Contract review: three options, and a ranking that decides
The first review of a confidentiality agreement against the firm’s own playbook. Two AI configurations are permitted here and so is doing it by hand, which is what makes the ranking worth watching: the manual route has the best adoption score of the three and still does not come first.
A constructed case, from a master’s research project.
The firm and the matter are constructed for a master’s research project. They are not a real engagement, and no client information, personal data or participant was involved.
The products are real and named. The three scores are the researcher’s analytical judgements, recorded in that research. The governance score rests on each supplier’s public documents, captured 21 September 2026 and re-audited 4 October 2026; the adoption and efficiency scores are judged against stated anchors from the task and the research literature, not from supplier documents. They are one assessor’s judgements, not a rating anyone has agreed to.
No product was run for this page. There is no trial, no result, no timing and no record of use anywhere on it.
Nothing here is legal advice or a product recommendation. A firm must make its own assessment, against its own evidence, for its own task, and reapply it at least yearly, because products and terms change.
The scores, the decisions and the confirmation statuses are that research’s own, and the conditions and the work still required are given in its words where it states them. The decision and the ranking are computed as the page loads, by the same code the product itself uses, so neither can show an outcome that rule would not reach. The stage levels are computed too, from the level rule as the research states it, by a module written for these pages rather than by the working record, which does not yet set a level for each stage. Where a computed decision or a computed level differs from the one the research recorded, both are shown and the difference is named. Nothing the research does not record is attributed to it: where it is silent, this page says so.
Evidence: Supplier pages, court rules, legislation and regulator guidance, read as documents and given a confirmation status. Captured 21 September 2026, re-audited 4 October 2026. Of 35 answers across five products, three were Confirmed, 28 were Public claim only and four Not confirmed. A confirmed component of an otherwise incomplete answer counts as Public claim only. The three complete confirmations concern two default no-training commitments and one certificate’s scope; no supplier confirmed storage, processing location, retention during use or model choice.
Step 1 Define the task
Say exactly what you are producing, for whom, and what it has to be right about.
Contract review
The task
First review of a confidentiality agreement against the firm’s playbook.
The premise
A seven-position playbook: the firm’s own standard clause positions.
The material
A non-confidential specimen agreement.
What is examined
A missing protection, and prompt injection (embedded instructions) as a governance condition.
The candidates
NetDocuments ndMAX in Word; CoCounsel drafting in Word, custom playbook; Manual review against the playbook.
The case is constructed: the agreement is a non-confidential specimen, and no participant, client information or personal data was involved in assessing it. Every option below, including doing the work by hand, is judged against this one description.
Step 2 Assess and decide
Three separate questions, asked of each option and never added up. The scores are the researcher’s recorded judgements; the decision beneath each set is computed from them as this page loads.
NetDocuments ndMAX in Word
Selected, provisionally
A document-system assistant, working in Word. Its Word routine and prompt library already existed.
Adoption
4 of 5
1Access, information or reviewer absent.
2Preparation dominates, or comprehension depends on one person.
3Feasible within existing resources, with manageable setup.
4Good fit; changes announced or detectable, and reassessment assigned.
5Strong fit and sustained capability, with negligible added burden.Doing it by hand sits here
Preferred on adoption, since its Word routine and prompt library already existed.
What this judgement rests on — these inform it, they do not add up to it
1Minimum quality unreachable, or checking impossible.
2Forecast: more total work than the manual route without a matching quality gain.
3Reference point: quality and total work comparable with careful manual work.Doing it by hand sits here
4Forecast: a task-relevant advantage that survives all required checking.
5A strong advantage after all required work, shown by an observed result.
The contract-review tools tied on efficiency: reuse of the playbook outweighs the checking. This selection rests on a low-confidence efficiency estimate.
What this judgement rests on — these inform it, they do not add up to it
Task-level capability evidenceE1
Quality against careful manual workE2
Checkable for support and omissionsE3
Total human workE4
Study against practice evidenceE5
Repeatability and cost of variationE6
Governance
4 of 5
1A mandatory condition confirmed to fail, or stated by the supplier to be impossible.
2An applicable mandatory AI-specific answer is Not confirmed.
3Applicable mandatory conditions met, with at least one answer Public claim only.
4Controls cover the task; applicable mandatory answers Confirmed, or none applies.Doing it by hand sits here
5Prevention, detection, response and recovery integrated with existing frameworks, including a supplier commitment to notify incidents.
The case used specimen material, so no mandatory AI-specific condition arose and every option scored 4 for governance. Controlled prompts cover the use.
What sets the limit — the weakest one decides
Security management and scopeG1
Privacy and cross-border handlingG2
Encryption, access control, supplier chainG3
Retention, deletion, incident responseG4
Training reuseG5
Storage locationG6
Inference locationG7
Confirmation status of G5 to G7G8
Connection to existing frameworksG9
Task-specific required controlsG10
Three separate answers. They are never averaged. The overall rating is the lowest of the three, and it cannot be higher than due diligence, so a strong answer cannot hide a weak one.
Decision, computed here from the three scoresUse with conditions
Decision recorded in the researchUse with conditionsThe same answer.
Every permitted AI option is used with conditions: its human-work plan and the lawyer’s professional responsibility still apply.
The supplier evidence behind the governance score
Default prohibition on training with customer contentPublic claim only
Storage location and Australian optionPublic claim only
Model-processing locationNot confirmed
Retention during usePublic claim only
Human access and service-provider safeguardsPublic claim only
Standards certification or audit evidencePublic claim only
Customer model choicePublic claim only, no choice
Public documents captured 21 September 2026, re-audited 4 October 2026. Confirmed means published binding terms or scoped independent evidence for the exact answer, not the firm’s own signed agreement. Public claim only means the supplier states it, or a binding answer is incomplete. Not confirmed means no answer for the assessed configuration.
The work still required
An AI option always carries its plan’s conditions.
The selection rests on a low-confidence efficiency estimate: mark it provisional, and record times on the first matters.
Prompt injection (embedded instructions) is a governance condition of this case.
The assessment holds only at the date of its evidence; reapply the framework at least yearly, and sooner on a change of model, subprocessor, terms, tool function, template, task, law, guidance or reviewer.
CoCounsel drafting in Word, custom playbook
Permitted, not selected
A legal research tool, drafting in Word against a custom playbook.
Adoption
3 of 5
1Access, information or reviewer absent.
2Preparation dominates, or comprehension depends on one person.
3Feasible within existing resources, with manageable setup.
4Good fit; changes announced or detectable, and reassessment assigned.
5Strong fit and sustained capability, with negligible added burden.Doing it by hand sits here
Feasible within existing resources, with manageable setup: the custom playbook is new, and setting it up lowers adoption. NetDocuments was preferred on adoption because its Word routine and prompt library already existed.
What this judgement rests on — these inform it, they do not add up to it
1Minimum quality unreachable, or checking impossible.
2Forecast: more total work than the manual route without a matching quality gain.
3Reference point: quality and total work comparable with careful manual work.Doing it by hand sits here
4Forecast: a task-relevant advantage that survives all required checking.
5A strong advantage after all required work, shown by an observed result.
The contract-review tools tied on efficiency: reuse of the playbook outweighs the checking under controlled use.
What this judgement rests on — these inform it, they do not add up to it
Task-level capability evidenceE1
Quality against careful manual workE2
Checkable for support and omissionsE3
Total human workE4
Study against practice evidenceE5
Repeatability and cost of variationE6
Governance
4 of 5
1A mandatory condition confirmed to fail, or stated by the supplier to be impossible.
2An applicable mandatory AI-specific answer is Not confirmed.
3Applicable mandatory conditions met, with at least one answer Public claim only.
4Controls cover the task; applicable mandatory answers Confirmed, or none applies.Doing it by hand sits here
5Prevention, detection, response and recovery integrated with existing frameworks, including a supplier commitment to notify incidents.
The case used specimen material, so no mandatory AI-specific condition arose and every option scored 4 for governance.
What sets the limit — the weakest one decides
Security management and scopeG1
Privacy and cross-border handlingG2
Encryption, access control, supplier chainG3
Retention, deletion, incident responseG4
Training reuseG5
Storage locationG6
Inference locationG7
Confirmation status of G5 to G7G8
Connection to existing frameworksG9
Task-specific required controlsG10
Three separate answers. They are never averaged. The overall rating is the lowest of the three, and it cannot be higher than due diligence, so a strong answer cannot hide a weak one.
Decision, computed here from the three scoresUse with conditions
Decision recorded in the researchUse with conditionsThe same answer.
Every permitted AI option is used with conditions: its human-work plan and the lawyer’s professional responsibility still apply.
The supplier evidence behind the governance score
Default prohibition on training with customer contentPublic claim only
Storage location and Australian optionNot confirmed, the Core plan’s Australian claim does not cover this configuration
Model-processing locationNot confirmed, the Core plan’s Australian claim does not cover this configuration
Retention during usePublic claim only
Human access and service-provider safeguardsPublic claim only
Standards certification or audit evidenceConfirmed, certificate scope
Customer model choiceNot confirmed
Public documents captured 21 September 2026, re-audited 4 October 2026. Confirmed means published binding terms or scoped independent evidence for the exact answer, not the firm’s own signed agreement. Public claim only means the supplier states it, or a binding answer is incomplete. Not confirmed means no answer for the assessed configuration.
The work still required
An AI option always carries its plan’s conditions.
Prompt injection (embedded instructions) is a governance condition of this case.
Manual review against the playbook
Permitted; the reference point
The review done without AI, against the firm’s seven-position playbook: the task’s reference point, and a candidate in every comparison.
Adoption
5 of 5
1Access, information or reviewer absent.
2Preparation dominates, or comprehension depends on one person.
3Feasible within existing resources, with manageable setup.
4Good fit; changes announced or detectable, and reassessment assigned.
5Strong fit and sustained capability, with negligible added burden.Doing it by hand sits here
The manual route scores 5 for adoption, since the firm already works this way.
What this judgement rests on — these inform it, they do not add up to it
1Minimum quality unreachable, or checking impossible.
2Forecast: more total work than the manual route without a matching quality gain.
3Reference point: quality and total work comparable with careful manual work.Doing it by hand sits here
4Forecast: a task-relevant advantage that survives all required checking.
5A strong advantage after all required work, shown by an observed result.
It scores 3 for efficiency, the reference point for every tool.
What this judgement rests on — these inform it, they do not add up to it
Task-level capability evidenceE1
Quality against careful manual workE2
Checkable for support and omissionsE3
Total human workE4
Study against practice evidenceE5
Repeatability and cost of variationE6
Governance
4 of 5
1A mandatory condition confirmed to fail, or stated by the supplier to be impossible.
2An applicable mandatory AI-specific answer is Not confirmed.
3Applicable mandatory conditions met, with at least one answer Public claim only.
4Controls cover the task; applicable mandatory answers Confirmed, or none applies.Doing it by hand sits here
5Prevention, detection, response and recovery integrated with existing frameworks, including a supplier commitment to notify incidents.
No mandatory AI-specific condition applies to it, and its safeguards cover the task.
What sets the limit — the weakest one decides
Security management and scopeG1
Privacy and cross-border handlingG2
Encryption, access control, supplier chainG3
Retention, deletion, incident responseG4
Training reuseG5
Storage locationG6
Inference locationG7
Confirmation status of G5 to G7G8
Connection to existing frameworksG9
Task-specific required controlsG10
Three separate answers. They are never averaged. The overall rating is the lowest of the three, and it cannot be higher than due diligence, so a strong answer cannot hide a weak one.
Decision, computed here from the three scoresUse
Decision recorded in the researchUseThe same answer.
The work still required
Read every document and check every statement relied on against the playbook, judge any departure from it, and have a named person approve the exact output.
Step 2b Choosing between what is permitted
Permitted options rank by efficiency, then adoption, which is a barrier to efficiency: of two products that would perform equally well once adopted, the rule prefers the one easier to implement and train people on.
Order
Option
Adoption
Efficiency
Governance
Decision, computed
01
NetDocuments ndMAX in WordPreferred today
4
4
4
Use with conditions
02
CoCounsel drafting in Word, custom playbook
3
4
4
Use with conditions
03
Manual review against the playbook
5
3
4
Use
Doing it by hand is the easiest of the three to run — adoption 5 against 4 and 3 — and still ranks third, because the rule asks efficiency first and never averages the three scores.
The preferred option rests on a low-confidence estimate of what it saves, so the selection is provisional: the firm records what its first matters actually take, and the order is recomputed from that.
What changes if the efficiency estimate is one point out
Nothing here is waiting on a confirmation: no mandatory AI-specific condition arose, so no governance ceiling binds. What the selection does rest on is a low-confidence efficiency estimate. The research reports that each AI selection changed when its efficiency fell from 4 to 3, so here is that move, re-ranked by the same code.
NetDocuments ndMAX at efficiency 3
CoCounsel drafting in Word, custom playbook — Use with conditions (now preferred)
Manual review against the playbook — Use
NetDocuments ndMAX in Word — Use with conditions (was preferred before the move)
The decisions do not move; the selection does. That is what “provisional” means in the record: the firm records its own times on the first matters, and the order is recomputed from them.
Step 3 Plan the human work
Being permitted says nothing about how much of the work a tool may do. That is the level, and it is set stage by stage by the rule, from the facts of the task rather than from the product.
NetDocuments ndMAX in Word
The first pass is yellow: the case record holds no setup test meeting the green capability prerequisite. It could become green only after a recorded capability test and validation of the playbook — a future condition, not a demonstrated result.
Plan
Written
Recorded at green
none
Recorded at yellow
first pass against the playbook; redraft
Recorded at red
interpret terms; accept departures
Stage
Level
What the rule read
What that stage then requires
Frame the task, choose sources
RedDiffers from the headline level
The lawyer’s own judgement: framing, weighing authority, deciding materiality or accepting a departure.This stage asks the same thing at every level, so the level cannot buy anything here.
Lawyer
Produce the first draft
Yellow
Nothing in the task puts this stage below yellow.The first pass runs against the firm’s seven-position playbook, its own standard clause positions. Green also asks for task-level capability evidence, and the research records none for this task: the efficiency score of 4 is a forecast under its own anchor, not an observed result, so it cannot serve as that evidence.
AI and lawyer, lawyer directing
Check high-risk points against the documents (such as dates, figures, key advice)
Yellow
Substantive work leaving the firm is at least yellow.What is being checked is the agreement the firm will advise on, which leaves the firm.
Each one
Check the rest, including omissions
Yellow
Substantive work leaving the firm is at least yellow.The missing protection the case examines can only be found by working from the playbook and the agreement, not from the draft.
Every statement relied on; every document read
Judge law, authority, jurisdiction and materiality
RedDiffers from the headline level
The lawyer’s own judgement: framing, weighing authority, deciding materiality or accepting a departure.This stage asks the same thing at every level, so the level cannot buy anything here.
Lawyer
Approve the exact output
Yellow
Substantive work leaving the firm is at least yellow.The output leaves the firm, so the approver is not the drafter.
Named person, not the drafter
Release (send, file, sign)
Yellow
Substantive work leaving the firm is at least yellow.What is released is the output that leaves the firm.This stage asks the same thing at every level, so the level cannot buy anything here.
Human
The headline level is Yellow, taken from the first-draft stage, and every stage that differs from it is marked. This plan uses 2 of the four levels.
The recorded level, set beside the computed one
Produce the first draft: the research records Yellow, the rule computes YellowYellow
A match is not independent support, and nothing here treats it as any. The facts the rule reads were taken from the same description of the task the research worked from, so the two are not independent of each other; and where the research marks a plan “Set by rule”, the level it recorded is this same rule’s output, so a match there would only repeat it. The research itself reports the rule reproducing its written levels as consistency, not validity. Only the first-draft stage is compared, because that is the one stage the research states a level for in its own words.
CoCounsel drafting in Word, custom playbook
The first pass is yellow: the case record holds no setup test meeting the green capability prerequisite. It could become green only after a recorded capability test and validation of the playbook — a future condition, not a demonstrated result.
Plan
Set by rule
Recorded at green
none
Recorded at yellow
first pass against the playbook; redraft
Recorded at red
interpret terms; accept departures
On the same task the rule reads the same facts, and it reaches the same levels as NetDocuments ndMAX in Word, stage for stage:
Frame the task, choose sourcesRed
Produce the first draftYellow
Check high-risk points against the documentsYellow
Check the rest, including omissionsYellow
Judge law, authority, jurisdiction and materialityRed
Approve the exact outputYellow
ReleaseYellow
The headline level is Yellow, taken from the first-draft stage, and every stage that differs from it is marked. This plan uses 2 of the four levels.
The recorded level, set beside the computed one
Produce the first draft: the research records Yellow, the rule computes YellowYellow
A match is not independent support, and nothing here treats it as any. The facts the rule reads were taken from the same description of the task the research worked from, so the two are not independent of each other; and where the research marks a plan “Set by rule”, the level it recorded is this same rule’s output, so a match there would only repeat it. The research itself reports the rule reproducing its written levels as consistency, not validity. Only the first-draft stage is compared, because that is the one stage the research states a level for in its own words.
Manual review against the playbook
The manual route has no level to choose: the rule gives it Manual (no AI) at every stage. It is the reference point every tool is compared against.
Stage
Level
What the rule read
What that stage then requires
Frame the task, choose sources
Manual (no AI)
The manual route: no AI at any stage.This stage asks the same thing at every level, so the level cannot buy anything here.
Lawyer
Produce the first draft
Manual (no AI)
The manual route: no AI at any stage.
Lawyer or staff
Check high-risk points against the documents (such as dates, figures, key advice)
Manual (no AI)
The manual route: no AI at any stage.
Each one
Check the rest, including omissions
Manual (no AI)
The manual route: no AI at any stage.
Every statement relied on; every document read
Judge law, authority, jurisdiction and materiality
Manual (no AI)
The manual route: no AI at any stage.This stage asks the same thing at every level, so the level cannot buy anything here.
Lawyer
Approve the exact output
Manual (no AI)
The manual route: no AI at any stage.
Named person
Release (send, file, sign)
Manual (no AI)
The manual route: no AI at any stage.This stage asks the same thing at every level, so the level cannot buy anything here.
Human
The headline level is Manual (no AI), taken from the first-draft stage, and every stage that differs from it is marked. This plan uses 1 of the four levels.
Steps 4 & 5 What the plan requires on every matter
These are the requirements the framework attaches to a matter that uses a permitted option. They are obligations, not events.
A person with the documents makes the plan’s checks, above all at the high-risk points: dates, figures, key advice.
Approval is refused unless no client commitment excludes this use, and unless the level and the reason for it are recorded.
At yellow and red the checks are counted and the approver is not the drafter; in a sole practice, counted checks suffice.
At red the lawyer confirms writing the output.
A named, competent approver approves this version, for its stated use, and a supervising lawyer approves a junior’s work.
Release — sending, filing or signing — is a separate human act.
The record keeps who drafted, who checked what, how many items and for how long, what changed and who approved, with the tool, the model, the date and any required AI-use statement.
At least yearly, and on a change of model, subprocessor, terms, tool function, template, task, law, guidance or reviewer, the assessment goes back to the start and the earlier checks lapse.
Nothing on this page is a record of use.
The list above is what the plan would require on a matter. No product was run for this page: no document was read by one, nothing was checked, nothing was corrected and nothing was timed. The research traced its plans through the constructed cases and reports that a planned check would catch each defect written into them; it does not report that a lawyer found one, and neither does this page. The plans are not costed in hours.
Reading this back
Doing it by hand scores highest on adoption and still ranks third, because efficiency is asked first and the three scores are never averaged.
Two limits travel with every number on this page. The scores are one assessor’s judgements, formed from public supplier documents captured 21 September 2026, re-audited 4 October 2026; a firm holds different evidence, and the supplier’s own signed agreement is better evidence than any of it. And an assessment holds only at the date of its evidence: products, plans, models and terms change, which is why the framework sends a task back to the start at least yearly.
Nothing here is advice, a recommendation or a rating any supplier has agreed to. A firm applying the framework to this task would reach its own three scores, and may well reach different ones.
