By Sagar Shankaran, Founder of CallSphere
A cloned controller can reroute a payroll run or a Bill.com vendor payment. The verification step, dual control and engagement-letter passphrase that stop it.
Key takeaways
Run the scenario in your head before you read the rest. It is 3:12 on the second Friday of the month. Your client accounting services team is two hours from submitting payroll for eleven clients through Gusto. The phone rings and it is Dana — the controller at the client you have had for nine years, the one whose voice everybody in the bullpen knows, the one who always sounds slightly rushed. She says their bank flagged fraud on the operating account and they had to open a new one overnight. She reads out a new routing and account number, apologises for the timing, says she is walking into a meeting and will email the void cheque later.
Your CAS associate has heard that voice a hundred times. Would she change it?
In 2024 the honest answer was "probably, and nobody would have blamed her." In July 2026 that answer is a liability, because copying a voice well enough to fool someone who knows it now takes a short sample and costs almost nothing. Dana has a thirty-second clip on the company's website, on a webinar recording, on the voicemail greeting she has had since 2019.
Two things moved at once this year, and they get discussed as one. The first is that voice copying got cheap and good — not "robotic but recognisable" but good, including the pauses, the throat-clear, the way somebody says a particular client's name. A short sample is enough.
The second is that live speech-to-speech systems now hold a real conversation, answering in roughly the time it takes a person to draw breath — around two-tenths of a second — while looking something up or booking something mid-call. That is genuinely useful when it is on your side of the phone. It is also what lets a fraudulent call improvise instead of reading from a script. When your associate says "wait, which account is this for?", the caller answers correctly, because it can look it up.
Caller verification means proving the person on the phone is who they claim to be using something the voice itself cannot supply — a callback to a number already on file, a passphrase agreed in writing, or a code delivered through the client portal. Anything a good impersonator could have researched is not verification.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent for financial services in your browser — 60 seconds, no signup.
Most partners think of their firm as low-risk here because they are not a bank. Then they list what a phone call can actually cause to happen, and the list is longer than expected:
Notice that not one of these requires breaking into a system. They all work through the front door, on the phone, with a friendly voice and good timing — the second Friday, the day before the 15 September deadline, the afternoon a partner is out.
flowchart TD
A["Caller: 'This is Dana, the controller'"] --> B{"Is the request on the money list?"}
B -->|"No: send a copy of the organizer"| C["Handle the call, log it, done"]
B -->|"Yes: change bank details"| D["Ask for the engagement-letter passphrase"]
D --> E{"Passphrase correct?"}
E -->|No| F["Stop. Open a ticket. No change made."]
E -->|Yes| G["Call back the number of record, not the number that called"]
G --> H["Second approver in the firm signs off"]
H --> I["Change entered in Gusto or Bill.com under dual control"]
The whole defence is one sentence your team is allowed to say without feeling rude: "Absolutely — banking changes go through our callback step, so let me ring you back on the number we have on file and we'll get it done." Then hang up and dial the number in your practice management system. Not the number on the caller ID. Not a number the caller gives you. Not a number in the signature of the email that arrived this morning.
A real controller will be mildly relieved; she has the same problem with her own vendors. An impersonator produces friction: the number on file is "the old office line," she is "travelling," the bank needs it in twenty minutes, the payroll will be late and it will be your firm's fault. That pressure is the tell, and staff should be told explicitly that being made to feel rushed is itself the signal to stop.
Three more controls that cost nothing:
Owners want to know whether this is worth an hour of a partner meeting. Here is the arithmetic, with everything stated. Suppose your CAS group runs payroll for 40 clients, the average bi-weekly run is $86,000, and one fraudulent bank change goes through on a single client's run.
| Line item | Assumption | Amount |
|---|---|---|
| Payroll run redirected | One client, one cycle | $86,000 |
| Recovered through the bank | Assume 35% recovered — ACH recall is not guaranteed | ($30,100) |
| Emergency re-run so employees get paid | Client fronts it; firm often shares the pain | $0–$86,000 |
| Forensic review and legal | Illustrative | $12,000 |
| Firm time responding, unbilled | 60 hours at $195 | $11,700 |
| Client lost, first-year fee | Assume the relationship does not survive | $21,000 |
| Firm's exposure, one event | $100,600 |
Now weight it. If you judge the chance of one successful event across your CAS book at 4% in a year — a low estimate for a firm with no callback rule — the expected cost is about $4,024 a year. The callback step costs about four minutes per banking change. At, say, 30 such changes a year, that is two hours of staff time, call it $60. You do not need to believe the 4% to see where this lands.
No verification step survives a partner overriding it. The most common failure in this profession is not a clever attack; it is the managing partner telling the CAS associate to just process it because the client is important and it is Friday afternoon. Write the rule so it binds partners too, and say so out loud in the meeting where you adopt it.
Still reading? Stop comparing — try CallSphere live.
See the financial services AI agent handle a real call — complete, industry-specific, and live in your browser. No signup.
Technology that claims to detect a copied voice is also not something to bet a payroll run on. The copying is improving faster than the detecting. Treat any detection score as a hint, not a decision, and keep the callback regardless of what it says.
And the judgment about what is unusual stays human. A long-standing client genuinely does change banks. A controller genuinely does leave, and the new one genuinely does call with a legitimate request. The person who notices that this client has never once called about banking — and always emails from the same address, with the CFO copied — is doing something no rule captures. Keep experienced people close to the phone during busy season instead of burying them in returns.
Sometimes, partially, and often under a separate cyber or crime endorsement with its own limit and deductible rather than the main policy. Call your broker and ask specifically about social engineering and funds transfer fraud, and ask whether the policy requires a documented callback procedure — several now do, and having one on paper can be the difference between a covered claim and a denied one.
Less awkward than the alternative conversation. Introduce it at the engagement letter stage as a service you provide: "we verify anyone who calls asking us to move money, including you." In practice clients like it, because they are getting the same calls at their own office.
Same rule, different direction. The Service initiates contact by mail, and a revenue officer's visit or call follows a notice you can look up. If a call claims to be the IRS demanding a payment or asking your firm to confirm a client's banking details, verify through the number on the notice or through your practitioner priority line — never through a number the caller provides.
Some firms have, and it works if the portal is genuinely usable. The realistic middle is: instructions can arrive any way the client likes, but changes are only executed after a callback to the number of record plus a second approver. What matters is that the execution path is narrow, not that the request path is.
One paragraph, in the next batch of letters that goes out: banking and payment changes are executed only after a callback to the contact number already on file, and the firm will not act on instructions received solely by phone or email. Add a two-word passphrase field to your client record and populate it as you renew engagements through the summer. That is a week's work for the firm administrator and it closes the door that everything above walks through.
The other half of the problem is that the front desk is answering a hundred calls a day and cannot apply a rule consistently at 4:55 on a Friday. CallSphere builds AI voice and chat agents that answer the firm's phone line and website chat, book appointments and capture caller details around the clock — and because they follow the same script on every call, a request that touches money can be logged and routed to a named human instead of being handled on the spot. The agent takes the call. The decision to move money stays with a person who called back.

Written by
Sagar Shankaran· Founder, CallSphere
LinkedInSagar Shankaran is the founder of CallSphere, where he builds production AI voice and chat agents deployed across healthcare, hospitality, real estate, and home services. He writes about agentic AI, LLM engineering, and shipping voice agents that handle real calls in production.
See how AI voice agents work for your industry. Live demo available -- no signup required.
Consignee changes, bank details, hold releases and DEA calls: how a contract manufacturer verifies the caller when the voice itself proves nothing in 2026.
Client PDFs are attacker-supplied documents. How a CPA firm scopes AI agent permissions, and the irreversible tax actions that always need a named human.
Bank-detail changes, gift card activations and pallet releases all run on a voice an independent grocer trusts. Three controls that close the gap this week.
A cloned voice can redirect every payment a client's customers make. What outsourced AR teams should verify, how the callback works, and what the exposure is.
A cloned voice can ask your SOC to un-isolate a host or release a break-glass credential. The authority list, callback and client code that stop it at 2 a.m.
A cloned voice can release a repaired tractor or approve $9,000 of injector work at 9pm. The five-line callback rule a heavy-duty truck shop puts on the wall.
© 2026 CallSphere Inc. All rights reserved.
Made within San Francisco
Watch how CallSphere handles real customer calls, schedules appointments, and processes payments — live.
Try Live DemoBook a DemoCalculate Your ROI