Banking AI agent · Trade finance
Trade-Finance Document-Checking AI Agent
Examine a presentation against the credit's terms and the rules it names, cite each discrepancy to the document and field it came from — with compliance, refusal and waiver left to a named checker.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Determination
- Named checker
- After launch
- Agent Care
What this agent does
Lists the discrepancies, determines nothing
In
- 01
Take the presentation as it arrived — draft, invoice, transport, insurance and certificates — with the credit behind it.
- 02
Read the credit as amended, the amendments in force, and the rule set the credit says it is subject to.
Reason
- 03
Compare each document on its face against the terms, and against the practice the rules set for that document.
- 04
Cross-read the set for data that conflicts — description, quantity, marks, dates, ports, parties and amounts.
- 05
Count the dates: shipment, expiry, the presentation period, and the day the examination period started.
Decide
- 06
Cite each finding to the document, the field and the term or practice behind it, and mark what it cannot evidence.
- 07
Route sanctions, dual-use and financial-crime questions to the teams that own them, not onto the checker's list.
Out
- 08
Hand the checker a worked presentation — the findings, the evidence, the days left and what is still open.
- 09
Retain the presentation, the terms compared, the findings, the checker's decision and the reasons they gave.
Product statement
The agent examines and evidences. Determining compliance, accepting or refusing a presentation, issuing a refusal notice, waiving a discrepancy, amending a credit and paying stay with named bank staff.
Example workflow
One presentation, end to end
- 1Presentation receivedThe documents as presented, under a credit, its amendments and the rule set it names
- 2Terms assembledThe credit as amended, the documents it calls for, and the practice the rules set for each one
- 3Documents examinedEach document read on its face against the terms, with the dates, expiry and periods counted
- 4Set cross-readDescription, quantity, marks, dates, ports, parties and amounts compared across the documents
No human action required
Stages 1 to 4 run before a checker opens the file — the terms, the documents and the dates are compared first. Nothing in that stretch determines compliance, refuses a presentation, waives anything or starts a notice.
Reaches the checker at the top of the queue.
Reaches the checker marked unresolved, not as a finding.
Document checker or trade-finance officer
Reads the worked presentation, determines whether the documents comply, and owns the notice, the waiver request and the payment.
Confirm · Strike out · Escalate
Findings settled — handed back
- 6Checker decides, agent recordsOnly what a named checker determined, with the reasons given, written where access and policy allow
- 7Outcome evaluatedFindings the checker struck out, discrepancies raised after the notice, and notices sent inside the period
Struck-out findings
A finding the checker struck out is counted as a failure.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Determining that documents comply.
- Accepting or refusing a presentation.
- Issuing or wording a refusal notice.
- Waiving a discrepancy or approaching the applicant.
Automation boundary
Agent acts unaided
- ✓Compare each document on its face against the credit and the amendments in the set.
- ✓Cross-read the set for data that conflicts.
- ✓Cite the document, the field and the term behind each finding, and mark what it cannot evidence.
- ✓Count the days left in the examination period.
Write actions run only inside the approval boundaries agreed during implementation. A refusal notice is not one of them.
- Paying, accepting a draft or deferring payment.
- Amending a credit or agreeing an extension.
- Answering a sanctions or dual-use question.
- Changing rule sets, checklists or thresholds.
Example output
One presentation, annotated
Everything the agent raises is attached to the document and the field it came from.
Examination output · single presentation
Illustrative example- Credit
- Sight, under UCP 600
- Amendments
- Two in force at presentation
- Presented
- Day 12 of 21
- Findings cited
- Three, to field and term
- Confidence
- 81%
- Determination
- None made by the agent
As received
The documents as presented and the credit as amended — nothing here is inferred or verified as genuine.
Evidence used
- Cover dated after shipment
- Marks differ across the set
- Certificate bears no issuer
Why only three
A fourth reading was dropped — nothing in the credit or the rules carried it.
What the score decides
Where the presentation sits in the checker's queue — not whether the documents comply.
Action
ConfirmStrike outEscalateValue
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 presentation received
Under the credits and rule sets in scope
Check against the credit as amended
Compare each document on its face against the terms and the practice the rules set, then cross-read the set for data that conflicts.
Reach the checker with days left
Examination runs to a defined limit, so the findings, the evidence behind them and the days remaining arrive with the presentation rather than after it.
Separate a finding from a suspicion
A reading the credit does not carry, a term open two ways, an original nobody can confirm — written down as unevidenced instead of raised as a discrepancy.
Build an evidence trail
Retain the credit version compared, the documents examined, the findings and their citations, the checker's decision 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.
Trade & core systems
- Finastra Trade Innovation
- Surecomp
- CGI Trade360
- Temenos
Messaging & presentation
- Swift MT 7xx
- ISO 20022
- eUCP portals
- Bolero
- essDOCS
Documents & capture
- Scanned presentations
- OCR output
- Document images
- courier packs
Agent
Trade-finance document checking
- Compares the terms
- Cites each finding
- Routes what it must
Screening & financial crime
- Sanctions screening
- vessel and port data
- TBML monitoring
- case management
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 finding and a refusal
Each control wraps the one inside it. The agent does not determine compliance — it works the presentation a checker will determine on — and no finding leaves these layers as a refusal.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Decide whether a finding may stand
- L3
- Decides who may determine and refuse
- L4 – L5
- Send questions out, keep the record
- L6
- Pulls automation back when signals degrade
How Nestack evaluates it
Evaluate the findings a checker struck out — not only the ones raised.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the list of findings the checker opens
- E1Finding-quality evaluationDid each finding rest on a term in the credit or a practice in the rules?
- E2Coverage evaluationWhich discrepancies raised later were absent from the list?
- E3Retrieval evaluationDid it read the credit as amended, and every document in the set?
- E4Date and period evaluationWere shipment, expiry and the examination days counted correctly?
- E5Slice evaluationHow does performance change across specific presentation cohorts?
- E6Business outcomeFindings struck out, discrepancies raised late, notices sent inside the period.
Floor — the discrepancy not missed, and the one not invented
Failure modes
Where each failure originates in the agent
Seven failure modes plotted against the five stages of the agent lifecycle. The documents arrive as they were presented — the agent examines what is in front of it, so nothing here claims a forgery was caught.
Agent lifecycle
Direction of processing →
01 · Intake
2 modes
- TF-01
Amendment never reached the comparison
The credit is compared as issued, after a term was amended.
- TF-02
Pages missing from the capture
A document nobody scanned is treated as one the credit did not call for.
Stage gathers
The presentation, the credit and its amendments
02 · Examination
2 modes
- TF-03
A finding the rules do not carry
Wording that differs without conflicting is written up as a discrepancy.
- TF-04
Original taken on trust
A copy stands in for the original the credit called for, unsigned or unmarked.
Stage compares
Each document on its face against the terms
03 · Cross-read
1 mode
- TF-05
Conflict across the set not seen
Marks, quantity or the vessel differ between documents and nothing flags it.
Stage tests
The set for data that conflicts across it
04 · Handover
1 mode
- TF-06
Clock counted from the wrong day
The examination period runs from a day the rules do not start it on.
Stage hands
A worked presentation, with the days counted
05 · Change / Version
1 mode
- TF-07
Practice revised, findings not re-tested
A rule set or checklist changes and nobody re-measures what it raises.
Stage tracks
Rule sets, checklists and prompt changes
- Sev-1 · a real discrepancy is never raised
- Sev-2 · a refusal the bank cannot stand behind
- Sev-3 · the checker works from a partial list
Affected slices
The credit nobody reads as amended
A sight credit with one shipment and a standard set of documents is close to a checklist — one set of terms, one comparison. Failures gather where a term was changed after the credit was issued. 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 |
|---|---|---|---|---|
| Credits amended more than once | 4.8% | 3.3× | Review | |
| Charter-party and transhipment bills | 3.2% | 2.2× | Review | |
| Handwritten and stamped certificates | 2.7% | 1.9× | Watch | |
| Sight credits, one shipment | 1.2% | 0.8× | Normal |
Bar: mis-checked-presentation lift vs. sight-credit baseline · scale 0–4.0× · tick marks the 2.0× review threshold
2 of 4 slices over threshold
Evidence-linked improvement
What the checker struck out is the evidence
A finding nobody raised leaves no trace of itself. So the loop is fed from the other side — the strike-out, and the presentation that came back reworked.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Struck-out findings, late discrepancies or reworked presentations move in one credit cohort.
- 02
Diagnose
Taken back to the credit as amended — which term was compared, which field was read, which practice was cited.
- 03
Improve
The comparison rule, the checklist or the citation standard is re-approved by the trade-finance head and version-linked.
- 04
Verify
Re-examined over presentations already determined, including the ones a checker struck findings from.
- 05
Learn
The wording that carried two readings is written into the checklist, next to the finding it produced.
Learn → Detect
The return edge. A presentation can be reworked and presented again while the credit is still open, so each cycle re-examines sets the last one had already handed over.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, credit and document modelling, comparison and citation, evaluation, the checker queue, then production validation and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Presentation discovery and boundary definition.Week 1
- 02Trade-system and capture assessment.Week 1, Week 2
- 03Credit, amendment and rule-set modelling.Week 1, Week 2
- 04Document-set taxonomy and checklist design.Week 2
- 05On-its-face comparison per document type.Week 2, Week 3
- 06Cross-document data-conflict tests.Week 3
- 07Date, expiry and examination-period logic.Week 3, Week 4
- 08Citation standard and unevidenced marking.Week 4
- 09Referral routing for sanctions and crime.Week 4
- 10Two-sided error and slice evaluation suite.Week 4, Week 5
- 11Checker handover and trade-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.
| Capability✓ in scope · — not at this tier | PilotOne credit type, one document set | ProductionProduction queue integration | AdvancedMulti-branch / multi-rule-set |
|---|---|---|---|
| Introduced at Pilot | |||
| Document-by-document comparison to the credit | ✓ | ✓ | ✓ |
| Each finding cited to document, field and term | ✓ | ✓ | ✓ |
| Cross-document conflict and date checks | ✓ | ✓ | ✓ |
| Named-checker gate on refusal and payment | ✓ | ✓ | ✓ |
| Sanctions and crime questions routed out | ✓ | ✓ | ✓ |
| Audit trail and baseline evaluation | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Additional rule sets and document sets | — | ✓ | ✓ |
| Trade-system and case-workflow integration | — | ✓ | ✓ |
| Observability and evaluation | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multi-branch and multi-jurisdiction controls | — | — | ✓ |
| High volume and round-the-clock cover | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on the credit types and rule sets in scope, trade-system and capture integrations, document volume, approval 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 month of presentations and how each one was determined
→Presentation discovery and automation-boundary definition
Week 102Your credits, and the amendments in force when each was checked
→Credit, amendment and rule-set modelling
Week 203Access to your trade system and document capture
→Trade-system and capture assessment, then integration
Week 204The checklists your checkers work from today
→Document-set taxonomy and comparison logic per document type
Week 305The findings your checkers struck out, and why
→Evaluation suite, slices and regression cases
Week 406Who owns a sanctions or dual-use question in your bank
→Referral routing and the limits on what a list may say
Week 407Named checkers, and who may refuse or ask for a waiver
→Handover, determination routing, 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 presentations your checkers have already determined, so nothing live waits.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1The presentations in scope, who checks them and who may refuse
W2Trade-system, document-capture and credit-record access
W3Per-document comparison, conflict tests and date logic
W4Citation standard, referral routing and the evaluation suite
W5Presentations replayed against determinations you have already made
W6Checkers work the live queue under review, then handover
- Reading the band
- Date, expiry and period logic lands before evaluation starts, because a period counted differently halfway through a test invalidates every result behind it. The bars show that order, not a smooth ramp.
- At the end of W6
- Checkers have worked live presentations from the agent's list, the findings a checker struck out 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 checking agent around the credits you already examine.
Show us a month of presentations, the findings your checkers raised and who is allowed to refuse one. We'll re-examine that month against your own determinations and show you where the two lists differ — the findings a checker struck out, and the discrepancies that surfaced only after the notice had gone.
- Typical delivery
- 4–6 weeks
- Deployment
- Your trade-finance stack
- Determination
- Named document checker
- After launch
- Agent Care available
Nestack Agents · Trade-finance document checking
AGT-BK-10 · Agent Care available after launch