By Shashank Pattekar, International Banker
In January 2024, an employee in the Hong Kong office of engineering group Arup authorised transfers worth HK$200 million (US$25 million) after joining what appeared to be a videoconference with the company’s chief financial officer and colleagues. Hong Kong police found that the pre-recorded meeting had been assembled from publicly available video and audio, with the face, voice and corporate authority all fabricated.
The case presents banks with a profound question: How can they verify incoming digital evidence and account for decisions produced by their own artificial intelligence (AI) when the evidence itself may be false and the decision process may be difficult to reconstruct? Answering it requires a traceable chain showing where information originated, how it changed, what model used it, what action followed and who authorised it. Digital provenance preserves that chain.
Digital provenance is the verifiable history of a digital asset or process. For a bank, this means being able to trace four parts of the chain.
Digital provenance is the verifiable history of a digital asset or process. For a bank, this means being able to trace four parts of the chain:
- Who or what created a document, recording or instruction, and was it altered?
- Where did the underlying data come from, and how was it transformed?
- Which model and version processed it?
- What decision or action resulted, including any human approval or override?
This tracing process can be supported by key tools: cryptographic hashes act like digital fingerprints; signatures associate records with particular systems or signers; trusted timestamps help establish sequence; and lineage records and logs preserve the path through data, models and actions.
While a deepfake detector estimates the probability that content was manipulated, digital provenance supplies evidence that can be checked regarding origin, recorded history, integrity and attribution. It does not certify the accuracy of the original claim.
“Provenance information can help establish the truth about the origin, history and authenticity of digital content, by providing evidence for its creation, discovery, ownership and movement over time; but provenance information alone cannot tell you whether the digital content is true, accurate or factual,” the Coalition for Content Provenance and Authenticity (C2PA), the standards body behind the open Content Credentials specification, explained. “Content Credentials can include assertions about the real-world identity of the provider of those assertions and the digital asset to which they refer.”
As AI moves deeper into financial decision-making, the need for this evidence is becoming more urgent. The 2024 Bank of England (BoE) and Financial Conduct Authority (FCA) survey of 118 firms in the UK financial-services industry found that 75 percent used AI, up from 58 percent in 2022. One-third of use cases were third-party implementations, up from 17 percent, while 55 percent involved some degree of automated decision-making. Among firms using or planning to use AI, 46 percent reported only a partial understanding of the technology.
Banks are increasingly answerable for decisions involving data, models and execution that sit partly outside their direct control. “That clearly poses challenges for financial firms’ management and governance, and for supervisors,” Bank of England Deputy Governor Sarah Breeden acknowledged. “While the way a person makes decisions based on information can also be complex, opaque and hard to explain, a person can be held to account.” By preserving the inputs, model version, outputs, approvals and actions, provenance records can give managers, auditors and supervisors a basis for investigating such outcomes.
The US Financial Crimes Enforcement Network (FinCEN) has also observed increased suspicious-activity reporting involving deepfakes, particularly fraudulent identity documents used to evade verification. “While GenAI holds tremendous potential as a new technology, bad actors are seeking to exploit it to defraud American businesses and consumers, to include financial institutions and their customers,” said FinCEN Director Andrea Gacki. “Vigilance by financial institutions to the use of deepfakes, and reporting of related suspicious activity, will help safeguard the US financial system and protect innocent Americans from the abuse of these tools.”
Banks already have systems that can support such an audit trail. The Basel Committee on Banking Supervision (BCBS) describes the BCBS 239 set of guidelines as foundational to bank data management. It defines data lineage as tracing data from its origin to its final use, which helps confirm its quality. With AI, the trail must extend to the model and configuration, retrieved material, output, reviewer and resulting action.
For an AI-assisted credit decision, a bank should be able to reconstruct the application data and transformations, the model, output, customer explanation and any approvals or overrides.
For an AI-assisted credit decision, a bank should be able to reconstruct the application data and transformations, the model, output, customer explanation and any approvals or overrides. Similar records are needed when AI verifies identity, flags payments, prepares anti-money-laundering (AML) investigations, drafts investment material or contributes to regulatory reporting. The Financial Stability Board’s (FSB) June 2026 consultation reported that some large banks used centralised AI inventories, recording each use case’s purpose, materiality, dependencies, responsible personnel and approvals. For AI agents, records may also cover identifiers, permitted tools, intermediate steps, accessed data and final outcomes.
Regulation points in the same direction. The European Union’s (EU) Artificial Intelligence Act (AI Act) classifies covered systems that assess individuals’ creditworthiness as high-risk. A July 2026 amendment postponed the relevant requirements until December 2, 2027, including automatic event logging and retaining relevant logs for at least six months. When an applicant is denied credit, the bank and technology provider may therefore need a contemporaneous audit trail showing how the system operated, rather than reconstructing the decision after a complaint or supervisory investigation.
“From my perspective, financial institutions should be specific about how they use AI and whether it is material to their business operations or legal and regulatory obligations,” Michelle Bowman, chair of the FSB’s Standing Committee on Supervisory and Regulatory Cooperation (SRC), recently explained. “That actual use and materiality will help inform the type and intensity of the governance and controls that should be applied for the specific AI deployment.”
Content Credentials can provide a tamper-evident history of a digital file, including its stated origin, recorded edits and declared use of AI. Cryptographic signatures and content bindings allow recipients to detect unrecorded alterations, while a compliant edit can generate a new credential recording the transformation. Audit logs can capture AI activity, while knowledge graphs can link events to their sources, actors and related records.
Umanitek is developing both media and agent tools. TrueSeal, for example, analyses suspected deepfakes and generates a tamper-evident certificate, while the company’s Guardian Agent anchors certificates in the OriginTrail Decentralized Knowledge Graph (DKG), making their integrity and provenance independently verifiable. Its open-source Agent Blackbox screens prompts, tool calls, commands and sensitive-file access before an agent acts, then flags and logs suspected threats or blocks confirmed ones.
“This partnership lets us pursue an ambitious goal—protecting 1 billion users from fake account attacks—while staying true to our core principle: you shouldn’t have to surrender ownership of your biometric data or personal information to be protected,” Umanitek’s chairman, Chris Rynning, said of its partnership with social-intelligence company LunarCrush. “By building on decentralized infrastructure, we can detect and track impersonation at massive scale without centralizing sensitive data. That’s how you build protection that scales without compromising privacy.”
Although Umanitek has not publicly disclosed a banking deployment to date, its technology could potentially support customer verification, fraud investigations and the auditing of AI-agent activity. Its applicability to banking, therefore, remains prospective at this early stage of development.
But provenance systems have weaknesses of their own. An attacker can strip an entire C2PA manifest; a compromised signing key can attach a valid provenance to malicious content; and detailed logs can expose sensitive customer or model information. Banks also face interoperability, retention and legacy-system costs if different providers preserve incompatible evidence.
An April 2026 preprint by Enis Golaszewski and his co-authors analysing C2PA Version 2.2 found further weaknesses: “Our formal-methods analysis reveals that generators and validators fail to agree on the signature’s trusted timestamp, enabling timestamp alterations that cast doubt on provenance. Inadequate certificate revocation policies cause conforming validators to accept manifests signed with known-compromised certificates, violating all security goals.” The authors noted that C2PA Version 2.3 adopted some proposed improvements. Even so, banks should use provenance alongside verified identity, access controls, model validation, transaction limits and human approval, applying the strongest controls whenever an erroneous or unauthorised action could create material harm.
Ultimately, the decisive test is whether a bank can reconstruct what happened. When a customer is denied credit, a payment is blocked, or an AI agent transfers information or money, the bank should be able to establish what evidence entered the system, how it changed, which model processed it, what the system did, and who authorised the outcome. As appearances become less reliable and AI gains autonomy in financial workflows, banks will depend on that evidence to investigate fraud, explain decisions and assign responsibility. Digital provenance is essential whenever an untraceable decision costs more than preserving its history.