Regulatory reporting rarely feels like a problem until the last week before the deadline. Then it feels like the same week you had last quarter, and the quarter before that.

The return is due. Finance has the period-end numbers, but one figure moved after close and nobody is sure whether the extract was refreshed. Risk owns a schedule that has to tie to the general ledger, and the two do not agree by a few thousand. Someone is searching an old email thread to find why last quarter's exposure was reclassified, because the same question is back. The person whose name goes on the submission is asking whether the figures reconcile to source, and the honest answer is a hopeful "mostly." The return will probably go in on time. But it goes in tired, and it leaves very little behind that the next person could actually follow.

That last part is the real cost. A regulatory return is not only a set of numbers submitted by a date. It is a claim that those numbers are right, produced a way the firm can defend if a regulator asks how. When the workflow behind it is manual, the numbers can still be correct and the firm can still be exposed, because it cannot easily show its work.

The filing has to be defensible before it reaches the regulator

The job of regulatory reporting is to submit the right figures, on time, in a form the regulator accepts, with enough evidence that a reviewer can trust how each number was produced. In financial services, that last clause is the whole game. Accuracy on its own is not enough. The firm also has to keep ownership clear and keep every figure traceable back to the system it came from, without rebuilding that trail from memory each cycle.

The return is the visible artifact. The work that makes it defensible happens earlier and mostly out of sight: agreeing what each line means, pulling the source data, reconciling it, checking movements, explaining adjustments, getting the right person to sign, and keeping the evidence somewhere a reviewer can find it a year later. Fix that, and the return stops being a quarterly reconstruction.

Follow one return from trigger to submission

Before changing any tools, take one recurring return and map how it actually gets produced, not how the procedure document says it does. Start at the trigger, usually period-end or a regulatory calendar date, and follow it to the moment someone submits and the confirmation comes back.

Write down who starts the cycle, where each figure is pulled from, who reconciles it to the ledger or the source system, who explains the movements, who approves any manual adjustment, who signs the final return, and where the supporting files end up. For most firms this map is uncomfortable, because it shows how many hands the numbers pass through and how few of those handoffs leave a trace.

The map usually exposes the same thing. The data sits in the core banking or policy system, the general ledger, a risk engine, a regulatory calculation tool, a spreadsheet or two, and a shared drive. The review runs across finance, risk, compliance, and whoever is accountable for the submission. A single late feed or one unclear definition can hold the whole return. Do not automate any of it until the handoffs are visible. If the sequence is unclear on paper, connecting systems will only move the confusion faster.

Where the reporting cycle actually breaks

Regulatory reporting tends to fail in a small number of predictable places, and naming them is more useful than a general complaint about manual work.

The return starts too late to reconcile properly

If assembly only begins once close is done, there is time to build the return but not to question it. Reconciliations get done in a rush, movements get explained after the numbers are already in the template, and anything that does not tie gets a note promising to look next quarter.

Figures lose their link to source

A number leaves the ledger, gets adjusted in a spreadsheet, is typed into the return, and by the time anyone asks where it came from, the answer lives in someone's head. The figure may be right. The firm just cannot show it quickly.

Definitions and classifications drift

Exposure classes, risk weights, netting rules, product mappings, and counterparty categories all carry regulatory definitions, and those definitions change, both in the firm's own mappings and in the regulator's instructions. When a mapping shifts quietly, a movement that looks like performance is really a classification change.

Adjustments are invisible

Manual overlays and top-side adjustments are normal in regulatory reporting. The problem is when one lives as an unlabeled cell override that no reviewer sees and no one can explain later.

Sign-off happens after the story, not the checking

The accountable person is asked to sign once the return is basically built, before the reconciliations have actually been reviewed. One late correction then ripples through several schedules, and the sign-off becomes a formality instead of a real check.

The evidence is scattered when the query arrives

Months later a regulator asks a question about a line in a filed return. The extract, the reconciliation, the adjustment reason, and the approval are in four places, and reassembling them takes days that feel far longer than the original filing.

Work backward from the submission date

Most regulatory returns have a fixed deadline and a known set of inputs, which means the cycle can be planned backward instead of discovered forward. A simple working-back calendar sets the dates for data freeze, reconciliation, review, and sign-off so problems surface with time to fix them, not on submission night.

The point is not more meetings. It is to stop finding source breaks, unexplained movements, and missing approvals after the return is already being finalized. A calendar for a quarterly return might look like this, with the days adjusted to your close timetable.

Timing What happens What exists by the end
Deadline minus 10 to 8 Confirm the return, its scope, source owners, and any changed instructions or taxonomy for this cycle. Agreed scope and owner list.
Deadline minus 7 Freeze the source extracts where possible and flag any late feeds explicitly. Source snapshot and a list of what is still outstanding.
Deadline minus 6 Reconcile each figure to the ledger or source system and record breaks. Reconciliation with open breaks listed.
Deadline minus 5 Run validation checks and the regulator's own form rules. Exception list for review.
Deadline minus 4 Explain movements and adjustments against the same set of questions. First commentary and adjustment notes.
Deadline minus 3 Accountable reviewer checks reconciliations, adjustments, and open items. Reviewed return with open items closed.
Deadline minus 2 Final tie-out, format validation, and submission. Filed return and confirmation.
Deadline plus 1 Capture queries, breaks to fix, and anything to change before the next cycle. Follow-up list for the next return.

Keep one filing tracker for each return

The single most useful artifact is a tracker that shows, for each return, everything a reviewer needs to know without opening ten files. At a minimum it holds the return name and period, the due date, the owner, the required schedules, the source systems and figure definitions, the reconciliation state, any open adjustments, the reviewer, where the evidence sits, and the submission status.

Each return should move through states the team trusts: not started, source data pending, reconciliation in progress, exceptions under review, adjustments pending approval, ready for sign-off, submitted, and post-submission follow-up. The value is that anyone can see where a return stands and what is blocking it, instead of asking the one person who knows.

Return line Source and reconciliation Sign-off Status
Reporting population. Core system extract reconciled to the ledger count. Data owner approves before reconciliation closes. One unmatched account under review.
Exposure classification. Product mapping checked against the current instructions. Risk review before the schedule is finalized. Draft note waiting on approval.
Manual adjustment. Top-side overlay with reason and supporting file attached. Compliance review before filing. Adjustment logged, approval outstanding.
Final submission. Tracker, reconciliations, and approvals attached. Named accountable reviewer. Blocked until the classification note is signed off.

Reconcile every figure back to its source

Reconciliation is what separates a return the firm can defend from one it merely submitted. Every figure in the return should tie to a source: the general ledger, the core system extract, a risk calculation, or a documented adjustment. Where it does not tie, the break should be visible with an owner and a reason, not buried.

The discipline that helps most is keeping the tie-out attached to the number rather than in a separate reconciliation file that ages out of date. When a figure in a schedule comes from a ledger balance plus two adjustments, that chain should be readable without a phone call. This is also what turns a regulator query from a multi-day hunt into a lookup, because the trail from filed number to source is already assembled.

Reconciliation is not only arithmetic. It also asks whether the number makes sense for the business. If exposures grew sharply while the book did not, the tie may balance and the story may still be wrong. Recording that judgment, with the reason and the person who made it, is what keeps next quarter's team from rediscovering the same question.

Run validation while the return is built, not on the final night

Validation should happen as the return moves, not only when the pack is nearly done. Useful checks include expected record counts, completeness against the required schedules, period-over-period movement thresholds, source-to-return tie-outs, cross-schedule consistency, and the regulator's own form or taxonomy validation rules where they can be run early.

Cross-checks are often the most valuable because they catch errors a single number cannot reveal on its own. If total exposure moved but the ledger did not, if a risk-weighted figure fell while the underlying book grew, or if a schedule total no longer agrees with the summary it feeds, each of those deserves a look before submission rather than a question afterward.

When a check fails or an adjustment is accepted, the decision belongs in the tracker with its reason, owner, and supporting file, not in a spreadsheet comment that disappears next cycle. An accepted exception with a recorded reason is an answer ready for the next query. An accepted exception with no explanation is a problem waiting to be rediscovered.

Treat definition and taxonomy changes as their own risk

Regulatory returns run on definitions that are not fully in the firm's control. The regulator updates forms, validation rules, and reporting taxonomies. The firm updates its own product mappings, netting logic, and classification rules. Either kind of change can move a reported figure without anything real changing in the business.

The workflow needs a deliberate place to catch this. Before a cycle starts, someone should confirm whether the form version, the taxonomy, or any instruction has changed, and whether the firm's own mappings moved since last time. When a definition does change, the return should carry a note explaining the effect, so a movement driven by reclassification is never mistaken for performance, by the firm or by the regulator.

This is unglamorous work that tends to have no owner, which is exactly why it breaks. Naming a person responsible for tracking instruction and mapping changes, and recording those changes where the reviewer sees them, removes a whole category of late surprises.

Make sign-off a real decision, not a signature

In regulatory reporting the sign-off is not a courtesy at the end. It is the point where a named, accountable person takes responsibility for the return in front of the regulator. That only means something if the sign-off happens after the reconciliations and adjustments have actually been reviewed, and if the reviewer can see what they are approving.

A sign-off worth the name gives the accountable person a clear view of the reconciled figures, the open breaks and how they were resolved, the adjustments and why they were made, the definition changes this cycle, and anything still outstanding. When those are visible in one place, sign-off is a real check. When they are scattered, the signature is a hope.

This is also the line that AI does not cross. A model can assemble the pack and draft the explanations, but the person who signs the return owns the judgment and the accountability. That responsibility does not move to a tool.

Keep the audit trail assembled, not reconstructed

The audit trail is not paperwork you produce after the fact. It is the by-product of doing the cycle in the open: the source extract, the reconciliation, the adjustment reason, the validation result, and the approval, all kept together and tied to the filed figure. Assembled as you go, it costs almost nothing. Reconstructed under a deadline or a regulator query, it costs days and nerves.

The test is simple. Pick a line from a return you filed two quarters ago and try to show, in a few minutes, where the number came from, what was adjusted, who checked it, and who signed it. If that takes an afternoon of searching drives and inboxes, the trail is being rebuilt each time instead of retained.

Plan for resubmissions and regulator questions

Even a good cycle produces resubmissions and queries, and the workflow should treat them as normal rather than as emergencies. When a figure is restated, the change needs its own reason, approval, and link to the original, so the firm can explain not just the new number but why it moved. A query about a filed return should resolve against the evidence already attached to that figure, which is only possible if the trail was kept the first time. Handled this way, a resubmission is a clean correction with a clear history, not a second scramble.

Connect systems after the checks are clear

The systems involved are familiar: the core banking or policy platform, the general ledger, a risk or regulatory calculation engine, a data warehouse, spreadsheet models, a document store, and the submission portal. Reporting work often stalls because teams try to connect these before agreeing how a figure should get from source to signature.

Once the tracker and the checks are clear, connect only the fields the return actually needs. Build stable extracts, name the owner of each source, define when it refreshes, and store outputs where reviewers can find them. Where a figure is adjusted by hand, the connection should still show the reason and the approver, so automation makes the trail stronger rather than hiding a step.

Where AI helps inside regulatory reporting

AI is useful in this work when it reduces manual assembly and helps reviewers see what changed, and it is risky when it is asked to be the source of truth. The safe uses sit around the return, not at its center.

In practice, a model can assemble a first draft of the return from reconciled sources, match figures across extracts and flag where they disagree, draft first-pass commentary on movements against the firm's own thresholds, summarize what changed since last cycle, pull the prior explanation for a line so the team is not re-deriving it, flag missing evidence or approvals before submission, and turn a regulator query into a checklist of what to gather. Each of these speeds up the work while leaving the judgment with people.

The boundary matters and it is not negotiable. AI assembles, reconciles, and drafts. The accountable person reviews and signs the filing. Every figure keeps its link to source, whether a person or a model produced it. And none of this is a substitute for regulatory or legal advice on what the firm must report. The workflow makes reporting defensible; it does not decide the firm's obligations.

A worked example: a mid-size lender's quarterly prudential return

Consider an illustrative case, with the figures invented to show the shape of the work rather than any real firm. Say a mid-size lender files a quarterly prudential return that pulls exposures from the core banking system, risk weights from a calculation engine, and capital figures from the general ledger, with a handful of manual overlays each quarter. Today the return is assembled in a master spreadsheet over the final week, and each cycle a few figures move after close in ways nobody can quickly explain.

Mapping one cycle shows where the time goes. The exposure population comes from a core system extract that has to reconcile to a ledger count, and this quarter one account does not match. A product has been reclassified into a different exposure class, which changes its risk weight, but the note explaining that lives in an analyst's email. Two top-side adjustments are applied in the spreadsheet with no recorded reason. The accountable reviewer is handed the return the night before, after the story is written but before the reconciliations have really been checked.

A better version of the same cycle does not add software first. It gives the return a tracker, freezes the extracts earlier, reconciles each figure with breaks made visible, and records the reclassification and the two adjustments where the reviewer can see them. The illustrative shape of that tracker might look like this.

Return line Where it comes from This cycle's issue Resolution before sign-off
Exposure population. Core system extract, reconciled to the ledger count. One account unmatched. Break traced to a timing difference, corrected and noted.
Risk weight on a reclassified product. Calculation engine, using the current mapping. Class changed since last quarter. Reclassification note attached, movement flagged as definition-driven.
Capital adjustments. Two manual overlays in the return. No recorded reason. Each overlay given a reason, source file, and approver.
Final figure. Tracker, reconciliations, and approvals attached. Reviewer sees the pack late. Sign-off moved earlier, with open breaks closed first.

None of these numbers are real, and no firm should read them as a benchmark. The point is the change in shape: the same people, the same systems, but a return where every figure carries its source, its movement is explained, and the person who signs it can see what they are signing before the deadline, not after.

What to measure

Measure the cycle, not only whether the return went in on time. On-time submission hides a lot of strain. More telling are the things that show whether the cycle is getting calmer: source extracts received by the freeze date, reconciliation breaks by type and how long they stay open, adjustments applied and how many carry a recorded reason, days spent in review versus assembly, items still open at sign-off, and queries or resubmissions after filing.

If the question is whether the manual effort justifies fixing this at all, it helps to put a number on the current cost. The manual-work ROI guide gives a way to estimate it, and the Workflow Readiness & ROI Calculator tests whether this particular return is a sensible place to start.

Common traps when fixing the workflow

Most attempts to improve regulatory reporting fail in a few recognizable ways, and they are different from the failures in the cycle itself.

The first is building a dashboard before the figure definitions and owners are stable, which makes the confusion look polished without resolving it. The second is keeping evidence in personal drives and email threads, so the trail still has to be reconstructed when a query lands. The third is treating reconciliation breaks and adjustments as spreadsheet comments rather than tracked items with owners. The fourth is letting a model draft commentary that is not tied to reconciled source data, which produces confident text the firm cannot stand behind. The fifth is submitting on time but never capturing what should change before the next return, so the same breaks come back every quarter. And the most common of all is trying to fix every return at once, when one recurring return would prove the approach and teach what the others need.

The first month should stabilize one recurring return

For a first month, pick one return that causes repeated pain and carries real regulatory risk, and leave the rest alone. The aim is to run one full cycle through a cleaner workflow and learn from it, not to redesign reporting across the firm.

A sensible first month traces the chosen return from trigger to submission in the first week, then defines the tracker, the states, the required evidence, and the validation checks in the second. The third week connects the minimum source extracts and builds the reconciliation and exception view. The fourth runs the sign-off from the workflow, captures what came up, and agrees what changes before the next cycle. This is the same narrow starting point Ubisar recommends in How to Choose the First Workflow to Improve with AI.

How Ubisar would implement this workflow

In week one, Ubisar would take one recurring return and trace it from trigger to submission: source extracts, reconciliations, classification rules, adjustments, validation, reviewer sign-off, and the evidence behind the filed figures. The first output would be a filing tracker showing each source and its owner, the refresh timing, the reconciliation state, open adjustments, the reviewer, the deadline, and the sign-off status.

In weeks two and three, we would connect the minimum core system, ledger, calculation, spreadsheet, and document data needed to keep that tracker current. AI would help assemble the draft return, reconcile figures across sources, draft movement commentary, and flag missing evidence, while compliance and the accountable owner still review and sign. By week four, one return should move through the cycle with fewer last-night surprises and a trail that survives a later query.

At the end of month one, keep going if breaks surface earlier and the filed return is easier to defend, and stop or narrow it if the definitions or the accountable owners are still unresolved. This is the kind of work the AI, Data & Tech Implementation retainer is built for, because it means fixing the source data, the tools, and the sign-off together rather than any one of them alone. If you are weighing how to get help, AI consultant vs AI automation agency vs software compares the options, and if budget is the first question, What AI Implementation Costs in 2026 is the place to start. When you would rather see this on one of your own returns than read about it, tell us which return hurts most and we will map it with you. You can also browse related workflow guides for adjacent finance and reporting work.

Sources and useful references