Risk Management Strategies

Zero-Human KYC: When AI Agents Replace Onboarding Staff, How Crime and Cyber Policies Respond to Synthetic-Identity and Deepfake Fraud

As Indian lenders and fintechs hand KYC and onboarding to AI agents, synthetic-identity and deepfake video-KYC losses fall between commercial crime, bankers' blanket bond, cyber and tech E&O cover. Here is where each policy responds and what underwriters expect.

Tarun Kumar Singh
Tarun Kumar SinghStrategic Risk & Compliance SpecialistAIII · CRICP · CIAFP
10 min read

Listen to this article

Audio version • 10 min read

deepfake fraudsynthetic identityvideo kyccrime insurancebankers blanket bond

Last reviewed: October 2026

The Onboarding Loss That Has No Dishonest Employee

Indian lenders and payment companies are moving identity verification out of human hands. As reported by Inc42 on 30 September 2026, PhonePe's CTO said that in some KYC and onboarding workflows, "We have been able to reduce 20 agents to zero", that voice-agent costs had fallen to about one-sixth, and that the AI follows the same role-based access rules as staff for regulated operations. Large lenders are moving in the same direction through capital rather than only through build: Inc42 reported on 5 October 2026 that Bajaj Finance has earmarked Rs 1,500 crore to back AI platforms, that HDFC Bank and Canara Bank have invested in conversational AI firm CoRover, and that Axis Bank cut 3,100 employees in FY26.

The fraud side of the ledger is moving just as fast. Business Standard reported in June 2026 that an NBFC lost about Rs 15 to 20 crore after approving loan applications built on synthetic bank statements and deepfakes in video KYC. Nothing in the report points to an insider or a breached server. On the reported facts, the onboarding process did what it was designed to do, and the applicant who passed it did not exist in the form presented.

That combination is the problem this post addresses. Commercial crime cover, the fidelity guarantee heritage it grew from, the bankers' blanket bond and cyber policies were all drafted around a person or a system doing something wrong. When the verifier is an AI agent working exactly as configured, and the loss surfaces months later as a written-off loan, the classic triggers stop matching the facts. This is a different question from the payment-diversion scams covered in our earlier piece on deepfake payment fraud and crime and cyber cover. There the loss is money moved out of a treasury account. Here the loss is credit extended to a fabricated borrower.

How Synthetic-Identity and Deepfake KYC Losses Actually Form

A synthetic-identity onboarding loss is rarely a single event. It is a pipeline of small deceptions that each pass a control:

  1. Document fabrication. Bank statements, salary slips and GST returns are generated or edited so that income and cash flow look consistent with the loan amount requested.
  2. Identity assembly. A real identity number may be paired with a different face, or an entirely fabricated profile may be seeded with a thin but clean credit footprint.
  3. Liveness and face match. During video KYC, a deepfake or face-swap stream is injected into the session so that the face on camera matches the document photograph.
  4. Disbursal and servicing. The loan is disbursed into a mule account. The first few EMIs may even be paid to keep the account out of early-delinquency monitoring.
  5. Bust-out. Repayments stop across a cluster of accounts at roughly the same time, and the lender discovers that the borrowers cannot be traced.

The Business Standard case fits this pattern: synthetic statements to pass the credit test, deepfakes to pass the identity test. The financial loss is recognised as credit loss, not as theft, and that accounting treatment shapes the insurance argument from the first day.

Why automation changes the shape of the loss

When 20 human agents review applications, a fraud ring has to beat 20 slightly different judgments, and inconsistencies get noticed. When a single model reviews every application, a technique that beats the model once tends to beat it every time until the model is retrained. Losses therefore arrive in clusters, and the aggregate from one technique can be far larger than any individual file. That matters for how limits, deductibles and the definition of a single loss are written.

Where the Loss Lands: Crime, BBB, Cyber and Tech E&O

Each of the four policies a lender or fintech is likely to hold responds to a different story. The task for a risk manager is to know which story the facts will tell before a claim is lodged.

Commercial crime and the bankers' blanket bond

A commercial crime or fidelity programme traditionally indemnifies direct loss caused by employee dishonesty, plus named third-party insuring clauses such as forgery, computer fraud and funds-transfer fraud. A bankers' blanket bond (BBB), discussed in our guide to bankers' indemnity and blanket bond cover for financial institutions, has a similar architecture with clauses suited to banks and NBFCs, often including a forged or counterfeit documents clause.

The difficulty is that many bond and crime wordings carry a loan or credit exclusion: loss arising from the extension of credit is excluded unless it results directly from employee dishonesty or from reliance on a forged original document of a specified type. A synthetic PDF bank statement is not obviously a forged original of anything; it may be a wholly fabricated document. A deepfake video is not a document at all. If the onboarding was automated, there is no employee whose dishonesty can be pleaded. On a strict reading, the classic triggers simply do not reach the loss.

Cyber and technology E&O

Cyber policies are triggered by a security failure, unauthorised access, or a privacy event. Injecting a deepfake stream into a video-KYC session is arguably a manipulation of the lender's system, but most insurers will say the system was used exactly as intended by an external party presenting false information. That is social engineering, and first-party cyber wordings typically either exclude it or cap it at a low sublimit. The loss is also credit loss, which many cyber wordings exclude outright.

Technology errors and omissions cover belongs to the vendor, not the lender. If a third-party KYC or liveness provider's product failed to detect a deepfake it was contracted to detect, the lender's recovery route may be a claim against that vendor, which then tests the vendor's professional indemnity or tech E&O policy. Contract terms, liability caps and service-level definitions decide how much of that recovery is real.

Why Classic Triggers Stop Fitting Once Humans Leave the Loop

Insurance wordings carry decades of assumptions about who does what inside a financial institution. Removing people from onboarding breaks several of them at once.

  • Employee dishonesty needs an employee. An AI agent is not an employee under most definitions. If a staff member manipulated the model or its thresholds, that conduct can still be pleaded, but ordinary model error is not dishonesty.
  • Forgery needs a forged instrument. Many forgery clauses were written for cheques, drafts, securities and specified title documents. Synthetic income documents and video streams fall outside the listed instruments in older wordings.
  • Computer fraud needs fraudulent entry or change. Some wordings require that the fraudster entered or altered data in the insured's system without authority. An applicant uploading their own fabricated documents through a public onboarding app is, from the system's view, an authorised user doing an authorised thing.
  • Discovery needs someone to discover. Bonds and crime policies are usually written on a discovery basis. When losses surface as delinquency in a portfolio analytics report, the date of discovery can be argued, which affects which policy year responds and whether notice was timely.
  • Single-loss definitions assume one actor. A ring that beat the model hundreds of times may be one loss or hundreds of losses depending on the related-acts wording. One loss means one deductible but also one limit.

PhonePe's point that AI follows the same role-based access rules as staff is relevant here. It shows that the controls framework has been carried over to the automated workflow. Insurance wordings mostly have not been carried over in the same way. That mismatch is where uncovered losses sit.

What Underwriters Now Expect From Automated Onboarding

Underwriters pricing crime, BBB and cyber for lenders and fintechs are beginning to ask about onboarding in the same detail they ask about payments controls. Expect questions in at least these areas during renewal:

  • Liveness and injection detection. Whether the video-KYC stack can detect virtual camera injection and replayed or synthetic streams, not only whether a face is present and moving.
  • Document provenance. Whether bank statements are pulled through account aggregator or direct bank feeds rather than accepted as uploaded PDFs, and whether metadata and template checks run on any uploaded document.
  • Model governance. Who owns the KYC model, how thresholds are changed, whether changes are logged, and how quickly a newly observed fraud pattern can be blocked across the portfolio.
  • Human escalation. Which cases are routed to a person, on what triggers, and whether reviewers can override the model. A zero-human workflow for low-ticket loans with human review above a value threshold is easier to underwrite than zero humans at every ticket size.
  • Early-delinquency analytics. Whether first-payment default and cluster analysis are monitored as fraud signals, not only as credit-risk signals.
  • Vendor contracts. Detection obligations, liability caps, and audit rights in contracts with KYC and liveness providers.

Lenders building or buying AI through investment arrangements, as Bajaj Finance, HDFC Bank and Canara Bank are doing, should expect the underwriter to ask whether a related-party vendor is accountable on the same terms as an arm's-length one.

Closing the Gap: Wording Fixes and Programme Design

There is no single product that cleanly covers synthetic-identity credit loss in India today, but a programme can be shaped to capture more of it. The practical levers are in the policy wording and in how layers fit together.

  1. Ask for explicit synthetic and deepfake language. Seek amendments that extend forgery or fraudulent-document clauses to electronically generated documents and to audio or video impersonation used in identity verification.
  2. Narrow the loan exclusion. The goal is a carve-back for loans induced by fraudulent identity or documents, even where no employee was dishonest. Expect a sublimit and a higher deductible.
  3. Align the cyber social-engineering sublimit. If the cyber policy has a social-engineering or impersonation extension, confirm whether it applies to onboarding and not only to payment instructions.
  4. Fix the aggregation clause. Decide whether you want a ring-based attack treated as one loss or many, and negotiate the related-acts wording accordingly with both your crime and cyber insurers so they do not point at each other.
  5. Push risk back up the vendor chain. Require KYC and liveness vendors to carry tech E&O at a meaningful limit, and to name detection failure in their contractual indemnity.
  6. Price the residual as credit risk. Some synthetic-identity loss will remain uninsured. Provision for it in credit models and treat it as an expected loss rate, not an insurance event.

Lenders with a digital-only origination model should read this alongside our wider note on insurance for digital lending NBFCs and fintechs, which covers D&O and regulatory exposures that sit beside the crime question.

The Policy Direction and What Comes Next

Cyber fraud is now discussed at the highest political level. Inc42 reported on 8 October 2026 that Prime Minister Modi called for a global framework to tackle cyber fraud. Calls of this kind usually translate, over time, into tighter expectations on regulated entities around identity assurance, reporting and cross-border cooperation. Those expectations then show up in proposal forms and claims reviews.

For Indian lenders the near-term priorities are practical. Map each onboarding channel to the policy that would respond if it were beaten by a deepfake. Test that mapping with your broker against the actual wordings, not the policy summaries. Capture evidence of how each fraudulent application entered the system, because the route of entry determines whether the loss is pleaded as forgery, computer fraud or vendor failure. And keep a human reviewer in the loop at the ticket sizes where a single cluster of losses would hurt the balance sheet.

Automation of KYC is not going to reverse. The cost case PhonePe described, voice agents at roughly one-sixth of the earlier cost, is too strong. The insurance market will adjust, but buyers who wait for standard wordings to catch up will carry the gap themselves in the meantime.

About the Author

Tarun Kumar Singh

Tarun Kumar Singh

Strategic Risk & Compliance Specialist

  • AIII
  • CRICP
  • CIAFP
  • Board Advisor, Finexure Consulting
  • Developer of the Behavioural Underinsurance Risk Index (BURI)

Tarun Kumar Singh is a seasoned risk management and insurance professional based in Bengaluru. He serves as Board Advisor at Finexure Consulting, where he advises insurance, fintech, and regulated firms on governance, growth, and trust. His work spans insurance broker regulatory frameworks across India, UAE, and ASEAN, IRDAI compliance and Corporate Agency model reform, VC governance in insurtech, and MSME insurance gap analysis. He is the developer of the Behavioural Underinsurance Risk Index (BURI), a framework applying behavioural economics to underinsurance and insurance fraud risk.

Frequently Asked Questions

Does a bankers' blanket bond cover loans approved on deepfaked video KYC?
Often not in its standard form. Most bonds exclude loss from the extension of credit unless it results from employee dishonesty or reliance on a forged original document of a listed type. A deepfake video and a synthetic PDF statement may not meet either test, so lenders should seek explicit synthetic-document and impersonation language at renewal.
Will our cyber policy pay if a fraud ring injects deepfake video into our KYC app?
It depends on the wording, but many insurers will treat this as social engineering by an external party using the system as intended, rather than a breach. That usually means an exclusion or a low social-engineering sublimit, and many cyber wordings also exclude credit losses.
Can we recover from our KYC or liveness vendor?
Possibly, if the contract obliged the vendor to detect deepfakes or injection attacks and the failure can be shown. Recovery then depends on the contractual liability cap and the vendor's technology E&O or professional indemnity limit, so review both before you sign.
Is a fraud ring that beat our model hundreds of times one loss or many?
That is set by the related-acts or aggregation clause. Treating the ring as one loss means one deductible but also one limit; treating it as many means many deductibles. Agree the position with both crime and cyber insurers in advance so they do not each point to the other.
What controls do underwriters want to see for automated onboarding?
Injection and liveness detection in video KYC, document provenance through direct bank or account aggregator feeds, logged model governance, human escalation above set ticket sizes, early-delinquency fraud analytics, and vendor contracts with clear detection obligations.

Related Glossary Terms

Related Insurance Types

Related Industries

Related Articles

Sarvada Intelligence

Ready to see Sarvada in action?

Explore the platform workflow or start a product conversation with our underwriting automation team.

Explore the platform