AI can help attackers produce clear, tailored messages at scale. That makes spelling, tone and apparent familiarity weaker tests of whether a request is genuine. Finance and IT need a process that still works when an email, chat message or call sounds convincing.
The UK's National Cyber Security Centre assesses that threat actors are already using AI to improve existing techniques, including reconnaissance and access through social engineering. Its earlier near-term assessment of AI and cyber threats explains that generative AI can create convincing lures without the language mistakes that often revealed phishing.
The practical response is to verify high-consequence requests through a separate, trusted route. The workflow below gives Finance and IT the same six steps, while keeping their responsibilities clear.
Trigger the workflow based on the requested action
Do not ask employees to decide whether a message “looks AI-generated.” That judgment is unreliable and unnecessary. Trigger verification when a request could cause meaningful harm.
Finance should pause and verify requests involving:
- a new supplier or payee
- changed bank or payment details
- an urgent or unusual transfer
- a request to bypass an approval
- payroll, tax or other sensitive financial data
IT should do the same for:
- password or multifactor authentication resets
- account recovery or new administrator access
- requests to share confidential data
- changes to security controls or access rules
- installation of software, integrations or remote-access tools
Add the triggers to the relevant procedure and interface. A warning hidden in general security guidance will be hard to recall when someone appears to be waiting for an urgent payment.
Use one six-step verification workflow
1. Pause the action
Do not pay, reset, disclose or approve while checking the request. Preserve the original message and attachments. If it arrived by phone or video, note the time, number or account and what the caller asked for.
The pause should be an accepted control, not a personal confrontation. Employees need authority to delay a consequential request even when it appears to come from an executive or important supplier.
2. Classify the request
Identify the action, the asset affected and the worst credible consequence. This determines the level of verification.
A routine invoice already matched to an approved purchase order may follow the normal payment process. A changed bank account or administrator reset needs stronger checks. Treat urgency, secrecy and pressure to bypass controls as reasons to slow down, not as proof that the sender is fraudulent.
3. Verify through a trusted channel
Contact the requester using details that existed before the message arrived. Use the supplier record, employee directory, approved ticketing system or a known internal contact. Do not use a phone number, link or meeting invitation supplied in the request you are checking.
Ask about the specific action: “Did you request that we change the account ending in 1234?” A reply in the same email thread is not independent verification. The account itself may be compromised.
4. Obtain the required approval
Use a second person for high-impact actions. The approver should see the request, the verification evidence and the proposed action. The requester should not choose the approver or redirect the process to someone outside the normal chain.
For Finance, this may be the existing payment-approval rule. For IT, it may be a service-desk approval for privileged access or account recovery. Reuse controls people already understand rather than inventing an AI-only path.
5. Execute in the system of record
Complete the action through the approved finance, identity or service-management system. Apply the minimum access or change needed. Record who verified the request, which trusted channel they used and who approved it.
Avoid sending sensitive evidence back through the suspicious conversation. If a genuine requester has lost access to the normal system, follow the documented recovery process rather than improvising in chat.
6. Report and contain suspicious activity
If verification fails, report the message through the organisation's security channel. The NCSC's phishing guidance covers recognising and reporting suspicious emails, texts, websites, adverts and calls.
Finance may need to contact the bank or payment provider quickly if money moved. IT may need to secure an account, revoke sessions, preserve logs or begin the incident process. Employees should know where their responsibility ends and who takes over.
Prepare the trusted routes before an attack
Verification fails when the correct contact details are difficult to find. Finance should maintain supplier contacts separately from invoice emails and require controlled approval for changes. IT should keep the employee directory, service desk and account-recovery contacts current.
Agree how urgent checks will work when a usual approver is unavailable. Define a backup rather than allowing the requester to name one. Make the reporting route visible in email, chat and the service desk.
The same preparation should cover senior leaders. An executive request should pass through the same control as anyone else's. That protects both the employee who pauses it and the executive being impersonated.
Test the process, not the employee's ability to spot AI
Use a short exercise that begins with a realistic request to change payment details, reset an account or share a file. Observe whether the employee pauses, finds an independent contact, records the check and escalates uncertainty.
Measure the process steps separately. A failed simulation can reveal an unclear supplier record, an unavailable approver or a reporting route that is too slow. Those are control problems the organisation can fix.
Our guide to AI-generated phishing explains how attackers use generative AI. The verification workflow is the next layer: it makes the quality of the impersonation less important because the requested action still has to pass a trusted check.
Start by choosing the three Finance and IT actions that could cause the most harm. Write the trusted contact source and approval rule beside each one, then run a tabletop exercise. A process that works under pressure offers more protection than asking employees to judge whether a polished message sounds human.
