Use cases/Financial services/Regulatory reporting
AI regulatory reporting for banks and insurers
Regulatory reporting means the figures for each return are pulled from the source systems and reconciled before they reach the analyst. Every difference is listed and traced to the record behind it, and each figure carries a lineage record that shows where it came from. A person signs off every return.
Talk to our team →- Supports
- Regulatory reporting analyst
- Reads
- Core banking, General ledger, Loan system, Claims system, Prior returns
- Runs on
- Your servers, on premises
- Connects to
- Temenos, Infosys Finacle, Ultradata, Data Action
- Decides
- A person, every time
Architecture
Everything runs on your servers
We build it on pipelines in Airflow or Dagster, with dbt models over PostgreSQL and the source extracts. An open-weight model such as Llama or Mistral, served with vLLM on NVIDIA GPUs inside your environment, writes the explanation for each difference. Reconciliation checks run as tests, and the lineage record is stored with the return.
- SourcesCore banking / General ledger / Loan system / Claims system / Prior returns
- Your serversPull each figure → Reconcile sources → Trace differences
- ReviewRegulatory reporting analyst
- OutputFigures pulled from source systems / Reconciliation across sources / Differences traced to the source / Lineage record for each figure
It takes four jobs off the regulatory reporting analyst
- 01
Figures pulled from source systems
The figures for each return are pulled from core banking, the general ledger, the loan system or the claims system. Each figure is stored with the query and the extract it came from.
- 02
Reconciliation across sources
Figures that should agree are compared across systems and against the prior period. Every difference is listed with its size, and the analyst sees them in one view.
- 03
Differences traced to the source
Each difference is traced to the accounts, transactions or policies behind it. A draft explanation sits beside it, and the analyst confirms or rewrites it before sign off.
- 04
Lineage record for each figure
Every figure keeps a record of its source systems, the steps applied to it and who reviewed it. The record is kept with the return for the reviewer and the auditor.
It goes live in five steps
- Step 01
Collect past returns and their workings
Submitted returns with the spreadsheets and queries that produced them, and the differences found at each reconciliation. This is the test set, and access to it is restricted to the people who already have it.
- Step 02
Map each figure to its source
We sit with the reporting analysts and list, for each figure in a return, the system and the extract it comes from. Each mapping is written down and each reconciliation rule is encoded.
- Step 03
Test against past returns
The pipeline rebuilds figures for periods that were already submitted. Analysts compare the result with the filed return and the differences list with the one they found by hand.
- Step 04
Run in shadow
Analysts prepare the return as usual while the pipeline builds it beside them. They mark each difference the pipeline missed or explained wrongly, and those marks feed the test set.
- Step 05
Person signs off and the log records it
The analyst reviews every difference and the sign-off owner approves the return. The lineage record and the log show what each figure read and who reviewed it.
Software or a custom build
Your core systems hold the figures, and the submission to APRA stays on the channel you already use. A custom build does not replace either. Use the reporting tools in your core banking or general ledger platform first if a return draws on one system. A custom build earns its cost when a return draws on several systems, when reconciliation runs in spreadsheets, or when the lineage for a figure cannot be shown to a reviewer.
Temenos / Infosys Finacle / Ultradata / Data Action / Guidewire / SharePoint
In practice
We build this on servers you own. The pipelines, the model and the lineage store sit inside your environment, and ledger and customer data is not sent to a public AI service. The data layer falls under data engineering, and the explanation step under AI agents and workflow automation.
APRA states that authorised deposit-taking institutions and general insurers are required, under the Financial Sector (Collection of Data) Act 2001 and its reporting standards, to provide data to APRA. The data is defined in the reporting forms and instructions. APRA publishes reporting resources for authorised deposit-taking institutions and for general insurers. For authorised deposit-taking institutions it lists some forms as subject to audit requirements. Your reporting team confirms which forms and standards apply to your business.
How the output is checked is covered in LLM evaluation and observability. Where the human step sits is covered in where to put the human. The wider picture for banking and insurance has its own pages.
Questions
What is AI regulatory reporting?
It is the pulling, reconciling and tracing that happens before a person signs off a return. The pipeline gathers the figures from the source systems, lists every difference and traces it to its source. A lineage record is kept for each figure. A person signs off the return.
Does the AI sign off or submit a return?
No. A person signs off every return and submits it through the channel you already use. The system prepares the figures, the difference list and the lineage record for that person to review.
How is each figure traced back to its source?
Each figure is stored with the system, the extract and the processing steps that produced it. The reviewer opens the record behind any figure. The lineage is kept with the return for the auditor.
Do we need to replace our core banking or general ledger system?
No. We build beside the systems you already run, reading from them. They stay the source of record, and the submission process stays as it is.
Industries where this comes up
More financial services use cases
Document search / Credit file assembly / AML alert triage / Claims triage / Month-end close / Audit evidence matching / Advice file drafting / Lease abstraction / KYC and onboarding / Complaints handling / Underwriting submissions / Board and management reporting
Before you call
Tell us where your hours go.
We map the work, size the hardware and price what is worth automating.
Talk to our team →