By Sagar Shankaran, Founder of CallSphere
Clearances, hold-offs and account changes move by phone at utilities. How cloned voices abuse that in 2026, and the tag word plus callback that stops it.
Key takeaways
Here is the question, and it does not have a comfortable answer. At 11:20 on a Thursday night, a voice comes into the switching centre on the crew line. It says it is Danny, foreman on the crew working the tap off the Millbrook 12 kV circuit. It gives the clearance number. It sounds exactly like Danny — same clipped way of reading numbers, same slight rasp, same habit of saying "copy that" twice. It asks the system operator to release the clearance so the circuit can be picked back up before midnight.
How does the operator know it is Danny?
Until recently the honest answer was: because it sounds like Danny, and because nobody could fake that. That second half stopped being true.
More than most executives realise, and the list splits into two very different piles.
The first pile is money and property. A large commercial customer's authorised contact calls credit and collections to change the bank account autopay draws from, or to redirect a deposit refund, or to release a pending disconnection on a delinquent industrial account. A contractor calls to schedule a temporary service disconnect at a job site. Each of those changes something in the customer information system, and each is routinely handled by voice with knowledge-based questions — account number, service address, last payment amount — all of which are sitting in a breach dump somewhere.
The second pile is the one that keeps operations people awake. Clearances and hold-offs. A crew calls the distribution system operator to request a clearance on a section, the operator opens the switching order, the tag goes on, and the crew works the line believing it cannot be energised. Later, the crew calls back to release. Automatic reclosing gets blocked and unblocked by phone. Standby generation gets cut in and out by phone. Under a public safety power shutoff, re-energisation approvals move by phone at speed because everything else has failed. Every one of those is an action authorised by a voice on a line, in a room where the person on the other end is usually taken at their word once the numbers match.
Voice cloning is the ability to copy a person's voice from a short recording — a voicemail greeting, a clip from a public rate hearing, thirty seconds of a safety video — closely enough that people who know them well cannot tell the difference on a phone line. In 2026 it is cheap, fast and good, which means "it sounded like him" is no longer evidence of anything.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent in your browser — 60 seconds, no signup.
Utilities already have a discipline for exactly this moment. Three-part communication — the sender states the instruction, the receiver repeats it back, the sender confirms — is required for operating instructions on the bulk system and applied by most companies on distribution switching too. It is a good practice and it has prevented a great many accidents.
But look at what it actually protects against: mishearing. It confirms the instruction was understood correctly. It does not confirm who gave it. A cloned voice repeats back perfectly, because repeating back is the easiest part of the call to fake. The same goes for the numbers everyone treats as secret-ish: clearance numbers, switching order numbers, circuit designations and crew names circulate in emails, on printed switching orders and in job briefings with contractors. They are identifiers, not secrets.
Ten years ago the equivalent risk was an email that looked like it came from the CFO. The defence was procedural: call the person back on a number you already had. That worked because the fraud could not follow you onto the phone. Now it can. That is the whole change, and it is why any control ending in "verify by phone" needs a second look this year.
The other change is speed of preparation. Building a convincing voice used to take a long clean recording. Now a public commission hearing, a storm restoration press conference, or a lineman's own social media video is enough source material. Utilities are unusually exposed here, because so many of your people speak on the record in public proceedings.
flowchart TD
A["Caller asks system operator to release a clearance"] --> B{"Clearance number matches a live hold tag?"}
B -->|No| C["Refuse. Log the attempt. Notify switching supervisor"]
B -->|Yes| D{"Callback to the number on today's crew roster answers?"}
D -->|No| E["Tag stays on. Operator dispatches supervisor to the site"]
D -->|Yes| F{"Foreman gives the tag word issued when the tag was hung?"}
F -->|No| G["Treat as impersonation. Security and operations management notified"]
F -->|Yes| H["Operator releases clearance, repeats back, logs the word used"]
The fix is not a gadget. It is two additions to a procedure your operators already follow.
First, a tag word. When the clearance is issued and the hold tag goes on, the operator generates a single-use word or number pair, writes it in the switching log next to the clearance, and gives it to the foreman, who writes it on his copy. Release requires it. It is issued once, used once, and it is never spoken outside those two calls. A cloned voice can get the clearance number from a printed switching order; it cannot get a word that only exists in the operator's log and on one crew's paperwork.
Second, a callback. Not to the number that called in — caller identification is trivially faked — but to the number recorded on today's crew roster or in the work management system for that job. Two rings and a confirmation. Twenty seconds, and it defeats the entire class of attack, because the fraudster does not control the crew's phone.
On the customer side, the equivalent rule is blunter: no change to bank details, payout instructions or mailing address on a commercial account happens on an inbound call alone, ever. The rep takes the request, and the confirmation goes out to the contact method already on file, with a hold period before the change takes effect. Your call centre representatives will tell you this creates friction. It does. That is the point.
Do the arithmetic honestly, because the safety side and the money side have to be argued differently.
Still reading? Stop comparing — try CallSphere live.
CallSphere ships complete AI voice agents per industry — 14 tools for healthcare, 10 agents for real estate, 4 specialists for salons. See how it actually handles a call before you book a demo.
| Assumption | Value |
| Clearance issues and releases per year | 14,000 |
| Extra seconds per call, tag word plus callback | 20 |
| Additional operator time per year | 78 hours |
| System operator cost per hour, loaded | $79 |
| Annual cost of the control | $6,162 |
| Attempted fraudulent account-change calls per year (illustrative) | 8 |
| Share that succeed today | 25% |
| Average loss per successful change | $18,400 |
| Expected annual loss on the money side alone | $36,800 |
So on money alone the control pays for itself six times over. But note what is deliberately missing from that table: no dollar figure on the clearance side. A single wrongly released clearance with a crew still on the conductor is not an expected-loss calculation, and anybody who presents it to you as one is doing it wrong. The money argument is what gets the change funded. The safety argument is why you would do it even if the money argument failed.
Be honest about the failure mode you are creating. During a major storm restoration with mutual aid crews from three utilities, foremen who have been awake for nineteen hours will lose the slip of paper with the tag word on it. If the procedure has no defined path for that, operators will start waving it through under pressure, and you will have taught your control room that the rule is optional — which is worse than never having it.
So write the exception before you need it: no word, no release by phone; a supervisor goes to the site, or the crew comes to a known radio channel, or a second known contact vouches in person. Slow, deliberately. And when a call fails verification, the operator's job is not to investigate. It is to refuse, keep the tag on, and escalate. Judgement about whether a caller is a fraudster or an exhausted foreman is not a control-room decision at 11:20 p.m.
The other thing that stays human: the debrief. Every refused call gets reviewed by the switching supervisor the next morning. That review is how you find out whether you are under attack or whether your own procedure is broken.
Detection tools exist and they are improving, but treating them as the control is a mistake — the cloning side improves at least as fast, and a tool that is right most of the time is not something to hang a crew's life on. A single-use word plus a callback to a number you already hold does not depend on winning that race.
Knowledge-based questions fail against anyone who has done homework, and utility identifiers are unusually easy homework: account numbers appear on bills, circuit and switching order numbers appear on paperwork handled by contractors, and names and roles appear in public proceedings. The test has to be something issued for that one transaction, not something knowable.
That is the same technology pointed the other way and it is already happening: a cloned, confident voice telling a restaurant owner at 4 p.m. on a Friday that the power goes off in an hour unless a prepaid card number is read out. You cannot stop the calls, but you can make your own rule loud and unmistakable in every bill insert and on every hold message: the company never asks for payment by prepaid card and never demands payment within the hour.
List every action a person can cause by phone alone, in operations and in the call centre. For each one, write down what the person on the phone would have to know that a well-prepared stranger could not. Where you cannot answer, you have found the gap, and it is usually shorter to fix than the list is to write.
Last note, from the inbound side of the phone. CallSphere builds AI voice and chat agents that answer business phone lines and web chat, capture caller details and book appointments around the clock. Used sensibly, an agent on the public-facing line handles the outage reports and appointment calls consistently and hands anything touching account changes or credit holds to a trained human with the verification step attached — which is exactly where that decision belongs.
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.
The monthly IEEE 1366 reliability close takes 64 hours across three people. What goal-driven agents change, the arithmetic, and what stays with the engineer.
Voice cloning is cheap in 2026 and dental refunds are authorised by phone. The callback rule that stops a $2,400 loss, plus where a human still has to decide.
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 controller can reroute a payroll run or a Bill.com vendor payment. The verification step, dual control and engagement-letter passphrase that stop it.
A cloned voice can get a prepaid training package refunded and a 24-hour fob turned back on. The callback rule that closes both, and what one incident costs.
© 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