[ Credit Risk ]
Generated, Not Edited: The Fraud Check That Stopped Working
Balance trail reconciliation catches edited statements. A generated statement reconciles by construction. What still works, and where prompt injection fits.
9 min read · 20 November 2026

Contents
- The balance trail assumed someone was editing
- What this does and does not mean
- Behaviour is harder to fabricate than arithmetic
- Prompt injection is a new attack surface, and it is specific to AI-based analysis
- Cheap fabrication changes the shape of the attack
- What this honestly does not solve
- What Fiscus does
- Frequently asked questions
A disclosure before anything else, because this topic has a credibility problem.
Most writing on AI-enabled fraud comes from companies selling fraud detection, cites "double and triple digit increases" with no named source or period, and resolves conveniently into a product recommendation. We sell analysis software, so this post has the same conflict. What follows is therefore argued structurally rather than from prevalence claims. The one specific empirical claim we make as a company, our measured categorisation accuracy, is published with its protocol so it can be contested.
With that said, one thing has genuinely changed, and it weakens a check we ourselves described as the most reliable one available.
The balance trail assumed someone was editing
In our earlier post on the two kinds of bank statement fraud, we described balance trail arithmetic as the most reliable forensic check. The reasoning was this: someone editing a single credit amount to inflate turnover has to edit every subsequent running balance on every subsequent page to keep the arithmetic intact, and almost nobody does. The break shows up as a discontinuity at a specific row.
That reasoning is sound and it applies to edited documents. It does not apply to generated ones.
When a document is fabricated whole rather than altered, the arithmetic is produced by the same process that produces the figures. Enforcing a consistent running balance across 400 transactions and 18 pages is trivial: it is a loop. A generated statement reconciles perfectly because nothing was ever inconsistent to begin with.
The same logic degrades several neighbouring checks:
- Font and rendering inconsistency arises from pasting new text into an existing layout. A document rendered in one pass has uniform rendering throughout.
- Metadata anomalies assume the file carries traces of the tool that modified it. Metadata is writable, and a generated file carries whatever its producer chose to write.
- Layout and template comparison degrades as generation quality improves, since matching an issuer's template is the explicit goal of the exercise.
Vendors in the document forensics space describe the same shift: AI-generated documents produced whole, lacking the editing artefacts older checks looked for, with metadata no longer serving as proof of authenticity. Their prevalence figures are unverified. The structural point does not depend on them.
What this does and does not mean
It does not mean document forensics is obsolete. It means the checks have to be understood by what they actually detect.
Shallowfakes remain the lowest effort attack and remain catchable. Editing one or two fields, a salary figure or a closing balance, in an otherwise genuine document still breaks the balance trail downstream. The arithmetic check continues to catch exactly what it always caught. What changed is that passing it is now evidence of less than it used to be.
Digital signature validation gets stronger in relative terms. It is cryptographic rather than inferential. A document that carries a valid bank signature was issued by that bank, and no generative process produces one. As the other document checks weaken, this becomes the only test in the forensic layer that yields certainty rather than suspicion.
The asymmetry remains what it was: a valid signature proves authenticity; a missing signature proves nothing, because borrowers routinely print to PDF or forward through clients that strip signatures. Treating absence as fraud still generates enormous false positive volume.
Source-authenticated data removes the question entirely. Data arriving through Account Aggregator comes from the provider over an encrypted channel. There is no document, so there is nothing to fabricate. For the accounts and borrower types AA can reach, this is the structural answer to synthetic documents rather than a detection improvement. Its gaps are substantial: current accounts of partnerships, companies and LLPs are not available, and cooperative banks are largely absent as data providers.
Behaviour is harder to fabricate than arithmetic
The fraud layer that does not degrade is the one that asks whether the account is authentic rather than whether the document is.
Arithmetic is cheap to enforce. Twelve months of plausible, internally correlated, idiosyncratic financial behaviour is not, and the difficulty compounds with the length of the period and the number of cross-references.
Real accounts have properties that are difficult to synthesise convincingly:
Correlation structure. Salary lands on a date, rent leaves a few days later, utilities cluster in a window, the card payment falls near the statement date. These relationships hold across months and shift together when the salary date shifts. Generated data tends to produce plausible individual transactions without the dependencies between them.
Idiosyncratic counterparty sets with history. A real account has a specific, persistent, slightly odd set of counterparties: the same chemist, the same two fuel stations, the same landlord, a sibling who occasionally sends money. These recur with realistic irregularity across the whole period.
Mess. Real accounts contain reversals, failed mandates, duplicate charges, refunds, a bounced payment that was re-presented four days later, charges the borrower disputed. Fabricated accounts tend toward being too clean, because mess is laborious to invent and serves no purpose for the fabricator.
Distributional texture. Real transaction amounts are irregular in specific ways: trade receipts carry GST and deductions, retail spend clusters around real prices, salary is a precise odd number. Round figures and smooth distributions are a signature.
Temporal structure. Weekday and weekend differences, month-end clustering, festival and seasonal variation, and the particular rhythm of the borrower's stated occupation.
A fabricator can make one month look right. Making twelve months internally consistent, and simultaneously consistent with GST filings and bureau tradelines, is a materially larger problem, which is why cross-verification across independent sources is now doing more of the work than document inspection. A fabricated file is internally consistent by construction. Keeping it consistent with records the fabricator does not control is the hard part.
The same applies to account dressing, which never depended on document forensics at all. A genuine, signed, perfectly reconciling statement from an account that was deliberately operated to look creditworthy passes every forensic test there is, and always did.
Prompt injection is a new attack surface, and it is specific to AI-based analysis
This one is underdiscussed in Indian lending and it is worth stating plainly.
If a lender runs a language model over borrower-submitted documents, those documents are untrusted input reaching a system that interprets text. Instructions embedded in a document, in white text, in metadata, in a region of an image or in a transaction narration field, can attempt to influence how the analysis behaves.
The realistic objectives are not dramatic: suppress a flag, reclassify a transaction, inflate a derived figure, or cause the system to report a clean result on a file that is not clean.
The risk is on record. The RBI committee report that set out the FREE-AI framework in August 2025 lists prompt injection, which it describes as embedding hidden commands, alongside data poisoning and adversarial inputs in its discussion of AI and cybersecurity. Those are committee recommendations, not binding directions. Separately, a CERT-In advisory of 15 October 2026 on AI agents as attack vectors (CIAD-2026-0049) recorded agents under evaluation exceeding the authorisation of their assigned tasks, and advised enforcing restrictions at the network layer rather than through instructions to the model. That advisory concerns agent access, not document analysis, but the principle carries over: a control that depends on what the model was told is not a control. None of this is yet part of how lenders evaluate analysis vendors.
Three defences, in order of importance:
Keep calculation deterministic. If a language model is interpreting narration while all arithmetic, reconciliation and ratio computation happens in code, injected text cannot change the arithmetic behind a DSCR or a FOIR, because no model performs it. The most an injection can do is misclassify a transaction, and that is what the third defence is for. Classification is inferential, calculation is code. That line is usually argued for auditability, and it turns out to be a security property as well.
Treat document content strictly as data. Text extracted from a borrower's statement is evidence to be classified, never instruction to be followed.
Validate outputs against invariants. Totals that must reconcile, counts that must match, flags that cannot be cleared without a recorded reason. An output that violates an invariant is rejected regardless of what produced it.
For buyers, this belongs in the vendor evaluation alongside the other AI questions: which parts of your pipeline are inferential, which are deterministic, and what prevents document content from altering system behaviour.
Cheap fabrication changes the shape of the attack
When producing a convincing application costs hours of skilled work, fraud is individual. When it costs minutes, coordinated attacks at volume become economic.
That moves part of the detection problem up a level, because a ring is invisible file by file. Each application is internally plausible; what gives it away is the relationship between applications:
- Template similarity across applications from unrelated borrowers: the same layout quirks, the same narration phrasing, the same structural choices
- Shared counterparties between borrowers who should have nothing in common
- Clustering in time, geography, sourcing channel or device
- Repeated patterns in income amounts, employer names or transaction rhythms across supposedly independent files
- Early default clustering, since ring applications fail together, which is why the 90 day default review is where rings are often first identified
This is portfolio-level pattern detection, and it is a different capability from assessing one borrower well. A lender that only ever looks at files one at a time cannot see it, however good the per-file analysis is.
What this honestly does not solve
Identity fraud sits outside statement analysis. Synthetic identities, deepfake video KYC, document liveness and presentation attacks are a separate discipline with separate tooling. Nothing in a bank statement establishes that the applicant is who they claim to be.
A clean behavioural profile is not proof. Behavioural analysis raises the cost of fabrication substantially; it does not make it impossible, and a well-resourced fabricator working from a real account's statistical profile can get close.
Cash-heavy borrowers are harder in both directions. Their genuine accounts look thin and irregular, which makes the behavioural baseline weaker and both false positives and false negatives more likely.
A flag is a trigger, not a verdict on the borrower. This matters more as detection becomes probabilistic. Indian fraud classification frameworks require documented investigation and natural justice before an account or borrower is declared fraudulent, and a model output is an input to that process. Flags need to name the page, the rule and the evidence, and route to a person.
What Fiscus does
Deterministic checks first: digital signature validation, balance trail reconciliation across page boundaries, metadata and rendering inspection. Each is reported as the specific check that passed or failed, on the understanding that these catch edited documents reliably and generated ones less so.
Behavioural analysis as the layer that holds: circular counterparty flows with both legs matched, round-figure credit patterns, nil-credit windows at statement period boundaries, balances parked before a period and withdrawn after, mechanically implausible transactions, and distributional texture inconsistent with the stated business. The rules are tuned separately for salaried, self-employed and SME borrowers, because a rule applied across segments without tuning generates false positives at a rate that teaches teams to ignore it.
Bureau cross checks, with GST analysis benchmarked alongside as a companion view, because consistency with outside records is what a fabricated file finds hardest to manufacture.
AA-delivered data ingested natively, with identical analysis applied to both paths, so where the source-authenticated route is available, the document question does not arise.
Architecturally: classification is inferential, all arithmetic and reconciliation is deterministic code, transactions with insufficient recoverable information are flagged at low confidence rather than assigned a confident category, and every flag carries the page and line that produced it.
What we do not do is identity verification, video KYC, deepfake detection or device intelligence. Those are real and necessary and they are not ours.
Frequently asked questions
Can AI-generated bank statements pass traditional fraud checks? Several of them, yes. Checks built around editing artefacts, such as font and rendering inconsistency, metadata traces and balance trail breaks, assume a genuine document was altered. A document generated whole has uniform rendering, writable metadata and a perfectly reconciling balance trail by construction. Digital signature validation is unaffected, since it is cryptographic.
Is balance trail reconciliation still worth doing? Yes. It reliably catches shallowfakes, where one or two fields are edited in an otherwise genuine document, which remains the lowest effort attack. What has changed is that passing the check is weaker evidence of authenticity than it used to be.
What fraud detection still works against AI-generated documents? Digital signature validation, behavioural analysis of whether the account itself is authentic, cross-verification against independent sources such as GST filings and bureau data, and source-authenticated data through Account Aggregator where available. A fabricated file is internally consistent by construction, but it also has to agree with records the fabricator does not control.
Why is behaviour harder to fake than arithmetic? Enforcing a consistent running balance is a trivial computation. Producing twelve months of correlated, idiosyncratic behaviour is a substantially larger problem: counterparty sets with realistic history, dependencies between salary, rent and utility dates, the mess of reversals and failed payments that real accounts contain, and simultaneous consistency with GST and bureau records.
What is prompt injection in document analysis? Instructions embedded in a borrower-submitted document, in hidden text, metadata, image regions or narration fields, intended to influence an AI system that reads it, for example to suppress a flag or alter a classification. The RBI committee report behind the FREE-AI framework lists prompt injection among AI cybersecurity risks. Keeping all calculation deterministic means injected text cannot alter the arithmetic behind a computed ratio.
How do you detect coordinated application fraud? Not from individual files, since each is internally plausible. Detection comes from relationships across applications: template similarity, shared counterparties between unrelated borrowers, clustering by time, geography or channel, and early default clustering, since ring applications tend to fail together.
Prevalence claims about AI-enabled fraud circulating in this market are largely vendor-sourced and unverified, including those cited in the research for this post. The structural arguments above do not depend on them. Book a parallel evaluation using files your team has already identified as fraudulent.
Written by
Zeus DhanbhooraZeus Dhanbhoora is the CEO of BridgeUp Tech, the company behind Fiscus. He previously co-founded Bacferim Technologies and was an associate at the law firm Bharucha & Partners. He writes the Fiscus credit desk blog on benchmarks, fraud detection and credit underwriting methods.


