Regulatory reporting · Prep
Multi-workflowRegulatory reportingWhich regulatory reporting prep step to hand to AI agents first
Regulatory returns are signed by people, but most of the weeks before sign-off go on checking, tying out and explaining. Here is how to choose which of those steps to give to workflows first.
Valnora & Elric team5 min read
Compared in this guide
- AData completeness checks
- BReconciliations
- CVariance commentary
- DTemplate validation checks
A regulatory return is a deadline with a signature on it. In the EU, prudential returns such as COREP and FINREP follow the EBA’s reporting templates; other regimes have their own. Whatever the template, the work before sign-off looks similar: check the data arrived, tie it to the ledger, explain what moved and catch errors before the supervisor does.
Each of those steps can be a workflow in Agent Serve. None of them should produce or submit the return; your reporting tool and your people do that. The real question is which to build first. This guide compares four options and says when to pick each.
The four options
Four prep steps compared
| Prep step | What the workflow does | What you need | Where people stay | Start here when |
|---|---|---|---|---|
| Data completeness checks | Counts rows, finds missing fields and entities, compares with last period | Read access to the reporting data mart and last period’s snapshot | Data owners fix the gaps at source | Late or missing feeds are your biggest source of rework |
| Reconciliations | Ties reporting balances to the ledger and logs each break with an owner | Ledger and reporting data at the same grain, plus agreed tolerances | Owners explain and clear breaks; controllers approve adjustments | Breaks are found late, in spreadsheets, by one person |
| Variance commentary | Drafts explanations for movements above threshold, citing drivers | Clean, reconciled data and a library of past commentary | The reporting lead edits and approves every line | Reviewers and supervisors keep asking why figures moved |
| Template validation checks | Runs your own cross-checks between related lines before the reporting tool’s validations | A written list of checks your team already performs by hand | Analysts investigate every failed check | You pass submission validations but find errors in internal review |
All four are built in the workflow builder and read your data where it lives, through data sources such as Snowflake, BigQuery or a database connection. They differ in what they need from you and in how much judgement stays with a person.
Choose completeness checks when feeds are the problem
If your team regularly discovers on day three that a subsidiary’s feed was half empty, start here. It is the cheapest workflow to build and the least risky, because it only reads and counts.
A Schedule trigger runs on your reporting calendar. Database blocks pull counts and key fields from the data mart, a Function block sets them against last period’s snapshot, and Rules flag what fails: entities missing, null rates above normal, totals far outside their usual range. Failures go to the data owner through a Slack or Microsoft Teams block, with the exact rows attached.
It also builds the habit the Basel Committee’s risk data principles, known as BCBS 239, push for: knowing that your reporting data is complete before anyone relies on it.
Choose reconciliations when breaks are found late
Reconciliation is the step most teams still run in spreadsheets. A Function block compares reporting balances with general ledger totals line by line and writes every difference above tolerance to a breaks table, with an owner and a status.
The table changes the conversation. Instead of a reconciliation file passed around by email, each break has one row that its owner updates. A dashboard on the table shows open breaks by line and age, and the reporting lead can see on any morning what still stands between the team and sign-off. Adjustments still go through your controllers; the workflow only finds and tracks.
Choose variance commentary when reviewers keep asking why
Commentary is where language models help most and where review matters most. An Agent block takes each movement above your threshold, searches past commentary and known drivers in the Knowledge Base, and drafts an explanation that cites the data behind it. A Human block then pauses the run until the reporting lead edits and approves each line.
What the solution produces
Draft variance commentary (illustrative)
Record- Line
- Loans and advances to customers, stage 2 exposures
- Movement
- +8.4% quarter on quarter, above the 5% commentary threshold
- Draft explanation
- Increase driven mainly by commercial real estate exposures moving from stage 1 after the quarterly review of significant increase in credit risk.
- Evidence
- Stage transfer table for the quarter; credit committee minutes referenced in prior commentary.
- Confidence note
- Driver matches data. Committee reference needs confirming by the reviewer.
Build this one last if your data is not yet reconciled. A fluent explanation of a number that is wrong is worse than no explanation. The draft is also only as good as the drivers on file, so keep last period’s approved commentary in the Knowledge Base every cycle.
Choose template checks when internal review finds what validations miss
Reporting tools run the regulator’s validation rules. Most teams also keep their own list of cross-checks: a line that should never exceed another, a ratio that should stay in a band, a total that should match a management report. If that list lives in someone’s head, write it down and turn it into rules in a workflow that runs before internal review. Each failed check goes to an analyst to investigate.
How the workflows fit together
Three prep workflows ahead of reporting sign-off
Workflow 1 · Completeness
- 01Reporting calendarA Schedule trigger runs on the working days your reporting timetable sets.
- 02Pull source extractsSnowflake, BigQuery or database blocks read the reporting data mart.
- 03Completeness rulesRules flag low row counts, null fields and missing entities compared with last period.
Workflow 2 · Reconciliation
- 04Tie out to ledgerA Function block compares reporting balances with general ledger totals by line.
- 05Breaks tableEach difference above tolerance becomes a row with owner, amount and status.
- 06Notify ownersA Slack or Microsoft Teams block tells each owner about their breaks.
Workflow 3 · Variance commentary
- 07Prior commentaryA Knowledge search finds last period’s explanations and known drivers.
- 08Draft commentaryAn Agent block drafts an explanation for each movement above threshold, citing the data.
- 09Reviewer signs offA Human block waits for the reporting lead to edit and approve each explanation.
Most teams end up with all three or four, chained in the order of the reporting timetable. A Workflow block lets one run another as a step, so completeness can start reconciliation once it passes. Each workflow runs from a deployed version, so a change to a tolerance or a prompt is reviewed before it touches a live cycle, and every run is traced in Logs with each block’s inputs and outputs. That trace is useful evidence when internal audit asks how a figure was checked.
Controls and governance
Run logs
Every workflow run records each block’s inputs, outputs and errors, so a reviewer can trace how a result was reached.
Deployed versions
Workflows run from a deployed version, so a change is reviewed before it reaches production runs.
Audit trail
Proposals, approvals, declines and executed actions are written to the workspace audit log, which admins can export.
Roles and permission groups
Admin, write and read roles per workspace, plus permission groups that limit which models, blocks and integrations a team can use.
Row-level access
Row-level policies narrow what each person sees in tables, dashboards, exports and scheduled reports.
Self-hosting
Run Agent Serve in your own cloud or data centre, next to the systems it reads.
Bring your own keys
Use your organisation’s own model provider keys at organisation or workspace level.
More on how Agent Serve handles access, data and audit on the security page.
Where to start
Pick the return that causes the most late nights and list every manual check your team performs on it, in order. The first item on that list that is purely mechanical is your first workflow. Run it alongside the manual process for one cycle and compare results before you rely on it. If supervisors also ask how data flows into your returns, the data lineage field notes pick up from there.
What you need
- Read access to the reporting data mart and general ledger through a database or warehouse connection
- Your reporting timetable, so schedules match the working days that matter
- Agreed tolerances for reconciliation breaks and thresholds for variance commentary
- Past commentary and known drivers, uploaded to a Knowledge Base
- Named owners for each break category and a reporting lead to approve commentary
- A clear line: your reporting tool produces and submits the return, not the workflows