Public-company analysis · Banking
Citi

When data becomes a control failure.

Across 2020 and 2024, U.S. regulators assessed Citi $535.6 million in civil money penalties tied to risk management, data governance, data quality, and internal-control deficiencies. Citi reported $3.3 billion of transformation-related expense in 2025.

The visible price is not compute. It is the inability to prove where a number came from, what changed, and who accepted the exception.

Independent analysis of public information. Citi is not a Fitzroy client.

Financial district office towers reflected across the harbour
The pattern

A number can be correct and still fail as evidence.

Regulatory reporting is a chain of custody. Source systems create balances, risk engines transform them, finance applies adjustments, and control owners attest to the result. When lineage, validation, and ownership sit in separate tools, each reporting cycle reconstructs that chain by hand.

The 2024 actions make the operating weakness unusually specific. The OCC cited a lack of processes to monitor how data-quality concerns affect regulatory reporting; the Federal Reserve cited inadequate compensating controls while the underlying program remained unfinished.

Citi also reports substantial progress, with more than 80% of transformation programs at or near target state by the end of 2025. The Fitzroy proposition is narrower: in an analogous reporting domain, preserve the systems of record and make the path from source fact to reported number reproducible by design.

Public record
$535.6M
The OCC’s $400M penalty in 2020, plus $75M from the OCC and $60.6M from the Federal Reserve in 2024.
$3.3B
Citi’s 2025 transformation-related expense, up 14%, driven largely by data, regulatory reporting, and risk-and-control work.
Proposed reference architecture

Make every reported number reproducible.

The layer accepts both CDC streams and controlled batch extracts, preserves versioned evidence, applies explicit quality and reconciliation rules, and records lineage before routing exceptions through named ownership. The underlying systems remain the record; existing report-authoring and submission systems remain in place.

SOURCESCore bankingGL + subledgersRisk + finance martsMaster dataSystems of record, unchangedMSK ConnectCDC adaptersTransfer FamilyBatch extractsAmazon MSKOrdered streamsAmazon S3Versioned landingGlue Data QualityRule validationLambdaReconcile recordsVPCAmazon DataZoneLineage + catalogAthenaEvidence queriesStep FunctionsException reviewDynamoDBControl stateQuickSightControl dashboardsS3 Object LockDecision evidenceSNSOwner escalationSECURITY · OBSERVABILITYIAM Identity CenterKMS + tokenizationLake Formation policyCloudTrail auditSchema registry + replayCross-Region DR
The dashed line shows retried asynchronous escalation. VPC marks private reconciliation compute.

Preserve the chain.

MSK Connect captures database changes while AWS Transfer Family receives controlled batch extracts. Amazon S3 preserves the source state behind each run.

Test before it travels.

AWS Glue Data Quality applies explicit rules; Amazon DataZone records lineage independently. Failed records become visible exceptions, not report inputs.

Make ownership executable.

AWS Step Functions places named review and remediation around each exception; S3 Object Lock retains the decision record after an authorized owner closes it.

Economics

Price the control layer against recurring work.

A penalty is visible, but it is not a responsible savings assumption. The planning case is built from recurring reconciliation, report rework, and remediation effort that a finance team can test directly.

Modelled run rate

What the platform costs to run.

Modelled for one high-volume reporting domain: 15TB ingested across stream and batch, 1B source changes, 6,000 Glue DPU-hours, 100TB of Athena scans, 30M control executions, and 100 dashboard readers per month.

ServiceBasisMonthly
Streaming and managed ingressMSK + Connect, Transfer Family$3,600
Transformation and qualityGlue Data Quality, 6,000 DPU-hours$3,000
Evidence, lineage, governanceS3 Object Lock, DataZone, Lake Formation$1,100
Reconciliation and workflowLambda, Step Functions, EventBridge$1,250
Query and control stateAthena, DynamoDB$850
Security and reportingCloudWatch, KMS, endpoints, QuickSight$2,350
Total$12,150 / month
Illustrative impact model

What it is modelled to return.

Fitzroy planning case for one regulatory-reporting domain at a large bank.

LeverAssumptionAnnual
Reconciliation and attestations60 FTE at $175K; 25% capacity released$2.6M
Regulatory report rework24 cycles at $225K; 30% avoided$1.6M
Remediation and retesting$10M annual spend; 15% avoided$1.5M
Annualized opportunity$5.7M
Implementation
$900K planning case
Modelled payback
4 – 6 months

AWS list-price planning estimate as of August 2026; implementation and support excluded. The impact model is illustrative and is not a reported result for a company named on this page.

For the board

The control case, expressed as operating economics.

First-year net
$2.1M

Fifty-five percent of annualized benefit in year one, less the $900K build case and $146K first-year run cost.

Benefit-to-cost
3.0×

First-year gross benefit divided by implementation and first-year run cost.

Annualized opportunity
$5.7M

Capacity value and cost avoidance—not revenue, guaranteed savings, or a result achieved for Citi.

The annual run case is 2.5% of the modeled annual opportunity. The argument does not depend on avoiding a headline penalty; it depends on reducing the recurring work required to make important numbers trustworthy.

Facing a similar control challenge?

Start with an Enterprise Systems Assessment.

Flexible engagement scope.Start an assessment