Fiscus vs Perfios and traditional bank statement analysers.

Statement analysis has moved in three steps, and most tools on the market are stuck at the second one.

• Manual analysis. An analyst reads the PDF, keys transactions into a spreadsheet, categorises by eye and reconciles by hand. Thorough on a good day, inconsistent across analysts, and hours per case.

• Parser era analysers. OCR and templates automate extraction: transactions arrive as data, summaries get computed, turnaround drops. But categorisation runs on generic retail logic, fraud checks are shallow, and what the numbers mean for credit is still the analyst problem.

• Agentic analysis. The tool carries the work to the decision: segment aware categorisation, fraud verdicts with page level evidence, bureau cross checks, CAM ready output and a copilot that answers questions with citations. Extraction becomes an implementation detail rather than the product.

Credit is a trust business, so an honest comparison starts here. Parser era tools genuinely solved extraction: they read most native PDFs reliably, they are fast, and for a lender that only needs raw transaction data for a downstream model, they can be enough. If your book is single segment retail with clean salaried files, your analysts trust their own categorisation, and fraud screening happens elsewhere, a traditional analyser is a defensible choice.

The gap opens with business borrowers, multi account files, mixed segments and fraud exposure, which is most of Indian lending.

Extraction gives you rows. Credit needs to know which rows are income, which are obligations, which are the borrower moving money between their own accounts, and which are staged. That is categorisation, and it is where parser era tools quietly hand the work back: industry categorisation accuracy runs around 80 to 85 percent, which in practice means 30 to 45 minutes of manual reclassification per case before the numbers can be trusted.

Everything downstream inherits categorisation quality. FOIR computed on misclassified income is wrong. Obligation coverage that missed a BNPL repayment is wrong. A clean looking file whose circular credits were never netted is not clean, it is unread.

We ran the same raw statements through Fiscus and Perfios and compared categorisation transaction by transaction, about 200,000 in total. Perfios earned its position as the market leader, which is exactly why it is the benchmark worth publishing against. The comparison was carried out jointly with partner banks and lenders, whose review teams arrived at similar conclusions on their own cases.

Net result: 92.8% correct for Fiscus against 80.3% for Perfios, on identical inputs. The difference concentrates exactly where credit risk lives: loan disbursals labelled as generic transfers, bounced EMI charges read as ordinary bank charges, and subsidy credits mistaken for revenue.

Methodology: identical raw statements processed independently by both tools; outputs compared per transaction and reviewed jointly with partner banks and lenders. Perfios is a trademark of Perfios Software Solutions Pvt Ltd, which is not affiliated with Fiscus and has not endorsed this comparison. Results reflect the statements and period tested; run the evaluation on your own book.

Always, at the decision. Fiscus never auto declines: fraud flags arrive with page level evidence for an analyst to adjudicate, unusual but genuine files get a human read, and the committee makes the call with Copilot answering its questions along the way. The point of the machine layer is to make the human layer count: judgement applied to decisions, not to data entry.

Do not take the benchmark on faith, reproduce it. Pick 20 to 50 cases your team has already decided, including business borrowers, multi account files and the fraud cases you caught late. Run them through Fiscus statement analysis in parallel with your current setup, have your reviewers score categorisation and fraud flags, and time the end to end turnaround. The dashboard onboards within 48 hours, so the evaluation can start this week.