Banking AI agent · Legal
Contract-Review & Legal-Workflow AI Agent
Extract the clauses that matter from a negotiated agreement, compare each against your playbook and its fallbacks, and register the obligations — with the position, and any deviation, left to a named lawyer.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Legal position
- Named lawyer
- After launch
- Agent Care
What this agent does
Marks the deviations, settles none of them
In
- 01
Take the agreement as it stands — the executed document, its amendments, side letters and any restatement.
- 02
Read the bank's playbook: the preferred position, the named fallbacks, and where a term has to go to counsel.
Reason
- 03
Extract the clauses in scope and quote the text, with the definitions that give each term its meaning.
- 04
Compare each clause against the playbook position and the fallbacks the bank's lawyers have already approved.
- 05
Assemble the obligations, conditions and dates into a register, each one tied to the clause it came from.
Decide
- 06
Mark what is missing, non-standard or open to two readings, and keep that apart from what it could not find.
- 07
Route the term to the lawyer who owns it, with the clause, the playbook rule and the distance between them.
Out
- 08
Hand counsel a worked file — the clauses, the quoted text, the marked deviations and the register behind them.
- 09
Retain the version compared, the playbook rule applied, the lawyer's position and the reasons they gave.
Product statement
The agent extracts, compares and records. Nothing it produces is a legal conclusion or legal advice; accepting a deviation, answering the counterparty, negotiating, executing and judging enforceability stay with a named lawyer.
Example workflow
One agreement, end to end
- 1Agreement receivedThe document as executed or as drafted, with its amendments, side letters and any restatement
- 2Set and version resolvedWhich document governs today, and which defined terms carry across the rest of the set
- 3Clauses extractedThe clauses in scope, quoted to the clause number, read with the definitions that bind them
- 4Playbook comparedEach clause against the preferred position and the named fallbacks, with the gaps marked
No human action required
Stages 1 to 4 run before a lawyer opens the file — the set, the clauses and the playbook are compared first. Nothing in that stretch advises, accepts a deviation, answers the counterparty or signs anything.
Reaches counsel as a worked file.
Reaches counsel marked unresolved, not as a deviation.
Legal counsel on the matter
Reads the worked file, takes the position on each deviation, and owns the advice, the mark-up and the signature.
Adopt · Overrule · Escalate
Positions taken — handed back
- 6Counsel decides, agent recordsOnly the position a named lawyer took, with the reasons given, written where access and policy allow
- 7Outcome evaluatedReadings counsel overruled, clauses counsel added, register dates corrected and business outcome
Overruled readings
A reading counsel overruled is counted as a failure.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Any legal conclusion or advice on the agreement.
- Accepting a deviation from the playbook.
- Accepting a counterparty's drafting or position.
- Negotiating, or sending a mark-up or a reply.
Automation boundary
Agent acts unaided
- ✓Extract the clauses in scope and quote the text they come from.
- ✓Compare each against the bank's playbook and fallbacks.
- ✓Mark what is missing, non-standard or open to two readings, and mark what it cannot evidence.
- ✓Register the obligations, conditions and dates.
Write actions run only inside the approval boundaries agreed during implementation. A counterparty mark-up is not one.
- Signing, executing or releasing any document.
- Deciding whether a term is enforceable.
- Waiving or varying an approved fallback position.
- Changing the playbook, templates or access rules.
Example output
One agreement, annotated
Everything the agent marks is attached to the clause and the version it was read from.
Review output · single agreement
Illustrative example- Agreement
- Vendor outsourcing agreement
- Set
- Two amendments, one side letter
- Playbook
- Vendor, v4
- Deviations marked
- Four, quoted to clause
- Confidence
- 84%
- Position
- None taken by the agent
As received
The documents as executed and the playbook as approved — the agent neither drafts them nor re-versions them.
Evidence used
- Liability cap under fallback
- No step-in right located
- Side letter redefines a term
Why four, not six
Two more were dropped — a playbook preference is not a legal requirement.
What the score decides
Where the file sits in counsel's queue — not whether a deviation is acceptable.
Action
AdoptOverruleEscalateValue
Where AI adds value
The same four claims, placed at the point in the workflow where each one applies.
Where the value lands
Value 01 – 04
Every agreement in scope
As executed and amended
Compare to your own playbook
Read each clause against the preferred position and the fallbacks your lawyers have already approved, not against a general market standard.
Reach counsel with the file worked
Pulling clauses out of a set and quoting them is slow and mechanical, so the quotes, the marked deviations and the register arrive with the agreement rather than after it.
Separate a deviation from a preference
A term the playbook takes no position on, a clause open two ways, a definition that changes across the set — written down as unresolved instead of raised as a deviation.
Build an evidence trail
Retain the version compared, the clauses quoted, the playbook rule applied, the lawyer's position and the reasons given — on both paths.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Contract lifecycle
- Icertis
- Ironclad
- SirionLabs
- Agiloft
- CLM exports
Document & knowledge stores
- iManage
- NetDocuments
- SharePoint
- Microsoft 365
- Signed-document archives
Legal workflow
- Legal Tracker
- SimpleLegal
- Jira
- ServiceNow
- e-signature platforms
Agent
Contract review & legal workflow
- Extracts the clauses
- Compares the playbook
- Routes to counsel
Banking & counterparty records
- Loan IQ
- ISDA and CSA records
- Vendor registers
- Counterparty masters
Observability & evaluation
- OpenTelemetry
- Langfuse
- Supported monitoring/evaluation sources
Integration availability depends on the client's existing systems and API access.
Agent controls
Six layers between a clause and a position
Each control wraps the one inside it. The agent takes no legal position — it works the file a lawyer will take one on — and no reading leaves these layers as advice.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Decide whether a reading may stand
- L3
- Decides who may accept a deviation
- L4 – L5
- Hold the file's limits, keep the record
- L6
- Pulls automation back when signals degrade
How Nestack evaluates it
Evaluate the clause it did not find — not only the ones it did.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the marked file counsel opens
- E1Clause-extraction evaluationWas the clause found, and is the quoted text the text in the document?
- E2Set and version evaluationDid it read the document in force, with its side letters?
- E3Playbook comparison evaluationWas the clause matched to the right rule and the right fallback?
- E4Absence calibrationWhen it reported nothing, was there nothing there to find?
- E5Slice evaluationHow does performance change across specific agreement cohorts?
- E6Business outcomeReadings overruled, clauses counsel added, register dates corrected.
Floor — the non-standard term that reached a lawyer
Failure modes
Where each failure originates in the agent
Seven failure modes plotted against the five stages of the agent lifecycle. None of them is a wrong legal answer, because the agent gives none — they are wrong readings of documents, reaching a lawyer as though they were right.
Agent lifecycle
Direction of processing →
01 · Assembly
2 modes
- LW-01
Superseded version read as current
The original is quoted after an amendment or restatement moved the term.
- LW-02
Document never in the set
A side letter nobody filed varies terms the agent reads as settled.
Stage gathers
The agreement, its amendments and side letters
02 · Extraction
2 modes
- LW-03
Clause read outside its definitions
A term defined elsewhere in the set is read at its ordinary meaning instead.
- LW-04
Not found reported as not there
A clause it missed and a clause the agreement never had come back the same way.
Stage extracts
Clauses, the text quoted and the terms defined
03 · Comparison
1 mode
- LW-05
Playbook position stated as law
A preferred position is written up as though a rule required it.
Stage compares
Each clause against the playbook and fallbacks
04 · Handover
1 mode
- LW-06
Obligation dated from the wrong trigger
A notice period is registered from execution, not the event that starts it.
Stage hands
The marked file counsel opens, dates registered
05 · Change / Version
1 mode
- LW-07
Playbook revised, files not re-compared
Fallbacks move and agreements already marked are not read again.
Stage tracks
Playbook, template, model and prompt changes
- Sev-1 · a deviation never reaches a lawyer
- Sev-2 · the file misstates the documents
- Sev-3 · the set is partial, the file goes back
Affected slices
The agreements drafted on somebody else's paper
An NDA cut from the bank's own template has one version and a playbook written for it. Failures gather where the drafting came from the other side and a defined term travels across a set. Nestack reports by slice, not only in total.
Slice performance — reported separately, not only in aggregate
Illustrative example| Slice | Failure rate | Lift | Lift vs. threshold | Status |
|---|---|---|---|---|
| Counterparty-paper agreements | 5.1% | 3.5× | Review | |
| ISDA and CSA schedules | 3.9% | 2.7× | Review | |
| Scanned and side-lettered files | 2.8% | 1.9× | Watch | |
| Bank-template NDAs | 1.2% | 0.8× | Normal |
Bar: mis-marked-clause lift vs. bank-template NDA baseline · scale 0–4.0× · tick marks the 2.0× review threshold
2 of 4 slices over threshold
Evidence-linked improvement
The signed version tells you what was missed
A deviation the agent never marked surfaces in the executed document — the term counsel negotiated is not the one it read. That comparison feeds the loop.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Overruled readings rise in one agreement cohort, or a clause counsel added was never on the list.
- 02
Diagnose
Traced back through the set — which version was read, which definition applied, which playbook rule was matched.
- 03
Improve
The extraction pattern or the playbook mapping is re-approved by the general counsel's office and version-linked.
- 04
Verify
Re-run over agreements already executed, including the ones counsel overruled a reading on.
- 05
Learn
The wording that carried two readings is written into the playbook, beside the fallback it sat against.
Learn → Detect
The return edge. A playbook is revised while agreements are still in negotiation, so each cycle re-compares files the last one had already handed to counsel.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, the document set and the playbook, extraction and comparison, evaluation, counsel handover, then validation and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Agreement scoping and legal-boundary definition.Week 1
- 02Document-store and CLM assessment.Week 1, Week 2
- 03Playbook, fallback and escalation mapping.Week 1, Week 2
- 04Version, amendment and side-letter resolution.Week 2
- 05Clause extraction and quotation standard.Week 3
- 06Defined-term resolution across the set.Week 3
- 07Playbook comparison and deviation marking.Week 3, Week 4
- 08Obligation, condition and date register.Week 4
- 09Privilege, access and confidentiality controls.Week 4
- 10Absence-calibration and slice evaluation suite.Week 4, Week 5
- 11Counsel handover and document-system integration.Week 5, Week 6
- 12Observability, deployment and Agent Care handover.Week 6
12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallel
Final scope and sequence confirmed in discovery
Engagement tiers
What each tier includes
Rows are the capabilities named in each tier's scope. Higher tiers include everything below them. No tier moves the legal position off a named lawyer.
| Capability✓ in scope · — not at this tier | PilotOne agreement type, one playbook | ProductionProduction workflow integration | AdvancedMulti-entity / multi-jurisdiction |
|---|---|---|---|
| Introduced at Pilot | |||
| Clause extraction quoted to the document | ✓ | ✓ | ✓ |
| Playbook comparison with deviations and gaps marked | ✓ | ✓ | ✓ |
| Obligation, condition and date register | ✓ | ✓ | ✓ |
| Named-counsel gate on deviations | ✓ | ✓ | ✓ |
| Privilege and access controls on the file | ✓ | ✓ | ✓ |
| Audit trail and baseline evaluation | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Additional agreement types and playbooks | — | ✓ | ✓ |
| CLM, document-store and workflow integration | — | ✓ | ✓ |
| Observability and evaluation | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multi-entity and multi-jurisdiction playbooks | — | — | ✓ |
| High volume and portfolio re-papering | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on the agreement types and playbooks in scope, document-store and CLM integrations, document volume, access controls and deployment requirements.
- Separate from build
- Build pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.
What we need from you
What you bring, and what we build with it
Each input maps to a piece of build scope and a week in the delivery timeline.
You bring
We build with it
01A set of agreements and the mark-ups your lawyers returned
→Agreement scoping and legal-boundary definition
Week 102Your playbook — preferred position, fallbacks and who owns each
→Playbook, fallback and escalation mapping
Week 103Access to the document store, the CLM and the signed archive
→Document-store and CLM assessment, then integration
Week 204The amendments, side letters and restatements behind each file
→Version, amendment and side-letter resolution
Week 205The clauses you actually care about, and how you quote them
→Clause extraction, quotation standard and defined-term resolution
Week 306Your rules on privilege, access and what a matter file may hold
→Privilege, access and confidentiality controls
Week 407Named lawyers, and who may accept a deviation
→Counsel handover and review workflow, then supervised queues
Weeks 5–6
Nothing else is required
Deployment, documentation and Agent Care handover are ours.
Delivery timeline
Four phases across six weeks
Phases are drawn over the weeks they actually occupy. Week 5 replays agreements your lawyers have already marked up, so nothing live waits.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1The agreements in scope, the playbook and who owns each
W2Document-store and CLM access, then version resolution
W3Clause extraction, defined terms and playbook comparison
W4Obligations register, privilege controls and evaluation
W5Agreements replayed against mark-ups you already returned
W6Lawyers work the live queue under review, then handover
- Reading the band
- Version and side-letter resolution lands before extraction starts, because a clause quoted from a superseded document invalidates every comparison behind it. The bars show that order, not a smooth ramp.
- At the end of W6
- Lawyers have worked live agreements from the agent's file, the readings counsel overruled have been counted, and Agent Care takes over monitoring.
- Duration
- Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · Banking AI agent
Build a review agent around the playbook your lawyers already apply.
Show us a set of agreements, the playbook you compare them to and the mark-ups your lawyers returned. We'll re-read that set against your own positions and mark two things on each file — the clauses counsel added that the agent had not marked, and the readings a lawyer overruled.
- Typical delivery
- 4–6 weeks
- Deployment
- Your legal stack
- Legal position
- Named lawyer on the matter
- After launch
- Agent Care available
Nestack Agents · Contract review & legal workflow
AGT-BK-12 · Agent Care available after launch