What changes in the work?

Expert judgement needs
better-prepared work.

Your finance team already knows how to reconcile, question and decide. Agentic AI could change what reaches their desk: a draft working pack with reconciliations assembled, evidence attached and differences explained or flagged.

Follow one corrected source.

Working paper / 01

Synthetic GSTR-9C example · Aster Components · FY 2024–25 · All amounts in ₹ lakh

Preparation people currently coordinate

Find the schedule. Check the version. Locate the adjustment.
Update the working. Reconcile to the return. Revisit the review.

Agreed source · UB-01

Opening unbilled
revenue schedule

UB-01 version change, amounts in lakh rupees
Project	v1	v2
A	14	14
B	4	10
Total	18	24

v1 retainedv2 current

UB-01 / v2 / total 5B

Agentic AI Evidence & mapping

Locate agreed evidence; propose its mapping. Record source, version, change and reviewer questions.

Calculation toolsDraft / ₹ lakh

GSTR-9C · Table 5 working

Selected Table 5 workings before and after the source correction. Row descriptions are abbreviated.
Row / description	Prior	Revised
5BOpening unbilled revenue	18	24
5PAdjusted annual turnover	1,231	1,237
5QTurnover declared in GSTR-9	1,237	1,237
5RUnreconciled turnover Q − P	6	0

1,250 5A · books+24 5B−6 5E−31 5H=1,237 5P

Programmed calculations update the working. The +6 source correction is separate from the unchanged 5E credit-note deduction of 6; 5H remains 31.

Illustrated method, not an AI run or completed review. Selected rows and assumed treatment; not the complete prescribed form. Other adjustments are unchanged.

Where the effort goes

EY puts 53% of internal tax personnel’s time in routine compliance: data collection, workpapers and reconciliation.[1] This is workload, not time AI can save.

Where judgement stays

FRC found AI use concentrated in extraction, consolidation, reconciliation and validation, with caution around judgement.[2] Reported practice, not causal productivity gains.

Why this matters beyond one return

Test a capability
your finance team can reuse.

GSTR-9C is a bounded proving ground: evidence spread across files, repeated preparation, a prescribed output and results a tax professional can inspect.

Look for the same pattern elsewhere

Ledger Schedule Supporting record

Fragmented evidenceFacts across files

Working / current periodv2

Source + versionCalculationReview note

UB-01 / v25B · 24Reopened

Prior version retainedNew input → affected working

Recurring preparationPeriods repeat; inputs change

Defined outputsA known working to complete

Reviewable resultsNumbers, evidence, questions

Plausible output is not enough.

Financial Audit Bench [3]

>85%Applicable checks passedOn average, for each model

69%Task attempts fully passedBest model; every applicable check correct

Different denominators: individual checks versus complete task attempts. These are not two scores for the same measure.

FinBalance found plausible balances alongside incorrect source links and inconsistent ledger calculations.[4] Both are controlled, synthetic benchmarks—not BluePi or production evidence. Can a reviewer verify the whole working, including what the system missed?

What a BluePi evaluation should establish

Requirements, not feature claims

Trace it.: Can a reviewer reach the exact source and version behind a figure?
Check it.: Do programmed calculations reconcile, with missing evidence and unresolved treatment visible?
Change it.: When agreed inputs change, are affected workings refreshed and prior decisions flagged for review?
Review it.: Does gathering and rechecking fall without more review effort? Measure both against today’s process.

The sample illustrates a pack and correction trail, not production automation, measured savings or completed customer review. Success here would not prove another finance process.

The bigger opportunity

If these capabilities hold up, examine close reconciliations or reporting schedules. More frequent reconciliation may help investigate differences earlier. Each application needs its own evidence, controls and business case.