Lesson objective
Build a permitted-use inventory that converts a general policy into a list of specific approved things, each with a tool, a data class, conditions, an owner and a reopening trigger.
Plain-language explanation
Why this matters now
The permitted-use inventory is the part of the programme that does the most work, and it is the part almost every firm omits. It is the difference between a firm that can answer "what are you permitted to do with AI?" with a list, and a firm that answers with a paragraph about responsible use.
The reason it matters is structural. A general permission — "approved tools may be used on client work" — permits an unbounded set of activities, because the interesting variation is not in the tool. It is in what the tool is asked to do, what data it is given, and which matter it touches. The same licensed product used for an internal knowledge query, for a first-pass clause extraction and for drafting a section of a brief presents three different risk profiles and three different verification requirements. A permission granted at the level of the product cannot distinguish them.
Professional guidance: ABA Formal Opinion 512 (29 July 2024) addresses competence, confidentiality and informed consent — including that informed consent may be required before confidential information is input into a self-learning tool — candour, supervision, and billing only time actually expended. Each of those turns on what is being done, with what data, on whose matter. An inventory is simply the place where a firm records the answers in advance rather than working them out per lawyer, per matter, at speed.
Law / rule: Model Rules 1.1, 1.4, 1.6, 3.3, 5.1 and 5.3 supply the tests. They are model provisions adopted differently in each jurisdiction, with different numbering and wording; apply the version adopted where you practise.
A question to sit with
Ask your firm for a list of the AI uses it has approved. If what comes back is a list of products rather than a list of uses, then nobody has yet decided what anyone may do — only what the firm has bought.
What goes wrong
Permission granted at the product level. The firm approves a platform. The platform grows features. Two releases later the approval covers things the firm never assessed, and nobody notices, because the approval was never expressed in terms of what may be done.
Data classes left implicit. The inventory says a use is permitted and is silent on what may be entered. Individual lawyers then make an unguided judgement about client confidential material under deadline, which is precisely the judgement the programme existed to make in advance.
No conditions, so the permission is absolute. A use is approved without stating the verification required, the matter types excluded, or the client constraints that override it. The row then says yes in situations the approver never contemplated.
No owner, so nothing is maintained. A row with no owner is never reviewed, never retired, and is still on the list three years after the workflow it described stopped existing.
No reopening trigger. The approval was given on the basis of facts — a contractual term about training, a retention period, a tested accuracy level. Nobody wrote down which of those facts, if it changed, would reopen the decision, so when it changes nothing happens.
Embedded features are invisible. The inventory lists what the firm chose. The features that arrived inside products the firm already licenses are the ones creating exposure nobody assessed.
The practical answer
Write the inventory as a table with seven columns and treat each row as a decision.
| Column | What it holds |
|---|---|
| Use | What is actually being done, in a verb phrase: "first-pass extraction of specified clause types from executed agreements" |
| Tool and tier | The specific product, plan and deployment; the plan matters because terms differ between tiers |
| Data class permitted | Which of the firm's data categories may be entered, by exception rather than by default |
| Conditions | Verification required, matter types excluded, client constraints, volume limits |
| Named owner | A person, licensed where the judgement is professional |
| Evidence relied on | What the approver actually saw: contractual terms, a test result, a security review |
| Reopening trigger | The specific change that forces re-assessment |
Classify data before you classify uses. Best practice: three or four tiers is enough for most firms, and more tiers reduce compliance rather than improving it.
| Class | Examples | Default posture |
|---|---|---|
| A | Public material, published authority, firm marketing text | Permitted in any approved tool |
| B | Internal firm material with no client identifiers | Permitted in approved tools with a contractual exclusion from training |
| C | Client confidential material, matter documents, personal data | Permitted only where the inventory row says so, and only where the client obligations check has been completed |
| D | Material under a protective order, a client prohibition, or a heightened restriction | Not permitted without the programme owner's written approval on the specific matter |
Make the reopening trigger specific enough to fire. "If circumstances change" never fires. These do.
| Approval rested on | Reopening trigger |
|---|---|
| A contractual exclusion from model training | Any change to the supplier's terms, or a move between plans |
| A stated retention period | Any increase in retention, or a change in where data is stored |
| A tested level of performance on a defined task | A material model change, a supplier change, or a change in the task |
| The client having raised no constraint | Any new or amended outside counsel guidelines or engagement terms |
Distinguish confabulation risk from automation-bias risk in the conditions column. Professional guidance: the NIST Generative AI Profile, NIST AI 600-1 (26 July 2024), names confabulation and automation bias as generative-AI risks and links them: excessive deference to the system can worsen the risk of confabulated output. A row whose output is a set of assertions that look like authority needs a verification condition naming the primary source. A row whose output is a summary a reviewer will skim needs a condition addressing what the reviewer must independently establish, because the failure will be silent.
Build the inventory by discovery, not by asking. Best practice, drawn from the Map function of the NIST AI Risk Management Framework 1.0 (26 January 2023) and the NIST AI RMF Playbook, both voluntary and neither creating nor discharging any professional duty: find the actual uses. Review software charges on the finance ledger, inventory installed browser extensions, ask what features arrived in the last two releases of every platform the firm licenses, and read the release notes. Asking lawyers which AI tools they use produces a list of the ones they think are approved.
Remember this
- Permission belongs at the level of the use, not the product; the same tool doing three different jobs is three different decisions.
- Seven columns per row: use, tool and tier, data class, conditions, named owner, evidence relied on, reopening trigger.
- Classify data into three or four tiers first; more tiers reduce compliance rather than improving it.
- A reopening trigger must name the specific change that fires it — a terms change, a retention change, a model change, a new client guideline.
- An embedded feature the firm never chose is a use that is not in the inventory and therefore not approved; only the programme owner adds it.
Legal example
Fairhurst Obadele LLP is a fictional 95-lawyer firm that has approved three AI products and believes it has a governance programme. Its invented programme owner, a partner named Ingrid Achterberg, sets out to build the inventory and starts, as the lesson advises, with the finance ledger rather than with a survey. The ledger shows eleven recurring software charges she cannot account for, of which four are consumer AI subscriptions billed to individual expense claims. A browser-extension inventory from the IT team returns two more, one of which reads the content of any page it is open on. The release notes for the firm's document management platform show a summarisation feature switched on by default four months earlier. Before she has written a single row, the firm has gone from three AI tools to seventeen. Ingrid then writes the rows. The first is easy and instructive: first-pass extraction of six named clause types from executed agreements, on the licensed diligence platform, enterprise plan, permitting class C material, conditional on every extracted provision being confirmed against the source document by a named person and on the absence of any provision being treated as a finding to be checked rather than as an absence, owned by the head of corporate, approved on the evidence of a written supplier answer excluding customer content from training and a thirty-day retention term, and reopened by any change to those terms, any move between plans, or any material model change. It takes her forty minutes and it settles the question for a practice group of twenty-two people. The second row is harder. A senior associate has been using an approved research assistant to draft short sections of briefs, which nobody ever approved as a use. Ingrid does not ban it. She writes the row with a condition: permitted for first draft prose only, class B material only, no authority or quotation may originate from the tool, and every authority in the finished document is checked against the primary source by a named admitted attorney recorded per authority. The four consumer subscriptions get no row. They get a thirty-day wind-down, a named alternative, and a conversation, because banning without an alternative is what created them.
Responsible-use warning
Never approve a product when what you mean is a use. A product approval is granted once, on the facts of one demonstration, and then silently extends to every feature the supplier adds for as long as the licence runs. The firm ends up permitting activities it never assessed, on data classes it never considered, with verification conditions nobody wrote, and the approval record shows only a tick beside a vendor name. Approve the use, name the data class, state the conditions, and record what the approver actually saw, so that when the product changes the firm can tell whether the approval still stands.
Human-review requirement
A licensed attorney owns every output that reaches a client or a court, and a licensed attorney must own each inventory row, because the judgement is professional, not procurement. The reviewing attorney confirms that the data class permitted for the row is consistent with the duty of confidentiality as adopted in the relevant jurisdiction, that any client-specific constraint in engagement terms or outside counsel guidelines has been checked before the row applies to that client, that the verification condition is one a real reviewer can perform at the stated volume, and that the evidence relied on was actually seen rather than assumed from a supplier claim. Where the evidence column cannot be completed, the use is not yet approved.
Practical exercise
Build the first five rows of a permitted-use inventory. Use fictional or non-confidential material.
- Find the real uses before you list any. Review software charges on a ledger, list installed browser extensions, and read the last two sets of release notes for every platform the firm licenses. Write the total number of AI uses you discover and how many were previously known.
- Write your data classification in three or four tiers, with two concrete examples in each and a default posture for each. Resist a fifth tier.
- Draft the seven-column table headings, then complete the row for the use your firm relies on most, leaving nothing blank.
- Complete a second row for a use that exists but was never approved. Do not ban it; write the conditions under which it would be acceptable, and be specific about verification.
- Complete a third row for a feature embedded in a product the firm already licenses. Note who currently has access to it and whether the firm chose to switch it on.
- For each of your three rows, write the reopening trigger as a specific event a named person would notice. Then write who that person is and how they would notice.
- Write two rows you will refuse, with the reason and the evidence gap that would have to be closed. A refusal with a stated gap is a better governance artefact than a silence, and it goes into the decision record in module 3.
Knowledge check
-
Why is approving a product not the same as approving a use?
Answer guidance: Because the risk lies in what the tool is asked to do, what data it is given and which matter it touches, not in the product itself. One licensed product used for internal knowledge queries, first-pass clause extraction and brief drafting is three risk profiles and three verification requirements. A product-level approval also extends silently to every feature the supplier later adds.
-
What are the seven columns of an inventory row?
Answer guidance: The use as a verb phrase; the specific tool, plan and tier; the data class permitted; the conditions, including verification, excluded matter types and client constraints; the named licensed owner; the evidence the approver actually relied on; and the reopening trigger that forces re-assessment.
-
Why must a reopening trigger name a specific event?
Answer guidance: Because a trigger phrased as "if circumstances change" never fires. The approval rested on particular facts — a contractual exclusion from training, a retention period, a tested performance level, an absence of client constraints — and the trigger must name the change to one of those facts, together with the person who would notice it.
-
How should a firm discover what AI uses it actually has?
Answer guidance: By discovery rather than by survey: software charges on the finance ledger, an inventory of installed browser extensions, and the release notes of every platform the firm licenses. Asking lawyers which AI tools they use returns the ones they believe are approved, and misses both shadow use and features that arrived inside existing products.
Sources and further reading
- American Bar Association Standing Committee on Ethics and Professional Responsibility — Formal Opinion 512: Generative Artificial Intelligence Tools (2024-07-29). Accessed 2026-09-14.
- American Bar Association — Model Rules of Professional Conduct. Accessed 2026-09-14.
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (2023-01-26). Accessed 2026-09-14.
- National Institute of Standards and Technology (AI Resource Center) — NIST AI RMF Playbook. Accessed 2026-09-14.
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1 (2024-07-26). Accessed 2026-09-14.
- International Organization for Standardization and International Electrotechnical Commission — ISO/IEC 42001:2023 Information technology — Artificial intelligence — Management system (2023-12-18). Accessed 2026-09-14.