Version 4 is live with Human Dialer and AI Workbench
AutoNurtureAutoNurture
Back to blog

Voice AI for Utilities: The 27 Questions to Ask Before You Sign

87 percent of customers say reaching a human has to stay an option when a company uses AI for service. Here are the 27 questions to put in front of any voice AI vendor before a utility contact centre signs anything.

9 minChecklist
Two contact centre agents at their monitors under a large city map on the wall, one in the foreground wearing a headset and out of focus.

87 percent of customers say that when a company uses generative AI for service, the option to reach a human has to stay. In the same survey, 50 percent say the AI made the interaction easier.

Both numbers are true. Both belong in your vendor evaluation.

Voice AI for utilities works when your evaluation covers six things: escalation to a human, language quality, consent and disclosure, who owns the phone line, whether it reads your billing system, and proof on your own recordings. Below are 27 questions to put in front of any vendor, ours included, before you sign.

Most voice AI demos never get asked half of these. That is why they demo so well.

Why this checklist got longer

Two things changed at once.

First, disclosure. Article 50 of the EU AI Act, in force since 2 August 2026, requires that people interacting directly with an AI system are informed of it, unless that is obvious. The criteria are set out in the Commission's Article 50 transparency FAQ.

That obligation lands on the deployer as well as the vendor. The deployer is you.

Second, arrears. Ofgem's debt strategy update put domestic energy debt at 4.43 billion pounds by June 2025, up 20 percent in a year, with nearly three quarters of it held by customers who have no repayment plan in place.

No repayment plan means nobody has had the conversation. That is a contact problem before it is a credit problem.

Pull this from your dialer right now: how many outbound arrears attempts did your team make last month, against how many accounts were eligible? That gap is what you are buying.

Block 1. Escalation and handoff, questions 1 to 5

This block decides whether your customers end up hating the deployment.

  • 1. When a caller asks for a person, what exactly happens, and how many seconds does it take?
  • 2. Does the human get the transcript, the intent and the account on screen, or do they start again?
  • 3. Which trigger phrases force an escalation, and can my team edit that list without raising a ticket?
  • 4. Can a supervisor listen to a live AI call and take it over mid sentence?
  • 5. What happens when the handoff fails at 21:00 and nobody is on the floor?

If a vendor cannot demo a live takeover, you are looking at an IVR with better grammar.

Worth quoting in the room: that 87 percent comes from a Gartner survey of 3,566 customers.

Block 2. Language and voice quality, questions 6 to 9

A Lisbon caller and a Porto caller are two different accents on the same line.

  • 6. Which languages run on native voice models rather than translated scripts?
  • 7. What happens mid call when the caller switches language?
  • 8. How does it handle a meter serial read out digit by digit over a noisy line?
  • 9. What is the measured latency from caller silence to first word back?

Ask for a recording in your language, on your accents, with your tariff names and your meter types.

Voice AI that cannot pronounce your tariff name will not survive a Monday.

Block 3. Consent, disclosure and GDPR, questions 10 to 15

Practical, not scary. None of this is legal advice, so take it to your compliance team.

  • 10. Does the opener disclose that the caller is speaking with an AI system, in the caller's language?
  • 11. Where does the recording notice sit, and is it inside the first ten seconds?
  • 12. Which country hosts the data, and is that named in the DPA?
  • 13. Can retention be set per script rather than one blanket setting?
  • 14. Does an erasure request reach recordings and transcripts, not only CRM notes?
  • 15. Are quiet hours and do not call lists enforced by the dialer rather than by memory?

Ask for the disclosure line in writing, in every language you operate in. Vendors who have done the work hand it over in a minute.

Block 4. Telephony and the dialer, questions 16 to 19

Most evaluations skip this block. It is the one that shows up on the invoice.

  • 16. Who provisions the phone numbers, you or the vendor, and in which countries?
  • 17. Is there a separate telephony bill on top of the platform fee?
  • 18. Do human agents and AI calls share one number, one log and one dashboard?
  • 19. What is the failover if the voice provider has an outage during a storm spike?

Two stacks means two bills and a broken handoff at the worst possible moment.

Block 5. Billing systems and data, questions 20 to 23

A voice worker that cannot see the account can only take messages.

  • 20. Which utility billing platforms have a working connector, and can you name a live one?
  • 21. How does a payment plan agreed on the call get written back, and how fast?
  • 22. Are payment links tokenized so no card data touches the call?
  • 23. Can we export every transcript and recording as a flat file, on our own schedule?

Block 6. Proof, pricing and exit, questions 24 to 27

  • 24. Will you run a pilot on our own recordings and show us the transcripts, including the failures?
  • 25. What does pricing key off: seats, minutes, or resolved contacts?
  • 26. What does month one look like, and who does the script work?
  • 27. If we leave, what do we keep, and how long does the export take?

Ask question 24 first. A vendor who will not touch your recordings is selling you a demo.

How to score it in one table

BlockQuestionsDeal breaker if
Escalation and handoff1 to 5No live takeover in the demo
Language and voice6 to 9Translated scripts instead of native voice
Consent and GDPR10 to 15No written disclosure line per language
Telephony and dialer16 to 19Separate telephony bill and a second call log
Billing and data20 to 23No named connector for your billing platform
Proof and exit24 to 27Refuses a pilot on your own recordings

Score each block red, amber or green. Two reds and you stop.

A 40 seat example, start to finish

Picture a 40 seat water utility contact centre in northern Portugal in August. Abandonment at 22 percent by mid morning because a tariff letter landed on the 1st.

This is an illustrative example, not a customer of ours.

They ran this checklist across three vendors. Two failed question 4 on the spot, no live takeover. One failed question 17, telephony billed separately at a rate nobody could quote.

The survivor got a two week pilot on 200 real recordings: meter reads, outage updates and Day 5 reminders only. Nothing touching vulnerability.

Narrow scope, real recordings, two weeks. That is the shape of an evaluation that tells you something.

Where AutoNurture.AI sits against this checklist

Straight answer, and hold us to the same 27 questions.

AutoNurture.AI answers inbound calls without a phone tree, identifies the customer, resolves on the first call, and sends payment links, status updates or callbacks. When intent or risk spikes it pulls a human in mid call.

AI and human agents share one dialer, one number and one call log, with live handoff both ways plus supervisor whisper and barge, in 12 plus languages with native voice. Blocks 1, 2 and 4 live on one platform for AI and humans.

On the credit side it makes past due calls and SMS on cadence, negotiates payment plans inside your policy, sends tokenized payment links so no card data sits on the call, and writes back to your CRM. That work sits under utilities and B2C collections.

SAP IS-U, powercloud, Schleupen.CS and Salesforce Energy and Utilities Cloud are named on the integrations page, with a webhook and REST API for the rest. That covers block 5.

Data is EU hosted, GDPR ready, DPA on request, with configurable retention and right to be forgotten. The detail is in the Trust Center.

Automation takes the repeatable volume so your advisors and collectors get the calls that need judgement. Longer read: The Future of Voice AI in the Energy Sector.

Run the evaluation in 10 working days

  1. Day 1. Pull last month's call reasons and your Day 5 and Day 15 contact coverage. Write down the two numbers you want to move.
  2. Day 2. Send all 27 questions to every vendor in writing. Written answers only.
  3. Day 3. Drop anyone who will not answer questions 4, 10 or 17.
  4. Days 4 and 5. Live demo. You pick the call type and the language, and you force an escalation mid call.
  5. Day 6. Compliance pass on block 3 with your DPO or compliance lead.
  6. Days 7 to 9. Pilot on 200 of your own recordings, narrow scope, no vulnerability cases.
  7. Day 10. Score the six blocks and decide.

Ten days, no procurement theatre. The only real cost is pulling the recordings.

What to do next

Print the 27 questions. Send them to whoever is already in your inbox pitching voice AI.

Then, if you want to hear a voice worker run your own call types in Portuguese, German, Italian or English, book a demo. Twenty minutes, live, in your language, mapped to your queue.

Frequently asked questions

What is voice AI for utilities?

Voice AI for utilities is an AI voice agent that handles utility customer calls end to end: meter reading submissions, outage updates, billing queries, change of tenancy and payment reminders. It identifies the caller, resolves what it can, and hands over to a human when judgement is needed.

How do I evaluate a voice AI vendor for a contact centre?

Score six blocks: escalation to a human, language and voice quality, consent and disclosure, telephony ownership, billing integration, and proof. Insist on a pilot against your own recordings, and put the escalation path and the disclosure wording in the contract.

Does the EU AI Act require telling callers they are speaking to an AI?

Article 50 of the EU AI Act has applied since 2 August 2026. It requires that systems interacting directly with people are designed so those people are informed they are dealing with an AI system, unless that is obvious. Check how it applies to your setup with your compliance team.

Can AI voice agents handle arrears and payment reminder calls?

Yes for the repeatable part: Day 5 courtesy reminders, promise to pay follow ups, payment link delivery and plan confirmation inside your existing policy. Disputes, disclosed vulnerability, bereavement and anything outside policy route to a person immediately, with full context attached.

Will voice AI replace contact centre agents?

No. The volume it takes is fixed path and repetitive: meter reads, outage updates, first reminders, evenings and weekends. Human agents keep the negotiations, the complaints and the vulnerable customers.

Who owns the phone numbers when you deploy voice AI?

Ask it in writing. Some platforms provision local numbers per region themselves, others expect you to bring a telephony provider and bill it separately. One shared number and one call log across AI and human agents avoids a second dashboard.