so any later edit or deletion is detectable, which is what lets an insurer or a court trust that the record reflects what actually happened rather than what the firm wishes had happened. An editable spreadsheet proves intent to track; a tamper-evident log proves the facts.

The difference between a spreadsheet and a tamper-evident record is the difference between a claim and evidence. Anyone can produce a clean log after the fact. What an insurer assessing a claim, or a court examining a firm's conduct, wants to know is whether the record could have been edited to look better than reality.

A tamper-evident log solves that by making entries append-only and chaining them cryptographically, so removing or altering any entry breaks the chain and is detectable. RankShield Legal's platform records verification and privilege-isolation events to exactly this kind of tamper-evident transparency log, signed and sealed so the firm can show, not just assert, what happened and when. That capability is live for the citation and privilege primitives it produces.

For a firm, the payoff is at the worst moment. When a client alleges its data was mishandled or a court asks how a filing's citations were verified, an append-only record answers the question with evidence, which is a far stronger position than a supervisor's recollection.

How long to keep AI usage records

Tie AI usage-record retention to your matter-file retention, not a shorter technical default. Because the AI log may be the evidence that privileged data was handled correctly or that citations were verified, it should survive as long as the matter it documents could be questioned, which usually means the firm's standard file-retention period plus any malpractice limitations tail. Set the schedule deliberately rather than accepting a vendor's default.

Retention is a decision, not a default. Many tools keep logs for a short operational window and then purge them, which is fine for uptime monitoring and useless for professional-responsibility evidence.

Match the AI usage record to the life of the matter it documents. If a malpractice claim or a fee dispute could arise years later, the record that shows the firm followed its AI controls needs to still exist then. Align the schedule with your existing file-retention policy and your jurisdiction's limitations periods, and confirm the retention rules where they vary, since records requirements differ by state and by matter type.

Turning the audit into an ongoing control

A one-time audit is stale within a quarter, because staff adopt new tools continuously. Make it continuous: schedule the identity and expense pulls to repeat, require new AI tools to be logged before first use on client work, and review the inventory at a set cadence. The aim is a living inventory and record, not an annual snapshot that is wrong by the time it is filed.

The audit you run once tells you where you stood on one day. Given how fast staff adopt tools, that picture decays quickly, so the value is in making the audit repeat on its own.

Automate the parts that can repeat: schedule the single sign-on and expense reviews, and wire new-tool logging into the firm's intake so a tool used on client work is recorded before, not after. Then review the living inventory at a set cadence, treating governance as a managed process, the posture standards like ISO/IEC 42001 describe for AI management ^[5] and a realistic firm AI policy operationalizes ^[4]. A continuous inventory feeding a tamper-evident record is the operational goal; a firm-wide dashboard that aggregates it across every tool is the direction RankShield Legal is building toward, and that broader dashboard is on the roadmap rather than shipped today.

Why the tool inventory is always incomplete the first time

A first-pass AI inventory is a snapshot of what the firm knows about, which is reliably a subset of what it uses. Shadow AI enters through personal accounts, browser extensions, and features quietly added to software the firm already licenses. The third route is the one that defeats a procurement-based inventory entirely, because nothing was ever procured.

Firms usually build the first inventory from procurement records and a survey. Both methods share a blind spot: they capture tools the firm decided to adopt, and shadow AI is defined by the absence of that decision.

Personal accounts are the obvious route and the one most policies address. A lawyer using a consumer AI account on a personal device to summarise a document leaves no trace in firm systems, which is precisely why prohibition alone does not solve it. If the approved tool is harder to reach than the unapproved one, use migrates to the easier tool regardless of policy.

Browser extensions are the underestimated route. An extension with permission to read page content can see whatever the lawyer sees, including matter data in a document management system, and it installs without any procurement event.

The third route is the one that breaks inventories built from purchase records: AI features added to software the firm already licenses. Nothing was bought, no approval was sought, and a product the firm has used for years acquires a summarisation or drafting capability in an update. The inventory says the firm uses that product, which is true and now materially incomplete.

The practical consequence is to treat the inventory as a recurring process rather than a project with an end. Re-run discovery on a cadence, add a check for newly enabled AI features in existing products, and treat every finding as information about the gap between what people need and what the firm has approved, rather than purely as a compliance failure. A tool that keeps reappearing after being blocked is telling you about an unmet need.

What the log proves, and what it does not

A usage log establishes that an interaction occurred, when, and by whom. It does not establish that the output was correct or that anyone verified it. Those are different records, and conflating them produces a firm that can show extensive AI activity and nothing about whether it was checked.

The value of a log is bounded and worth stating precisely, because a firm that misunderstands what it has will rely on it for a question it cannot answer.

What a usage log does establish is real and useful. That a specific person used a specific tool at a specific time. What category of material was involved. Whether use was inside approved boundaries. In aggregate, which tools are actually used, which is what makes the next inventory better than the last.

What it does not establish is whether the output was accurate, whether a human reviewed it before it was relied upon, or who that human was. A log showing that a drafting tool produced a document tells you nothing about whether the citations in that document were checked.

This distinction has a sharp edge in litigation. A log that records extensive AI use, with no corresponding verification record, is not neutral evidence; it documents the volume of unverified reliance. The firm has proved the exposure and none of the diligence.

So build both records and keep them linked. The usage log answers what happened. A verification record answers whether it was checked, by whom, and when, which is the question a court or an insurer actually asks. Our guide on proving verification to a court covers the second record; this article covers the first. Neither substitutes for the other.