Decentralised News · 2027 planning series · Sources reviewed 22 November 2026
What Matters
A familiar face can make a payment request convincing without making it authorised. For 2027, businesses need a transfer process that verifies the request through an independently maintained contact route, checks the beneficiary separately and binds approval to the exact payment. DN’s worksheet measures whether those records exist; it does not detect deepfakes or certify that a payment is safe.
The meeting ended. The payment instructions continued.
In January 2024, a finance employee in Hong Kong received an email apparently from a UK chief financial officer, joined a fabricated conference and authorised transfers to five local accounts. The loss was about HK$200 million. The Hong Kong government later explained a detail missing from many retellings: investigators believed the meeting was prerecorded, with no interaction between the employee and fraudster. Instructions continued through instant messaging after the video ended. That is the account in the Security Secretary’s 26 June 2024 Legislative Council reply.
The case is useful because it separates three events that a convincing story can collapse into one. Someone appeared to give an instruction. An employee believed it. A payment system accepted an authorised user’s actions. The visual deception supplied credibility; the money moved through a business process.
The operational question is therefore precise: what independent evidence must survive between a persuasive request and the release of funds? A finance team can answer that question even when it cannot reliably identify every manipulated face or voice.
Evidence Block
Reviewed 22 November 2026 for the 2027 planning edition. The documented case above is a historical Hong Kong government account, not a reconstruction of private conversations or a claim about the investigation’s present outcome. The FBI and IC3 provide the US prevention and reporting references below. DN has not tested a deepfake detector, inspected the victim company’s controls or measured a loss-reduction rate. The control map, examples and worksheet are DN’s proposed operating framework.
1. Separate identity, instruction and authority
Identity asks who is communicating. Instruction asks what that person wants done. Authority asks whether the business permits that person to make that commitment under these circumstances. They require different evidence. A real chief financial officer can ask for a payment outside an approved delegation; a correctly signed employee message can repeat a supplier’s fraudulent bank details.
DN proposes that every exceptional transfer carry a compact decision record: the requesting role, the business obligation, payment amount and currency, beneficiary version, approval requirement, verification method and final release identifier. Each item should point to an existing controlled system or retained document. The record need not reproduce sensitive bank details in a shared chat.
The distinction matters when AI assistants prepare work. A meeting summary can accurately transcribe an impersonator. An invoice extractor can perfectly capture a criminal’s destination account. Extraction quality does not establish permission to pay. Give those outputs a preparatory status until the relevant person or system verifies the necessary evidence.
2. Make the callback route independent
The FBI’s business email compromise guidance recommends independently finding a company’s telephone number, verifying payment requests and checking changed account details. The principle is independence: a fraudster who supplies both the instruction and the verification contact controls both sides of the apparent check.
For a business, DN would implement that principle through a maintained contact register rather than an improvised search during an urgent transfer. Record who owns each supplier or executive contact, when the entry was verified, how changes are approved and what backup route exists. A callback to the number in the suspicious message is a continuation of that message, not a separate check.
Independence also requires examining account recovery. If the same compromised inbox can change the directory entry and receive the confirmation, the workflow has two screens but one trust source. Similarly, an incoming call displaying a familiar number should not automatically satisfy a requirement to initiate an independent verification. Record how the connection was established.
This does not make telephone calls infallible. It makes the failure assumptions visible. For an unusually large or unfamiliar transaction, the business may require additional confirmation through its existing authenticated approval system and a separate authorised reviewer.
3. Verify the beneficiary as its own decision
Payment approval and beneficiary maintenance should have separate records. A manager may approve a legitimate invoice while assuming the supplier’s bank account has already been checked. An accounts clerk may update that account while assuming a manager has verified the change. The gap appears when both assumptions pass silently.
DN’s proposed record associates the proposed payment with a specific beneficiary version. A version means the approved account details and the verification evidence in force at a recorded time. If material destination details change, the transfer returns to the relevant verification and approval stage. An earlier approval should not quietly attach itself to a newly edited destination.
The IC3 BEC resource specifically recommends secondary channels or two-factor authentication to verify requests for account-information changes. That recommendation supports treating the change as an event requiring its own check. It does not promise that any particular combination of channels guarantees authenticity.
For a crypto payment, the same business question remains even though the destination format differs. Keep the asset, network and full intended address within the controlled approval record. A successful test transfer proves that a route can receive value; by itself, it does not prove that the person controlling the route is the intended commercial counterparty.
4. Preserve independence between reviewers
Two approvals are stronger when the reviewers establish different facts. If both watch the same suspicious clip and click the same reassuring summary, the second click adds little new evidence. DN recommends assigning explicit review responsibilities: one person verifies the business obligation and requesting authority; another checks the approved beneficiary record and the exact release details.
A small business may not have two finance employees available at all times. It can still define who provides an independent check, which transactions wait and what documented emergency arrangement applies. The weak point is an exception that invents its own approver during the incident. The backup route should exist before somebody invokes urgency or secrecy.
Where staffing makes full separation impossible, record the limitation instead of describing the process as equivalent. Reduce the action scope, preserve the evidence and select a review arrangement appropriate to the consequence. This is a governance choice for the organisation, not a universal numerical threshold supplied by this article.
5. Bind the check to the transfer that actually leaves
The final review should display the fields the payment system will use. Amount, currency, destination, relevant network, payment reference and timing should match the approved record. A vague approval such as “please handle the supplier invoice” is difficult to compare with an actual outgoing transfer.
Do the comparison again after a material edit. A change in currency can alter economic value even if the displayed number stays constant. Splitting a payment can also change which approvals appear necessary. Track the related business obligation and cumulative releases so that one request does not become several unrelated-looking transactions.
DN’s existing AI Agent Payment Security framework examines technical binding between an agent’s intent and execution. This article addresses the human corporate process immediately before release: who was independently contacted, what beneficiary evidence was checked and which people authorised the exact transfer.
DN Payment Verification Record
Use the local worksheet to record five evidence gates: a documented obligation and authorised requester; independent confirmation through a previously maintained route; verified beneficiary details; the required independent review; and a match between approved and final payment fields. Each gate has three states: verified, failed or unknown.
A failed gate produces a hold. Any unknown gate produces an incomplete record. Five verified gates produce “records complete for the defined checks”, not “safe to pay”. The tool cannot inspect the evidence, contact a bank or detect a compromised directory. A payment remains subject to the organisation’s own authority and release process.
Use aliases and references to secure internal records. Do not enter a bank account number, wallet secret, identity document, confidential invoice or recording into the public worksheet. The useful output is the missing check and its responsible owner.
DN Payment Verification Record
Record five evidence gates and estimate the monthly verification workload. Unknown evidence stays unknown. Use aliases; do not enter bank details, private documents or credentials.
This instrument runs locally in your browser. It does not connect to accounts, fetch evidence or submit instructions. Use aggregate values and non-sensitive references. Downloaded records contain your inputs.
Machine-readable illustrative record
An illustrative record of the example assumptions, not an observed transaction or provider test. A null calculation timestamp identifies this static example.
{
"schema_version": "1.0.0",
"article_id": "N001",
"evidence_basis": "illustrative",
"source_review_date": "2026-11-22",
"computed_at": null,
"inputs": {
"obligation": "unknown",
"callback": "unknown",
"beneficiary": "unknown",
"review": "unknown",
"binding": "unknown",
"requests": "120",
"minutes": "8",
"hourly": "45"
},
"status": "Incomplete: evidence remains unknown",
"outputs": {
"methodology": "DN payment verification record v1.0",
"self_reported": true,
"state": "incomplete",
"gates": {
"obligation": "unknown",
"callback": "unknown",
"beneficiary": "unknown",
"review": "unknown",
"binding": "unknown"
},
"unknown_gates": [
"Obligation and requester",
"Independent confirmation",
"Beneficiary verification",
"Independent review",
"Approval-to-transfer match"
],
"failed_gates": [],
"monthly_requests": 120,
"minutes_per_request": 8,
"hourly_labour_cost": 45,
"verification_hours": 16,
"labour_cost": 720,
"authorises_payment": false
}
}A worked example: the invoice is genuine, the destination is not established
Hypothetical example. A business owes a supplier 48,000 currency units. The invoice matches an accepted order. A video message apparently from an executive says the supplier has changed banks and asks for immediate release. The finance employee verifies the amount and receives a second approval in the same conversation.
In the proposed record, the obligation can be verified while the independent callback and beneficiary change remain unknown. The second approval does not repair those missing facts merely because it arrived from another participant. The output is an incomplete record and a hold under the business’s stated release policy.
If the team reaches the supplier through its independently maintained route, verifies the account change, obtains the required separate approval and confirms the final payment fields, the record can become complete. This demonstrates the logic of the checks. It does not establish that the original video was genuine, or that no other compromise exists.
The example also explains why an aggregate safety score is inappropriate. Four completed checks out of five should not become an “80% safe” payment. A decisive missing destination check can be sufficient to stop release regardless of how much other paperwork is present.
The cost question: how much verification work should you plan?
A control that nobody has time to perform becomes a ceremonial control. DN’s companion arithmetic estimates the workload of verifying the exceptional-payment queue. Let monthly exceptional requests be N, additional verification minutes per request be M and loaded reviewer cost per hour be W. Planned monthly verification hours are N × M ÷ 60, and direct labour cost is those hours × W.
Suppose a team handles 120 exceptional requests a month, with eight additional minutes per request and a loaded cost of 45 currency units an hour. The planned burden is 16 hours and 720 units. If the actual queue requires 18 minutes per request, the burden rises to 36 hours and 1,620 units. These are illustrative workload figures, not estimates of losses prevented.
Do not convert the result into an investment return by inventing a fraud probability. The calculation helps a manager allocate staff and redesign avoidable friction. It cannot tell the manager how often an attack will succeed. Actual benefits require observed incident and near-miss records, credible comparisons and a clear definition of what would have happened without the control.
Separate elapsed delay from labour time. A callback can require five minutes of work and two days of waiting for an authorised contact. The operating design needs both measures because a team may otherwise remove a control to solve a service-delay problem.
Run a drill that tests the process
A useful exercise does not require cloning an executive’s face. With the organisation’s authorisation, a facilitator can present a labelled simulation: an urgent request, a changed beneficiary and a familiar-looking message. The participants practise finding the correct record, initiating the independent check and explaining why release is blocked or permitted.
| Exercise | Evidence the team should produce | Failure to record |
|---|---|---|
| New beneficiary on a genuine invoice | Independent supplier verification and beneficiary-version approval | The invoice’s authenticity is used to justify the new account |
| Executive requests secrecy | The pre-existing confidential-payment procedure and authorised reviewer | Secrecy removes the normal control owner |
| Approver is unavailable | An approved backup or an explicit wait decision | An improvised substitute approves their own exception |
| Payment details change after approval | Renewed comparison and approval of the altered fields | The old approval is reused for a different transfer |
| Request arrives near the bank deadline | A recorded choice under the established urgency policy | A cutoff becomes automatic permission to bypass verification |
Measure whether the team found the right route, whether evidence was independent, how long the process took and which step remained unresolved. Do not publicly rank employees by whether they spotted a visual artefact. The operating objective is to make a persuasive fraudulent request unable to satisfy the release procedure.
After the drill, fix the directory, backup delegation or payment-system display that caused the problem. Repeating awareness training cannot compensate for a supplier contact register that nobody owns.
Where detection fits, and where it does not
A deepfake detector can provide an additional signal for investigation. DN would evaluate it separately from payment authority: what media types were tested, how false positives and false negatives were measured, whether compression affected results, and whether performance persists on new attack methods. This article provides no detector ranking because it contains no comparative detector experiment.
Keep a detector’s output labelled as a signal rather than a release command. A clean result should not override missing beneficiary evidence. A suspicious result should trigger the established review route without requiring a finance employee to prove which model produced the media.
The same applies to ad hoc authenticity games. Asking a person to turn their head may supply context, but a satisfactory performance is not the business’s authorised payment record. The organisation can continue to require independent confirmation and its ordinary approval controls regardless of how plausible the screen looks.
If money has already moved
The IC3 incident guidance advises contacting the originating financial institution promptly to request a recall or reversal and filing a detailed complaint. The FBI also advises immediate contact with the financial institution. These are US references; businesses elsewhere should use their bank’s fraud channel and the appropriate local reporting authority.
Preserve original messages, timestamps, approval records and transaction identifiers through the organisation’s incident process. Do not keep following the suspected attacker’s recovery instructions. A request to pay a new “release fee” is not a bank recall procedure. For crypto transfers, contact the relevant service provider promptly where one is involved; no public worksheet can reverse an onchain transfer.
A recall request is an attempt to recover funds, not a promise of reimbursement. Assign one incident lead, coordinate payment and security teams, and retain the distinction between what was requested, what the bank accepted and what was actually recovered.
DN Alpha Thesis
As persuasive synthetic media becomes easier to distribute, businesses may place more value on evidence that a payment process can independently verify. The commercial opportunity is not limited to spotting altered pixels. It includes maintained counterparty records, usable approval systems, controlled beneficiary changes and clear incident evidence.
This is a conditional operating thesis, not a forecast of a particular vendor’s revenue. It would weaken if independent checks added little protection in representative testing, or if their cost and delay consistently exceeded the benefit under a defined risk model. Teams should measure the complete workflow rather than assume that another authentication screen creates independent evidence.
Methodology, scope and maintenance
DN reviewed primary government accounts and prevention guidance, then designed a proposed decision record around a corporate payment-release problem. The historical case is used to explain the distinction between appearance and authority. We do not attribute unreported internal failures to a named company, claim that a specific control would certainly have stopped the incident, or treat a historical investigation update as a current verdict.
The worksheet performs deterministic status logic and workload arithmetic. Inputs are self-reported. It has no connection to bank, identity, accounting or wallet systems. Evidence completeness is limited to the selected gates and does not establish the authenticity of the underlying records. The organisation must define its own approvers, thresholds, retention arrangements and lawful operating process.
Review the record after changes to banking permissions, supplier onboarding, directory ownership, staff roles or AI tools that process payment requests. Retain the date and owner of each review. A stale callback number or a departed backup approver can turn yesterday’s documented process into today’s unresolved exception.
For related controls, see DN’s property payment workflow and the labour cost of human rescue. This edition contains no affiliate recommendation, paid detector ranking or security certification.
Frequently asked questions
Can a video call prove that a payment is authorised?
No. Appearance, identity, authority and the final payment details require separate evidence. The business should apply its established verification and approval process even when a caller looks familiar.
Does the DN worksheet detect deepfakes?
No. It records five self-reported evidence gates and calculates verification workload. It does not inspect media, contact counterparties or certify a payment.
Why is a callback different from replying to the request?
A callback can provide an independent check when the team initiates it using a previously maintained, separately controlled contact route. A number supplied in the suspicious request does not provide that independence.
What does a complete record mean?
It means the user marked all defined evidence gates verified. It does not prove those records are authentic or grant authority to pay. A failed gate produces a hold; an unknown gate remains incomplete.
Are two approvers always independent?
No. If both rely on the same compromised message or media, the second approval may add no independent evidence. Define what each person verifies and how their evidence is obtained.
What should a business do after a suspected fraudulent transfer?
Contact the originating financial institution promptly through its established fraud channel and follow the applicable local incident and reporting process. Preserve original evidence. A recall request does not guarantee recovery.
Primary sources and review record
- Hong Kong Government: LCQ9, Combating frauds involving deepfake. Historical January 2024 prerecorded conference, five local accounts and approximately HK$200 million loss; update published 26 June 2024. Do not represent as current investigation status. Reviewed 2026-11-22.
- FBI: Business Email Compromise. Independent contact, payment/account-change verification, urgency caution and immediate financial-institution contact. Reviewed 2026-11-22.
- IC3: Business Email Compromise. Secondary verification of account changes; contact originating institution for recall/reversal and file detailed complaint. US-specific reporting reference. Reviewed 2026-11-22.




