Use case · Contract review
Read the supplier's paper against your own rules before a lawyer opens it.
The agent reads inbound NDAs, MSAs and supplier agreements clause by clause, compares each one with your negotiation playbook, and marks where the drafting departs from your positions. It proposes your approved fallback wording and sends to counsel only the points that genuinely need legal judgement.
For procurement directors and legal operations leads at organisations that routinely sign supplier agreements on the counterparty's paper.
Clause by clause
A supplier MSA, read against your playbook.
Fictional clauses from a supplier-drafted services agreement, with the agent's assessment against a typical buyer-side playbook.
8.1 The Supplier's total aggregate liability arising under or in connection with this Agreement shall not exceed the Charges paid in the three (3) months preceding the claim.8.3 Nothing in this Agreement limits either party's liability for death or personal injury caused by negligence, or for fraud.9.1 The Customer shall indemnify the Supplier against all losses arising from the Customer's use of the Services.14.2 This Agreement shall renew automatically for successive twelve-month terms unless either party gives notice no later than ninety (90) days before renewal.16.4 The Supplier may engage sub-processors without prior notice to the Customer.16.6 Personal data may be transferred outside the UK where the Supplier considers it appropriate.21.1 This Agreement is governed by the laws of England and Wales, and the courts of England have exclusive jurisdiction.After the upload
Three passes over one agreement.
The agent structures the document first, then compares, then drafts. Each pass leaves a record a lawyer can inspect.
Conceptual flow, not a live system or measured result.
Structure the document
A draft arrives through the procurement intake form, a shared mailbox or your CLM. The agent converts Word or PDF while keeping clause numbering, existing tracked changes and schedules. It identifies the agreement type and counterparty, splits the text into clauses, and resolves defined terms and cross-references, so a phrase such as "subject to Schedule 3" is followed to the schedule and the two are read together rather than in isolation.
Compare with the playbook
Each clause is mapped to a playbook topic such as liability, indemnity, termination, data protection or governing law. The agent compares the wording with your preferred, acceptable and fallback positions and your walk-away point. Every finding quotes the clause and cites the playbook rule it relies on. Topics the playbook requires but the draft omits, such as audit rights or a data processing schedule, are reported as absences, not silently passed.
Redline and route
Where a deviation can be fixed with approved fallback wording, the agent produces a tracked-change redline in Word with a comment explaining the reason. Deviations beyond the fallback, or on protected topics such as indemnity, liability and IP ownership, go to a named lawyer with a one-page summary. Agreed contracts move to signature through your e-signature tool, and each outcome is recorded against its playbook topic so recurring deviations can be reviewed.
Three ways to review
Where each approach earns its place.
| Criterion | Manual review | CLM rules & templates | Playbook-driven agent |
|---|---|---|---|
| Turnaround on supplier paper | Depends on queue length and lawyer availability | Fast on your own templates; limited on third-party drafts | First-pass review soon after upload, including on supplier drafts |
| Consistency between reviewers | Varies with experience and workload | High, for the clauses the rules anticipate | Same playbook every time, with the rule cited for each finding |
| Unusual or non-standard drafting | Strongest: an experienced lawyer reads intent, not just words | Weak: reworded clauses often pass keyword and template checks | Recognises reworded clauses well; unfamiliar structures still go to a lawyer |
| Explainability | Reasoning sits in the reviewer's notes or email | Rule matches are clear but shallow | Each flag links the clause text to the position it breaches |
| Cost to maintain | Rises with contract volume | Rules and templates need upkeep as positions change | Playbook upkeep plus re-running the evaluation set after changes |
| Negotiation and relationship judgement | Belongs here | Not addressed | Not addressed: the agent prepares, people negotiate |
| Best suited to | High-value, bespoke or contentious agreements | High volumes on your own paper | Recurring supplier paper under a stable playbook |
Failure modes
Five ways contract review goes wrong, and the fix for each.
Hallucinated clause: the agent reports wording that is not in the document, or misquotes it.
Every finding quotes the source text and clause number. Findings whose quote cannot be found verbatim in the document are discarded before anyone sees them.
Missed cross-reference: a clause looks acceptable on its own but is changed by a definition, schedule or order-of-precedence clause elsewhere.
Definitions and cross-references are resolved first, order of precedence is a mandatory topic, and the evaluation set includes contracts where the problem sits in a schedule.
Confidentiality: counterparty drafts and commercial terms are processed somewhere they should not be.
Processing runs in your tenancy or an approved enterprise model deployment that excludes training on your data. Access mirrors your matter permissions.
Over-reliance: reviewers start accepting the agent's redlines without reading them.
Protected topics always need a named person's decision, a senior lawyer samples approved contracts regularly, and the review screen lists what the agent did not check.
Playbook drift: negotiating positions change but the rules the agent applies do not.
The playbook is versioned with an owner, each review records the version used, and changes trigger a re-run of the evaluation set.
Phased adoption
From shadow reviews to a live queue.
Each phase ends on evidence, not on a date.
- Phase 1
Codify the playbook
Turn positions, fallbacks and escalation rules for one agreement type, usually NDAs, into a structured, owned playbook.
Move on whenLegal signs off the playbook and a labelled evaluation set of past contracts.
- Phase 2
Shadow review
Run the agent alongside lawyers on live NDAs without its output reaching any counterparty.
Move on whenEvery agent–lawyer disagreement is reviewed and legal accepts the remaining error types.
- Phase 3
Assisted review
Reviewers work from the agent's redlines; protected topics still route to counsel.
Move on whenEdit rates and escalation reasons stay stable over an agreed period, and reviewers confirm it saves effort.
- Phase 4
Extend by agreement type
Add MSAs, then supplier and data processing agreements, each with its own playbook section and evaluation set.
Move on whenEach new type passes the same shadow and assisted stages before it goes live.
Where we would say no
Cases where an agent is the wrong tool.
- You have no written playbook and positions depend on who happens to review. The agent needs rules to apply; agreeing them comes first, and that is legal work rather than software.
- Nearly all of your agreements are on your own templates with little negotiation. Template controls and a clause library in your CLM will do the job for less.
- Your contracts are few, high-value and heavily bespoke, such as joint ventures or major outsourcing deals. Experienced counsel should read those end to end.
- You want contracts approved with no person involved. Sign-off stays with someone who holds delegated authority.
Questions from legal and procurement
Points buyers raise before a pilot.
Does contract text leave our environment?
It does not have to. The agent can run in your own cloud tenancy, calling a model through an enterprise deployment in the region you specify. Documents are read from and written back to your existing repositories, with access following your current matter permissions.
Is the model trained on our contracts?
Under the enterprise terms we deploy with, your inputs are not used to train the underlying model. During scoping we take your security and privacy teams through the provider terms and data flows, so the position is confirmed in writing, not assumed.
How accurate is it?
We do not quote a single figure; it depends on your playbook and counterparties. We measure the agent against a labelled set of your past contracts, report results per playbook topic, and treat missed deviations on protected topics as far worse than false flags.
Can it work inside Word and our CLM?
Redlines are produced as tracked changes with comments in Word, so lawyers edit in the tool they already use. We integrate with common CLM and repository platforms for intake and filing, and with e-signature tools for agreements that clear review.
What drives the cost?
The number of agreement types in scope, how detailed the playbook is, how deeply we integrate with your CLM and intake process, and the effort to build and maintain evaluation sets. Page volume affects model usage costs but is rarely the largest element of the total.
Will our lawyers accept it?
Adoption goes best when lawyers own the playbook and see the shadow-phase results themselves. The agent does the repetitive first read; negotiation and judgement stay with them. Where they overrule it, those overrides are logged and reviewed, and they drive the next round of playbook changes.
Go deeper
The pieces behind this workflow.
What would you like to move forward?
Start with the outcome that feels closest. We'll help give the next step a useful shape.
One conversation. A clearer direction.
Tell us what you want to improve. A few sentences are enough to start.
shayan@coldai.orgLast reviewed 25 September 2026.