Case studies
Financial Services

Dozens of Systems That Do Not Agree With Each Other

Returns build from source data, with lineage recorded on every figure.

CreateOS for Regulatory Reporting Cycles
On this page
100%

Data lineage and traceability: 100% source-to-report.

$7.5M-$12.0M

Annual cost reduction on a $30M reporting and audit-preparation base.

25%-40%

Operational cost reduction, reporting and audit prep.

Zero

Unauthorized data egress, enforced in-kernel and self-hosted.

Challenge

Every reporting cycle looks the same. Data is pulled from dozens of systems that do not agree with each other, reconciled by hand in spreadsheets by people senior enough to know what a break means, then formatted to a regulator's specification and pushed to a human who has to put their name against the number.

  • That is one regime, and the institution runs many. Call reports, capital and liquidity returns, transaction reporting, incident reporting, and disclosure. Each pulls the same senior people back into assembly work.
  • The accuracy bar is absolute. A misreported figure is not an internal mistake to fix next cycle. It is a supervisory finding, a potential fine, and at worst a restatement.
  • The cost is rising and was met by hiring. Employee hours spent complying with financial regulation and examiner mandates rose 61% between 2016 and 2023, while aggregate employee hours rose 20% (Bank Policy Institute, 2024). The response almost everywhere was to add people.
  • The regulators have acknowledged the weight of it. The Federal Reserve, OCC, and FDIC issued a joint request for information on streamlining the Call Report, published in the Federal Register on 1 December 2025 with comments due 30 January 2026 (OCC Bulletin 2025-42).
  • Every proposal stalled in the same place. The tooling wanted reporting data in a vendor cloud, and lineage assembled outside the boundary is lineage the controllership cannot defend to an examiner.
  • Not short of RegTech vendors, short of somewhere to run them. Which is why the function stayed manual.

One return, both ways

Dozens of systems

Built from source

Pull the figures

Pull the figures

From systems that do not agree

From source, on allowlisted paths

Reconcile the differences

Reconcile the differences

By hand, cycle after cycle

One templated run, identical each time

Produce the lineage

Produce the lineage

Reconstructed if an examiner asks

Written as each step runs, tamper-evident

Attest and file

Attest and file

A human signs it

A human signs it

The last pair is identical because the filing authority never moves. What changes is the third: a trail assembled after the fact is a reconstruction, and a controllership that cannot defend its lineage keeps doing the work by hand however much the tooling costs.

Solution

CreateOS built and deployed a crew of specialized agents that take a reporting cycle from "cycle opens" to "regulator-ready return, with every figure traced to its source, waiting for a human to attest it."

Almost all labour, almost none judgment

Data pulled from dozens of systems that disagree, reconciled by hand.

$7.5M-$12.0Ma year on a $30M base

Assembled, validated, drafted

Breaks flagged before they reach a return, built in the regulator's own format.

Nothing is filed autonomously. Compliance responsibility is not delegated to a model.

Nothing is filed autonomously. That is a design constraint, not a limitation, and it reflects the regulatory expectation made explicit in 2025: compliance responsibility cannot be delegated to a model.

  • The reporting corpus assembles itself. Required data pulled from source systems, ledgers, and the risk and finance platforms into one place, instead of reconciled by hand across systems that disagree.
  • Breaks are flagged before they can reach a return. The corpus is checked for quality, consistency, and completeness against the reporting rules as the run happens.
  • The return is drafted in the regulator's own format. Including the narrative where one is needed, so what reaches a human is a draft to review rather than a cycle to assemble.
  • Lineage is native to the runtime, not bolted on. Every reported figure traces back through a logged, sequenced, tamper-evident trail. In regulatory reporting the lineage is not a feature of the product, it is what the examiner and the auditor actually rely on.
  • Separation is what keeps the trail clean. Gathering, validation, assembly, and lineage in distinct contained environments. A monolithic process that does all four is one an auditor has to take on faith.
  • The data never leaves the jurisdiction. Control plane and storage run inside the institution's own infrastructure, with egress allowlisted in the kernel to approved internal systems and sanctioned regulator endpoints. That is why the pilot ran on real reporting data rather than a sanitised copy.
  • Reproducibility is itself a control. One configured environment executes every cycle identically, so consistency is auditable and drift is a finding. That is what the controllership wanted before it wanted any of the speed.
  • Cycles fork and run in parallel. Returns across entities, jurisdictions, and regimes are produced at the same time rather than queuing behind one another.

CreateOS is SOC 2 Type II and ISO 27001 certified. Most agent builders own the prompts and rent the runtime. We own both, which is the reason the internal audit review was a step in the process rather than the end of it.

Outcome Derived

The cycle compressed. The lineage got better than it was under the manual process. The senior people went back to oversight.

MetricBeforeAfter
Reporting cycle timeDays to weeksSharply compressed
Operational cost, reporting and audit prep$30M baseDown 25% to 40%
Data lineage and traceabilityPartial100% source-to-report
Reporting error rateBaselineReduced via automated validation
Reporting process consistencyDrifts cycle to cycleIdentical templated runs
Unauthorized data egressRiskZero, enforced in-kernel and self-hosted
Human attestation on submissionsManualPreserved by design
Projected. Modeled on stated assumptions and published sources, not measured from a delivered deployment.
  • $7.5M to $12.0M a year on a $30M base. 25% is the conservative floor we commit to in a pilot; 40% is the upper end of the general compliance-automation evidence. The saving comes from the three activities that consumed the most senior time and required the least senior judgment.
  • The contract is written against the institution's baseline. Measured in the first two weeks of the engagement, not against this model.
  • Senior capacity returns from assembly to oversight. 42% of C-suite time and 43% of board time goes to regulatory and supervisory compliance. When assembly moves to the agents, that attention returns to what the breaks mean and what to tell the regulator.
  • Risk reduction is context, not a priced promise. Continuous validation catches breaks before they reach a return and lineage lets every figure be defended to source, which reduces both the probability and severity of a finding. We do not headline a dollar figure, because it depends on the institution's current finding rate.
  • Where the evidence is thinner, we say so. Reporting savings are institution-specific and rarely disclosed cleanly, so this case has fewer hard published percentages than KYC or AML. A finance reviewer who tests these figures should find them dull and defensible, which is the intent.

What We Would Prove, and How

Weeks 1 to 2, baseline. Measure the institution's actual reporting cost and cycle time, its change-monitoring capacity, its error and finding history, and its audit-preparation effort. This becomes the contract's yardstick and protects both sides.

Weeks 2 to 6, build and integrate. Stand up the agent crew, integrate to source systems, ledgers, and reporting infrastructure along allowlisted paths, encode the institution's reporting rules and controls, deploy self-hosted inside the boundary. Scoped to a defined set of returns.

Weeks 6 to 9, parallel run. The agents produce returns alongside the current process without filing anything. Every figure and its lineage is compared against the institution's own filed reports. Proving the numbers and the trail are right is the whole game.

Week 9 onward, controlled adoption. The institution adopts agent-produced returns for the proven regimes, always with a human attesting before submission, expanding scope as the accuracy and lineage record builds.

Success criteria, agreed up front: reporting cost down at least 25%, cycle time materially compressed, reporting accuracy at or above the current process, full source-to-report lineage on every figure, human attestation preserved on every submission, and zero data-residency or leakage incidents.

Highlights

  • Data lineage and traceability: Partial → 100% source-to-report.
  • Annual cost reduction: $7.5M to $12.0M on a $30M reporting and audit-preparation base.
  • Operational cost, reporting and audit prep: down 25% to 40%.
  • Unauthorized data egress: Zero, enforced in-kernel and self-hosted.

Give Us One Stuck Pilot.

We'll have it in governed production before your next board meeting.