---
title: "How it works"
url: "https://ailawframework.com/how-it-works"
type: "how_it_works"
site: "AI Law Framework"
---

# How it works

> How the AI Law Framework decides whether a firm may use a particular tool for a particular legal task, and who does what once it may.

Canonical page: https://ailawframework.com/how-it-works

Site index for agents: https://ailawframework.com/llms.txt

## One task. One tool. One decision.

This framework answers a narrow question: may this firm use this particular tool for this particular piece of legal work — and once it may, who drafts, who checks what, and who signs it off?

It is deliberately not a rating of AI products. The same tool can be right for one task and barred for the next, because what changes is the work, not the software.

### The five steps

Every task goes through the same five, in two phases.

- **Building the toolkit**: Steps 1 to 3, before any work on a matter starts: record the products the firm holds or is considering, assess each one for this task and decide, then set the plan for the human work. Once for each task — not once for each matter — and reviewed at least yearly, or sooner if the task, the product or its terms change.
- **Daily use**: Steps 4 and 5: the lawyer gets the task list and the checklist for the matter, does the work to the plan, and the checks and the approval are recorded. Every matter that uses the selected tool.

1. #### Define the task
   Say exactly what you are producing, for whom, and what it has to be right about.
   - **Phase**: Building the toolkit
   - **You bring**: The instructions, the sources you may use, the jurisdiction, and who is responsible for each part.
   - **You end up with**: One task, described closely enough that two people would assess it the same way.
   - **Why it is there**: Every option, including doing it by hand, is judged against the same description. A vague task cannot be assessed at all.
2. #### Assess and decide
   Score the tool on three separate questions, and let the answers decide whether it may be used.
   - **Phase**: Building the toolkit
   - **You bring**: What the firm can run, what the work would save, and what your evidence about the supplier actually shows.
   - **You end up with**: One of three decisions: use, use with conditions, or do not use for this task — which reads “until confirmed” while a mandatory answer is outstanding. A permitted AI option is always use with conditions; only the manual route is a plain use.
   - **Why it is there**: The same tool can be fine for one task and barred for the next. The decision belongs to the task, not to the tool.
3. #### Plan the human work
   Decide, stage by stage, who drafts, who checks what, and who approves.
   - **Phase**: Building the toolkit
   - **You bring**: The permitted options, what the task risks getting wrong, and what the professional requirements demand of it.
   - **You end up with**: A plan with a level for every stage, the checks named, an approver and a release.
   - **Why it is there**: A tool being permitted says nothing about who does what. This is the part that tells a junior what they are responsible for.
4. #### Do the work
   Carry out the task under the plan, in your usual systems.
   - **Phase**: Daily use
   - **You bring**: The plan, the sources, and the tool if one is permitted.
   - **You end up with**: The output, and a record of how it was produced.
   - **Why it is there**: The plan is only worth something if the work actually follows it.
5. #### Check, approve and record
   Make the plan’s checks, above all at the high-risk points, have a named person approve the exact output, and release it as a separate act.
   - **Phase**: Daily use
   - **You bring**: The output, the documents behind it, and the plan you accepted.
   - **You end up with**: A record of who checked what, how many items and for how long, what changed, who approved it and when it went out.
   - **Why it is there**: This is the record that answers, later, what was checked and by whom.

### The three questions

Step 2 asks three separate questions about one configuration doing one task. Each is scored from 1 to 5 against fixed wording. **They are never averaged.** The overall rating is the lowest of the three, and due diligence is a gate: the rating cannot be higher than the due diligence score.

#### Adoption

Can the firm put this into practice and keep it working?

Adoption feasibility

#### Efficiency

Does it save total human work at the required quality?

Efficiency likelihood

#### Governance

Can the firm’s governance carry the risk this use creates?

Governance suitability

**Why not just average the three?**

Because averaging lets a good answer hide a bad one. A tool that scores 5 for saving work, 4 for being easy to run and 1 because a mandatory condition fails averages to 3.33 — which reads as “moderate”, and conceals the one answer that should have stopped it.

So the three stay separate and the weakest one governs. The overall rating is that lowest score. It is not an average. A score of 1 on any question means the tool is not used for that task, however strong the other two are. Recording a different overall rating does not change that recommendation.

#### What each score means

| Score | Adoption | Efficiency | Governance |
| --- | --- | --- | --- |
| 1 | Access, information or reviewer absent. | Minimum quality unreachable, or checking impossible. | A mandatory condition confirmed to fail, or stated by the supplier to be impossible. |
| 2 | Preparation dominates, or comprehension depends on one person. | Forecast: more total work than the manual route without a matching quality gain. | An applicable mandatory AI-specific answer is Not confirmed. |
| 3 | Feasible within existing resources, with manageable setup. | Reference point: quality and total work comparable with careful manual work. | Applicable mandatory conditions met, with at least one answer Public claim only. |
| 4 | Good fit; changes announced or detectable, and reassessment assigned. | Forecast: a task-relevant advantage that survives all required checking. | Controls cover the task; applicable mandatory answers Confirmed, or none applies. |
| 5 | Strong fit and sustained capability, with negligible added burden. | A strong advantage after all required work, shown by an observed result. | Prevention, detection, response and recovery integrated with existing frameworks, including a supplier commitment to notify incidents. |

Doing the work by hand is always one of the options, and it is scored too: adoption 5, because the firm already works this way; efficiency 3, the reference point every tool is measured against; governance 4, because no AI condition applies to it.

### What your evidence allows

The governance question is the one where your evidence, not your opinion, sets the limit. Each mandatory condition sits in one of four states, and the weakest one caps the mark.

| What you can honestly say | Recorded as | Caps governance at | Decision | What to do |
| --- | --- | --- | --- | --- |
| We checked — it fails. | Confirmed as failed | 1 / 5 | Do not use for this task | The condition is not met, or the supplier says it is not. Nothing lifts this except the condition being met. |
| Nobody has confirmed it. | Not confirmed | 2 / 5 | Do not use for this task until confirmed | Record the question, the document that would answer it, who is getting it, and what you will do meanwhile. An unanswered question is not a soft yes. |
| The supplier says so. | Public claim only | 3 / 5 | Use with conditions | Name the binding document you need and when the condition lapses. A higher opinion does not lift a ceiling; a document does. |
| We hold the document. | Confirmed | no limit | Use with conditions | Your own judgement now sets the mark. An AI option is still used with conditions, because its plan’s checks apply. Confirmed means published binding terms or scoped independent evidence for the exact answer — not the firm’s own signed agreement. |

You do not invent the governance number. Your evidence sets the limit and your judgement works inside it. To lift a ceiling, get the answer in writing.

#### The decision

- **Use** — the manual route, with nothing outstanding. Only doing the work without AI reaches a plain “Use”.
- **Use with conditions** — every permitted AI option, because its human-work plan and your professional responsibility still apply. Governance held at 3, a document, setting or check still outstanding, or a 2 for being hard to run or not worth the work with a recorded condition that removes its cause, each adds a condition of its own.
- **Do not use for this task** — a score of 1 on any of the three, or a 2 for being hard to run or not worth the work with no condition that fixes the cause. Where governance is held at 2 only because an answer is unconfirmed, this reads **Do not use for this task until confirmed**: record the question, the document that would answer it, who is getting it, and what you will do meanwhile.

Three decisions, and only these — “until confirmed” is the third one worded for an answer you are still waiting on, not a fourth. A decision belongs to one task and one configuration: reassess when either changes.

### Who drafts

Being allowed to use a tool says nothing about how much of the work it may do. That is set separately, by the level. The four levels are defined by who drafts and who checks, and a plan takes its headline level from the first-draft stage.

#### Green

AI drafts, lawyer checks

The tool writes the first draft inside a template you have already tested. You check it.

**Use it for:** Procedural or internal work, inside a validated template or playbook, with recorded task-level capability evidence such as a test of the tool at setup. Without that evidence, first drafting starts at Yellow.

#### Yellow

Lawyer and AI draft together

You and the tool draft together, with you directing. You check every statement you rely on.

**Use it for:** Substantive work that leaves the firm. This is the floor for anything a client or another party will read.

#### Red

Lawyer writes; AI may check, clarify or collect data

You write the text. The tool writes nothing into the document. Where permitted, it may check, clarify or collect data, and anything it checks or collects is verified.

**Use it for:** Anything sworn or certified; the stages that need your own judgement, such as weighing authority or deciding what is material; and anything a court rule, a client commitment or firm policy says AI may not write.

#### Manual (no AI)

No AI at any stage

No AI is involved at any stage. This is always available and is the reference point every tool is compared against.

**Use it for:** Any task, at any time, and required wherever a court rule, a client commitment or firm policy bars any AI use at all. Choosing it needs no justification.

#### Rules that hold whatever the task

- **Affidavits, witness statements and statutory declarations are Red in every forum, and no AI writes any part of them.** They record the deponent’s own words. A scaffold or a structure shapes those words just as a draft does. Where the forum permits it, the tool may flag a gap between the affidavit and the documents, and the lawyer takes any addition from the witness, in the witness’s own words; where the forum bars any AI use, the stage is Manual (no AI).
- **Substantive work that leaves the firm is Yellow at the lowest.** Someone outside the firm will rely on it.
- **Sending, filing or signing is always a separate human act.** Approving a version and releasing it are different decisions, and the second is the one with consequences.

#### What raises a level

Any of these moves the task up one level, at most to Red.

- The tool, model or template has changed materially and the change has not been assessed.
- The person doing the work is a junior without supervision.
- The tool can send, file, or change a system by itself.
- You are unsure. In doubt, raise.

### Stage by stage

A task has seven stages. The level decides what each one requires. A plan takes its headline level from the first-draft stage, and any stage that differs is marked.

| Stage | Green | Yellow | Red | Manual (no AI) |
| --- | --- | --- | --- | --- |
| Frame the task, choose sourcesDecide what you are producing and what you may rely on. Yours at every level. | Lawyer | Lawyer | Lawyer | Lawyer |
| Produce the first draftGet the first version of the document onto the page. This is the stage the headline level is named after. | AI, in a validated template or playbook | AI and lawyer, lawyer directing | Lawyer; AI writes nothing | Lawyer or staff |
| Check high-risk points against the documents*(such as dates, figures, key advice)*Check each one against the document behind it — a court date read by eye, in case the tool misread it. Required at every level, green included. | Each one | Each one | Each one; any AI check or data verified | Each one |
| Check the rest, including omissionsWork from the documents and the instructions, not only from the draft: an omission leaves nothing behind in the draft to check. | Matter-specific content; template tested at setup; end check | Every statement relied on; every document read | Every statement relied on; every document read | Every statement relied on; every document read |
| Judge law, authority, jurisdiction and materialityRight jurisdiction, current law, the authority actually says it, and it matters to this matter. Your own judgement at every level. | Lawyer | Lawyer | Lawyer | Lawyer |
| Approve the exact outputA named person approves this version, for this use. | Named person | Named person, not the drafter | Named person, not the drafter | Named person |
| Release*(send, file, sign)*Sending, filing or signing is a separate human act. | Human | Human | Human | Human |

What gets checked follows what the task risks getting wrong and what the professional requirements demand of it — court and conduct rules, law society guidance, the practice area and what the firm has promised the client — more than it follows the product. So the checks concentrate at the high-risk points, where an error would matter most: dates, figures and key advice. They stay proportionate, because a plan so onerous that the tool saves no time defeats its own purpose, and they change as the guidance changes.

Checking the rest is worked from the documents and the instructions, not only from the draft. An omission leaves nothing behind in the draft to check, so reading the draft alone can never find one.

### Choosing between options

Once you know which options are permitted, they are ranked: **first by whether they save work, then by whether the firm can run them.** Doing it by hand sits in the same list.

Ranking is never a total. Two options are compared on one question at a time, in that order, so a strong answer on one can never make up for a weak answer on another.

The time each plan actually takes a person is not yet costed here, so the order does not account for it. Record your own times on the first few matters.

### Worked through, end to end

Three pages carry one piece of work through all five steps. On each of them the decision and the ranking are computed as the page loads, from the scores above them, so none of the three can show an outcome these rules would not reach; on the two that name real products, the level for every stage is computed as well.

- [A constructed example, start to finish](https://ailawframework.com/how-it-works/example)A chronology in which the firm, the matter, the products and the numbers are all invented, so the method can be shown without anything being assessed.
- [Contract review](https://ailawframework.com/how-it-works/example/contract-review)Doing it by hand scores highest on adoption and still ranks third, because efficiency is asked first and the three scores are never averaged.
- [Medical chronology](https://ailawframework.com/how-it-works/example/medical-chronology)An unconfirmed condition caps governance at 2, and the rule answers “do not use for this task until confirmed” rather than guessing.

The second and third name real products. Their scores are one assessor’s analytical judgements from a master’s research project, formed from each supplier’s public documents at a stated date, and no product was run for either page. Neither is advice, and neither is a product recommendation.

### When to do this again

Any of these lapses the checks and sends the task back to planning.

- The model changes.
- A subprocessor changes.
- The supplier’s terms change.
- The tool gains or loses a function.
- The template changes.
- The task changes.
- The law changes.
- Guidance changes.
- The reviewer changes.

### Words used here

- **Task**: One piece of legal work with a deliverable. The unit everything here is decided about. Not a matter, and not a tool.
- **Configuration**: The exact product, plan and settings you would actually use — not the vendor’s name in general.
- **Pillar**: One of the three separate questions asked about a configuration doing a task. They are never added up.
- **Anchor**: The sentence that describes what a score of 1 to 5 means. You pick the sentence your evidence supports; the number follows.
- **Ceiling**: A limit your evidence places on the governance mark, however strong your own judgement is.
- **Level**: How much of the work a tool may do on a task: Green, Yellow, Red or Manual (no AI).
- **Stage**: One of the seven parts of doing a task, from framing it to releasing the output.
- **Phase**: The two halves of the method: Building the toolkit, done once for a task, and Daily use, done on every matter that uses the tool it selected.
- **Order of preference**: The order permitted options are ranked in: efficiency first, then adoption. Manual work is always one of the options.
- **Work record**: The saved record of what was checked, by whom, what changed, and who approved it.

Definitions taken from court practice notes and statutes are in the [glossary](https://ailawframework.com/glossary.md), each with its source.

### What this is, and what it is not

The framework is a way of deciding and a way of recording. It does not certify that a tool is safe, that your work is correct, or that a saving is real. Every judgement on every screen is one a person made and the record says so. The [security and privacy frameworks](https://ailawframework.com/frameworks.md) page says whether the Essential Eight, the Privacy Act, ISO 27001, IRAP and the conduct rules bind a private firm, and where they stop.

[Browse the practice areas](https://ailawframework.com/practice.md) · [See the software catalogue](https://ailawframework.com/software.md) · [A worked example](https://ailawframework.com/how-it-works/example) · [Open work records](https://ailawframework.com/work)