AI Agent Disclosure on Utility Calls: The 10 Line Audit
Article 50 of the EU AI Act has applied to contact centres since 2 August 2026, and the first ten seconds of your call flow is where it lands. Here is the 10 line AI agent disclosure audit, the opener lines and the consent checklist.

A caller in Porto rings about a 94 euro gas bill. The line is answered on the first ring, the balance gets explained, a payment link lands by SMS, and she hangs up satisfied. Two days later a neighbour tells her she was talking to a machine.
That second conversation is the one your call flow has to be designed for.
AI agent disclosure means telling callers in plain language that they are speaking with an AI system, and giving them a route to a human without a fight. Article 50 of the EU AI Act covers systems that interact directly with people. In a utility contact centre that obligation lands in the first ten seconds of the call.
The European Commission sets out specific disclosure obligations so that people know when they are dealing with a machine and can make an informed decision.
Trade coverage put the operational date on it. Contact centre disclosure duties applied from 2 August 2026.
None of this is bad news for your queue.
Gartner surveyed 3,566 B2B and B2C customers and found that half say their interactions are easier when companies use generative AI. Callers push back when the AI becomes a wall between them and a person. That is a design choice you control.
Disclosure is a call design problem
Most teams treat this as one line pasted at the top of a script. Then nobody listens back to the recording.
The disclosure is a set of moments across the call where the caller asks the same question in different words. Who am I actually talking to, and can I get a person.
Answer it before they have to ask.
A disclosure the caller has to ask for twice is not a disclosure.
Stop here for thirty seconds. Open your ACD and pull one number: the share of AI handled calls where the caller asked for a human. If you cannot pull it, that is the first thing to fix.
The six moments on a utility call where disclosure matters
Six moments. Each has a different job. Map your current flow against them.
- Inbound answer. First ten seconds. Company name, then the fact that this is an AI assistant, then the offer of a person.
- Outbound reminder or arrears call. Company first, reason second, AI third, opt out fourth. In that order.
- Identity and account verification. Say why you are asking and say that the call is recorded. Two separate sentences.
- Payment or promise to pay. Say that the commitment is being logged and confirmed in writing, and by when.
- Escalation. Say a colleague is joining and that they already have the account open. Then do it.
- Call close. Summarise, say what happens next, and repeat how to reach a human.
Four openers you can steal
Steal these, then have your compliance team mark them up before anyone says them on a live call.
The inbound opener. Hello, you have reached [supplier]. I am an AI assistant and I can help with billing, meter readings and payments. If you would rather speak to a colleague, say so at any point and I will bring one in.
The outbound reminder opener. Good afternoon, this is [supplier] calling about the account at [address]. I am an AI assistant. The reason for the call is an invoice of [amount] that is now [n] days past due.
The are you a robot answer. Yes, I am an AI assistant working for [supplier]. I can sort this now, or I can bring in a colleague. Which would you prefer? No deflection and no cute phrasing.
The handover line. I am bringing in a colleague now. They can see the account and everything we have covered, so you will not have to start again.
Say it early, say it plainly, and offer the human before the caller has to ask.
What this looks like on a real floor
An illustrative example, not a customer.
Picture a 40 seat contact centre at a regional water utility in northern Italy. Roughly 18,000 calls a month, four in five about billing, meter readings and change of tenancy.
They put an AI voice agent on the after hours queue and the Monday morning spike. Containment looks good. Complaints stay flat.
Then the quality team pulls fifty recordings. The disclosure is there in Italian. It is missing from the English and Romanian flows, because those openers were written three weeks later by someone else.
Nobody decided to skip it. The second language flow was built in a hurry and the opener got copied from the wrong branch.
Disclosure rarely breaks on the flow you tested. It breaks on the flow you added afterwards.
The 10 line AI agent disclosure audit
One hour, one spreadsheet, ten lines. Mark each one pass or fail against a real recording.
- Play your current opener and time it. If the AI disclosure arrives after ten seconds, move it.
- Check the disclosure exists in every language you answer in, not only your main one.
- Check the outbound opener names the company before it names the assistant.
- Confirm the route to a human is stated out loud and works on the first ask.
- Say are you a robot mid call and listen to what comes back.
- Check the recording notice and the AI disclosure are separate sentences.
- Confirm the disclosure survives a transfer, so the human does not have to repeat it.
- Check voicemail and answer machine messages carry the same disclosure.
- Check the SMS or email follow up says the call was handled by an AI assistant.
- Check the transcript stores the disclosure line so you can evidence it call by call.
Audit the recording. The script is only what you meant to say.
Where an AI voice worker fits
Split the work into fixed path and judgement.
Fixed path. After hours cover, meter reading submissions, balance and payment questions, Day 5 and Day 15 reminders, change of tenancy intake. High volume, strict wording, no discretion needed.
Judgement. Hardship, disputes, bereavement, anything carrying a vulnerability flag. Keep your agents there and give them back the time the fixed path was eating.
AutoNurture.AI answers inbound and runs outbound in 12 plus languages, and the assistant discloses that it is an AI and offers a human handover on the call. You can hear that on a recorded German collections call where the assistant discloses it is an AI and an unpaid 140 euro invoice becomes four instalments of 35 euro.
When the caller wants a person, the hybrid human and AI dialer moves the call to a human seat in under two seconds with the transcript and the account already on screen.
Disclosure wording deserves the same testing as any other opener. The workbench lets you A/B test openers across prompts and voices and promote the winner, so you can find phrasing that keeps callers on the line without burying the fact.
Dial windows per market, retry logic and do not call rules run from the same place, which matters when one outbound list crosses three regulators.
For the collections and service workflow end to end, the utilities and energy playbook walks it through.
Consent, recording and GDPR, the practical checklist
Practical, not alarming. Walk it once, then get it signed off.
- Write down your legal basis for the call and for the recording before you dial.
- Keep the recording notice and the AI disclosure as separate sentences. They answer different questions.
- Set retention on recordings and transcripts, and be able to evidence the deletion job.
- Confirm where audio and transcripts are processed, not only where they are stored. EU hosting and data residency is a fair question to put to every vendor, including us.
- Version the disclosure line so you can show which wording ran on which date.
- Read do not call and vulnerability flags from the source system before dial.
Your obligations depend on your market and your regulator. Check with your compliance team before you change anything customer facing.
How to tell whether disclosure is costing you anything
Five numbers. All five come out of your dialer export.
- Seconds to disclosure. From answer to the sentence. Target under ten.
- Human request rate. Share of AI handled calls where the caller asks for a person.
- Handover time. From the request to a human on the line.
- Containment after disclosure. Share of calls resolved by the AI once the caller knows.
- Complaints per thousand AI handled calls, tracked weekly.
If human request rate jumps after a wording change, look at the wording before you look at the disclosure itself.
If containment holds and complaints stay flat, you have your answer.
What to do next
Pick one queue. Run the ten line audit against five real recordings this week and mark the fails.
Then book a demo and bring your own opener. Twenty minutes, a live call in your language, and you can hear the disclosure and the handover in the same call.
Frequently asked questions
What is AI agent disclosure?
AI agent disclosure is telling a person that they are interacting with an AI system rather than a human, in plain language and early enough to matter. In a contact centre that means naming the company, stating that the assistant is AI, and offering a route to a human agent in the opening seconds.
Does the EU AI Act require telling callers they are talking to AI?
Article 50 of the EU AI Act sets transparency obligations for AI systems that interact directly with people, and the European Commission describes these as disclosure obligations so people know they are dealing with a machine. How that applies to your specific call flows is a question for your compliance team.
When should the disclosure happen on a utility call?
In the first ten seconds, before identity verification and before any account detail is discussed. On outbound calls, say the company name first, then the reason for the call, then that the assistant is AI, then the opt out.
Does disclosing AI hurt containment?
Not in a well built flow. What drives callers away is being unable to reach a person. Measure human request rate and containment after disclosure before and after any wording change, and let the numbers decide.
Does a recording notice cover the AI disclosure?
Treat them as two separate sentences. A recording notice tells the caller the call is being captured. An AI disclosure tells them who, or what, they are speaking to. Keep both and keep them short.
Should every AI call offer a human?
Design for it. A working route to a person protects the rest of the experience, and it should work on the first ask rather than after three attempts.
What should an AI voice agent hand to a human?
Hardship, disputes, bereavement, any vulnerability flag, and any case where the caller asks. Fixed path work such as balance questions, meter readings and scheduled reminders is what automation is for.


