Banking AI agent · Regulatory reporting
Regulatory-Reporting & Compliance-Monitoring AI Agent
Assemble a return with each figure traced to source, reconcile it to the last submission and to the ledger, then monitor thresholds across the book — filing and attestation stay with the reporting officer.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Submission
- Officer only
- After launch
- Agent Care
What this agent does
Assembles the return, attests to nothing
In
- 01
Draw the figures a return needs from the core, finance, risk and product systems that hold them.
- 02
Pick up the taxonomy version, the reporting instructions and the submission filed last period.
Reason
- 03
Map each cell to the source field it came from, the transformation applied and the person who owns it.
- 04
Reconcile the return to the general ledger and to the figures filed for the period before it.
- 05
Explain each movement against the drivers the underlying data carries, and mark those it cannot.
Decide
- 06
Run the taxonomy's own validation rules, and separate a rule that fails from a figure that is wrong.
- 07
Route unexplained movements, threshold breaches and stale mappings to the officer who owns them.
Out
- 08
Retain the lineage, the validation run, the reconciliation and the officer's decision on each cell.
- 09
Write into the reporting system only inside the approval boundaries agreed during implementation.
Product statement
The agent assembles, reconciles and monitors inside the approval boundaries agreed during implementation. Submitting a return, attesting to one, and deciding that a breach is not reportable stay with the reporting officer and the accountable executive who own them.
Example workflow
One return, end to end
- 1Period openedThe reporting calendar, the taxonomy version in force and the submission filed last period
- 2Figures assembledCell by cell from core, finance, risk and product systems, each against the source field it came from
- 3Reconciled and explainedTo the general ledger and the prior submission, with a driver named for each movement or marked as missing
- 4Validation runThe taxonomy's own rules, the client's own checks, and a failing rule kept apart from a doubtful figure
No human action required
Stages 1 to 4 run without a person in the loop — assembly, reconciliation and validation finish before the officer is asked to open the pack. A source system that has not closed its period ends that stretch on the spot.
Reaches the officer's queue ready to be reviewed.
Goes to the owner of that figure first.
Reporting officer
Reads the lineage, the reconciliation, the validation result and what the agent could not explain, then decides what is submitted.
Approve · Correct · Escalate
Approved — handed back
- 6Pack queued for the officerWritten only where access and policy allow; the submission and the attestation are not the agent's to make
- 7Outcome evaluatedCells corrected, movements left unexplained, rules failed, alerts dispositioned and amendments filed after the fact
Corrections
Cells the officer re-traces or re-explains are counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Submitting a return to a supervisor.
- Certifying or attesting to a return.
- Signing anything an officer attests to.
- Deciding that a breach is not reportable.
Automation boundary
Agent acts unaided
- ✓Assemble the return cell by cell, each figure traced to its source field.
- ✓Reconcile it to the ledger and to the last submission.
- ✓Run the validation rules and say what failed.
- ✓Explain the movements it can, name the ones it cannot, and route breaches to the officer.
Writes run only inside the approval boundaries agreed during implementation. Submitting a return is not one of them.
- Amending or resubmitting a filed return.
- Overriding a validation failure with an explanation.
- Changing a reporting interpretation or a threshold.
- Closing a monitoring alert without a disposition.
Example output
One prepared return, annotated
Everything the agent assembles is attached to the source field or the filing it came from.
Prepared return · one reporting period
Illustrative example- Return
- Quarterly prudential return
- Period
- Quarter now closed
- Cells traced
- 612/640
- Handed to
- Officer — cells unexplained
- Confidence
- 89%
- Submission
- Not made by the agent
As received
The return as the calendar raised it, and the cells as the systems held them at close — nothing here is restated.
Evidence used
- Source field and owner
- Prior submission, as filed
- Ledger balance at close
Why the rest stopped
Two extracts were taken before the sub-ledger closed, so those cells carry no driver to name.
What the score decides
Confidence decides how much the officer re-traces, not whether the return is right.
Action
ApproveCorrectEscalateValue
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
Each return as it falls due
Prudential, statistical and conduct returns
Work to the version in force
Build against the taxonomy version, the reporting instructions and the interpretations that apply to that period — not the ones the last return was built to.
Hand over a return already traced
Each cell set against its source field, reconciled to the ledger and to the last submission, so the officer opens on the movements rather than on the build.
Name what it cannot explain
A driver the data does not carry, an extract taken before a system closed, a mapping left over from the last version — named as such, rather than passed on as a finished return.
Build an evidence trail
Retain the source field behind each cell, the transformation, the validation run, the reconciliation, the officer's correction and what moved since the last submission — on both paths.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Core & finance
- FIS
- Fiserv
- Jack Henry
- Temenos
- General ledger and sub-ledgers
Risk & treasury
- Credit risk
- market risk
- Capital and liquidity calculators
Reporting & taxonomies
- Reporting platforms
- taxonomies
- Validation rule sets
Agent
Regulatory reporting & monitoring
- Traces each figure
- Reconciles and explains
- Routes breaches
Monitoring & cases
- Limit monitoring
- threshold breaches
- Case management
- issue registers
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 figure and a submission
Each control wraps the one inside it. A return clears every layer before the officer is asked to read it, and the attestation sits outside all six.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Decide whether a figure may stand
- L3
- Decides which rules it is built to
- L4 – L5
- Keep the filing with a person and the trail intact
- L6
- Pulls automation back when signals degrade
How Nestack evaluates it
Evaluate the return an officer attests to — not only whether it validated.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the pack the officer reads
- E1Final-output evaluationWas the figure right for that cell, period and version?
- E2Lineage evaluationDoes each cell trace to the source field it claims?
- E3Reconciliation evaluationDoes the return agree with the ledger and the last submission?
- E4Explanation qualityWere movements explained by drivers the data actually carries?
- E5Slice evaluationHow does return quality change across specific reporting cohorts?
- E6Business outcomeHow much did the officer re-trace, and what came back after filing?
Floor — the submission the officer has to stand behind
Failure modes
Where each failure originates in the agent
Seven failure modes plotted against the five stages of the agent lifecycle.
Agent lifecycle
Direction of processing →
01 · Assembly
2 modes
- RR-01
Cell traced to the wrong field
A mapping survives a change in the source system.
- RR-02
Extract taken before close
A source system had not finished the period.
Stage gathers
Cells, source fields, mappings and the last submission
02 · Reconciliation
1 mode
- RR-03
Movement explained away
A driver is asserted that the data does not carry.
Stage explains
Movements against the ledger and the period before
03 · Validation
1 mode
- RR-04
Passes the rules, still wrong
A figure built to a superseded instruction validates.
Stage validates
The return against the taxonomy's own rule set
04 · Monitoring / write
2 modes
- RR-05
Alert left undispositioned
A breach alert ages with nobody recorded against it.
- RR-06
Pack queued as if signed
A return is marked ready and the attestation step is not recorded.
Stage presents
Breach alerts and the pack the officer reads
05 · Change / Version
1 mode
- RR-07
Instruction change not picked up
The template moves and the build stays where it was.
Stage tracks
Taxonomy, instruction, threshold and model changes
- Sev-1 · a boundary is crossed without a person
- Sev-2 · a wrong figure reaches the return
- Sev-3 · assembly degrades, the pack is held
Affected slices
A repeat return is mostly a re-run
A return whose sources and instructions have not moved comes back much as the last one. What fails is a new taxonomy, a spreadsheet-fed cell, or a book bought since the mappings were 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 |
|---|---|---|---|---|
| First return on new taxonomy | 4.2% | 2.8× | Review | |
| Cells fed by spreadsheets | 3.2% | 2.1× | Review | |
| Newly acquired portfolios | 2.6% | 1.7× | Watch | |
| Steady-state repeat returns | 1.2% | 0.8× | Normal |
Bar: corrected-cell lift vs. steady-state baseline · scale 0–4.0× · tick marks the 2.0× review threshold
2 of 4 slices over threshold
Evidence-linked improvement
A return can be wrong long after it validated
A rule set that passes says the return is well-formed, not that the figure behind it is right. What settles it is the officer's correction, or the amendment.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Corrected cells, failed rules or unexplained movements move in one return or one cohort.
- 02
Diagnose
Followed down to the cell — the source field, the mapping, the extract window or the instruction it was built to.
- 03
Improve
The mapping, the reconciliation rule or the explanation standard is re-approved by the reporting officer and version-linked.
- 04
Verify
Re-assembled over periods already submitted, including the ones that validated first time.
- 05
Learn
The cell that broke is kept as a case, and the returns built on the same mapping are re-tested before release.
Learn → Detect
The return edge. An amendment is a supervisory event in its own right, so each change is tested against periods already filed before it reaches a live return.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, sources and lineage, reconciliation and validation, evaluation, integration, then production validation and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Return scope and attestation-boundary definition.Week 1
- 02Core, finance and risk source assessment.Week 1, Week 2
- 03Taxonomy version and instruction mapping.Week 1, Week 2
- 04Cell-level lineage and owner mapping.Week 2
- 05Ledger and prior-submission reconciliation.Week 2, Week 3
- 06Movement explanation and driver rules.Week 3
- 07Validation runs and failure handling.Week 3, Week 4
- 08Threshold and control-breach monitoring.Week 4
- 09Evaluation suite and cohort regression cases.Week 4, Week 5
- 10Officer review pack and queue handoff.Week 5
- 11Reporting-system write-back and change control.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 return, one period | ProductionProduction reporting cycle | AdvancedMulti-entity / multi-regime |
|---|---|---|---|
| Introduced at Pilot | |||
| Cell-level lineage retained as an audit trail | ✓ | ✓ | ✓ |
| Reconciliation to ledger and last submission | ✓ | ✓ | ✓ |
| Validation runs against the version in force | ✓ | ✓ | ✓ |
| Threshold and breach alerts routed to officers | ✓ | ✓ | ✓ |
| Submission and attestation left to officers | ✓ | ✓ | ✓ |
| Baseline evaluation | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Core, finance and risk system integration | — | ✓ | ✓ |
| Reporting-system write-back and handoff | — | ✓ | ✓ |
| Observability and evaluation | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multiple returns and reporting regimes | — | — | ✓ |
| Multi-entity consolidation and enterprise controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on returns in scope, source systems, reporting regimes, cell volume, monitoring scope, 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
01The returns in scope and who attests to each one
→Return scope and attestation-boundary definition
Week 102Access to the core, finance and risk source systems
→Core, finance and risk source assessment
Week 203The taxonomy version and the instructions you build to
→Taxonomy version and instruction mapping
Week 204Whatever maps your cells to source fields today
→Cell-level lineage and owner mapping
Week 205The last submissions and the ledger they were tied to
→Ledger and prior-submission reconciliation
Week 306Returns you amended, and the findings behind them
→Evaluation suite, cohort regression cases and failure-mode testing
Week 407Named reporting officers, and how alerts are dispositioned
→Officer review pack and queue handoff, then the pilot return
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 carries both the cohort testing and the first supervised return.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Return scope, attestation boundary and the reporting calendar
W2Source access, taxonomy version and cell-level lineage
W3Reconciliation, movement explanation and validation runs
W4Evaluation suite, cohort testing and breach-monitoring rules
W5Reporting-system integration, a supervised return and corrections
W6Officers read from live packs, then Agent Care starts
- Reading the band
- Nothing is submitted in these six weeks. The week 5 return is built alongside the one your team files, and the two are compared cell by cell.
- At the end of W6
- A live period has been assembled, reconciled and validated in parallel with your own process, then 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 reporting agent around the returns your officers already attest to.
Show us one return, the systems its figures come from and who signs it off today. We'll rebuild a period you have already submitted, put each cell next to the source field it should trace to, and show you which movements your own pack could not explain.
- Typical delivery
- 4–6 weeks
- Deployment
- Your reporting and core stack
- Submission
- Reporting officer, always
- After launch
- Agent Care available
Nestack Agents · Regulatory reporting & compliance monitoring
AGT-BK-11 · Agent Care available after launch