An AI voice agent for utilities customer service is not tested on a quiet Tuesday. It is tested at 9 p.m. on storm night, when thousands of customers call about the same outage in the same hour.
Most utility contact centres are sized for the average day and overwhelmed on the bad ones. This guide is for customer operations, IT and contact-centre leaders at electric, gas and water utilities. It covers which calls to give the AI, a reference architecture built around the outage system, the safety rules that belong in code, a storm-night capacity model and a checklist for any vendor.
Why the utility phone line is under pressure
Three recent data points explain why this matters now.
- Outages are getting longer. US electricity customers averaged 11 hours of interruptions in 2024, nearly twice the annual average of the previous decade, and Hurricanes Beryl, Helene and Milton accounted for 80% of those hours (U.S. Energy Information Administration, December 2025). In the first half of 2025, 45% of utility customers reported an outage (J.D. Power, via T&D World, October 2025).
- Satisfaction is at a low. J.D. Power's 2025 residential electric study scored overall satisfaction at 499 on a 1,000-point scale, the lowest across its residential utility studies, as average monthly electric costs rose 34% since 2020 to $189 (127,103 interviews, January–November 2025; J.D. Power, December 2025).
- Customers are not hearing from their utility. In J.D. Power's 2026 Utility Brand Appeal Index Study (142,626 customers, August 2025–June 2026), only 41% of electric customers recalled receiving any communication from their utility, and brand appeal was 57 points lower among customers who had an outage (T&D World, July 2026).
Put together: more outage hours, customers already unhappy about their bills, and a communication gap that the phone line ends up absorbing. J.D. Power also reports that satisfaction rises sharply when a utility contacts customers directly after power is restored. That is a design requirement, and we come back to it below.
You cannot hire for the four storm nights a year. You can build for them.
Which calls to resolve, assist or keep human
The same rule we use in every regulated sector applies here: automate the calls that need one lookup and a clear answer, and keep judgement, money disputes and anything safety-related with people.
| Call type | AI role | Why |
|---|---|---|
| Report an outage / “is my power out?” | Resolve | Identify the premise, check it against known outages, create an outage report if new. The biggest volume on storm night. |
| Restoration time | Resolve | Read the estimated restoration time from the outage system, never estimate it. |
| Bill amount, due date, payment confirmation | Resolve | Read-only billing lookup after identity verification. |
| Start, stop or move service | Assist | Capture the request and documents; completion runs through your existing back-office checks. |
| Payment arrangement request | Assist | Only offer plans your rules allow; anything outside them goes to a person. |
| High-bill complaint or dispute | Keep human | Needs judgement and often a meter or usage review. |
| Disconnection, hardship, medical-need accounts | Keep human | High stakes and rules that vary by state and regulator. |
| Gas smell, downed wire, sparking, fire, flooding near equipment | Never AI | Immediate transfer to your emergency line with standard safety instructions. |
A reference architecture for an AI voice agent for utilities customer service
The model is the smallest part of the design. What decides whether the agent works on storm night is the system around it. Five layers:
- Telephony edge. The agent sits in front of, or beside, your current platform, taking all calls or only overflow. It must scale concurrent calls elastically; a fixed number of “AI ports” recreates the queue you are trying to remove.
- Deterministic safety gate. Before any conversation logic, a rule set listens for emergency keywords and transfers. More on this below.
- Identity and premise resolution. Callers phone from a mobile, not the number on file, and often stand outside the affected address. Resolve the account by phone number, account number or service address, and confirm before reading any account data.
- Systems of record, read-first. The outage management system (OMS) for known outages and restoration estimates, the customer information system (CIS) for billing, and your CRM for contact history. Writes start with two actions only: create an outage report and create a ticket.
- Notification loop. Opt-in to text or callback updates, and a restoration callback when the OMS closes the outage for that premise.
The outage feed is the product
Most utility deployments fail on data, not on speech. The agent is only as good as the OMS record it reads. Three rules:
- Read the restoration time, quote its age. “The current estimate is 11:30 p.m., updated 20 minutes ago” is honest. An estimate the model invented is how one storm call becomes three.
- Know when the address is outside the known outage. A customer with no power at an address the OMS shows as energised has a different problem, possibly a service-line fault or a new outage. That call should create an outage report and, where your process requires it, go to a person, not get reassurance.
- Degrade loudly. If the OMS stops responding under load, the agent should say it cannot see outage details right now, take the report, and offer a callback. Silent fallback to generic answers is the worst failure mode.
This “know when you don't know” behaviour cannot be left to a prompt. We explain why in Confidently Wrong: if a behaviour matters, it lives in code.
Safety routing belongs in code
For a utility, the dangerous failure is not a wrong bill amount. It is an AI agent politely collecting an account number from someone who smells gas.
Test the safety gate with real recordings, accents, background noise and indirect phrasing (“there's a funny smell by the meter”) before go-live, and keep testing after every change.
Storm night: an illustrative capacity model
The numbers below are assumptions to show the method, not a benchmark. Replace them with your own storm history.
| Assumption | Value |
|---|---|
| Customers who lose power | 60,000 |
| Share who call in the first 3 hours | 15% → 9,000 calls |
| Agents on shift | 40 |
| Average handle time for an outage call | 5 minutes → 480 calls per hour |
| Human-only capacity in 3 hours | 1,440 calls; about 7,560 left waiting, abandoning or redialling |
| With an AI agent closing 70% of outage calls | Every call answered; about 2,700 reach people, with the account already identified |
Two things stand out. First, even at 70%, the human queue still exceeds capacity, so the agent has to offer callbacks rather than park callers on hold. Second, the 70% is only reachable if the OMS has good data on known outages. Measure the share of storm calls where the premise matched a known outage in your last three events before you believe any containment number, ours included. For the cost of the calls you do not answer, see our analysis of call center overflow cost; for the telecom version of this model, see the AI voice agent for telecom customer service playbook.
Restoration callbacks and outbound rules
The J.D. Power finding on post-restoration outreach is the cheapest satisfaction lever you have, and voice AI makes it practical at scale: when the OMS closes the outage for a premise, call or text the customers who reported it and confirm power is back. If it is not, that is a nested outage, and the callback just found it for you.
Outbound needs its own legal review. In August 2016 the FCC said utilities are deemed to have consent for automated calls on matters closely related to utility service, such as outage and restoration notices, to numbers customers provided, while opt-out and the other TCPA requirements still apply (summary by Keller and Heckman). In February 2024 the FCC ruled that AI-generated voices are “artificial” under the TCPA (FCC, February 2024). Keep collections and marketing calls out of the outage workflow, log consent and opt-outs, and have counsel sign off. This is not legal advice.
A rollout that survives the first storm
- Weeks 1–2: call-flow audit. Pull call reasons from your last three major events and a normal month. Size outage, billing and start/stop volume.
- Weeks 2–4: read-only integrations. OMS and CIS lookups, outage-report creation, ticket creation, transfer targets including the emergency line.
- Week 4: storm drill. Load-test at your worst historical hourly volume, with the OMS deliberately slowed or offline.
- Weeks 5–6: blue-sky go-live. Start on normal-day billing and outage calls, review transcripts daily, then switch on storm mode once the safety gate and OMS lookups have held for a few weeks.
Track five numbers: answer rate during events, share of outage calls matched to a known outage, safety-transfer accuracy (from transcript review, not self-reported), callback completion after restoration, and wrong-answer rate on restoration times.
Vendor checklist
- How many concurrent calls can you take, and what happens at the limit? Show a load test.
- Is the safety transfer a rule outside the model? Show us the rule and the test set.
- What does the agent say when the outage system is slow or down?
- Where does the restoration time come from, and does the agent ever generate one?
- Which systems do you write to on day one, and who approves each new write action?
- Can we review every transcript, and is our call data used to train models?
- Who operates and tunes the agent after go-live, and what metrics are you accountable for?
For reference, Unicall's managed service starts from $1,000 a month, billed on ticket volume rather than agent seats, with implementation quoted after a free call-flow audit (pricing). Averages across our live deployments, mostly in e-commerce customer care, are about 60% of calls resolved autonomously and cost per ticket down 52% within 90 days, with sample sizes and periods varying by client. They are not a utility benchmark: your result depends on your call mix and the quality of your outage data. Our deployments typically go live in 4–6 weeks without replacing the existing phone system, and call data is never used to train models. To compare voice platforms on public information, see our comparison hub.
FAQ
Which utility calls can an AI voice agent handle?
The high-volume calls that need one lookup and a clear answer: outage reporting and status for a known outage, estimated restoration times read from the outage management system, bill and due-date questions, payment confirmations, and start, stop or move requests captured for back-office completion. Safety calls, disputes, hardship and disconnection conversations stay with trained staff.
Should an AI voice agent handle gas leak or downed power line calls?
No. A safety keyword such as gas smell, downed wire, sparking or fire should end the AI conversation and send the caller straight to your emergency line or dispatch, with your standard safety instructions. That routing should be a deterministic rule in code that runs before any other logic, not an instruction in the prompt.
How does an AI voice agent give accurate restoration times?
It should only read the estimated restoration time held in your outage management system for that address or account, say when it was last updated, and never estimate one itself. If the record has no estimate, the agent says so and offers a text or callback when one is set.
Can a utility use AI voice for outbound outage and restoration calls?
Possibly, but check with counsel first. In August 2016 the FCC said utilities are deemed to have consent for automated calls closely related to utility service, such as outage and restoration notices, to numbers customers provided; opt-out and other TCPA rules still apply. In February 2024 the FCC ruled that AI-generated voices count as artificial voices under the TCPA.
How long does it take to deploy an AI voice agent for a utility?
A first inbound scope such as outage status and billing questions can typically go live in 4 to 6 weeks with a done-for-you provider, without replacing the phone system. Read access to the outage management and billing systems drives most of the timeline, and the agent should be load-tested at storm-level concurrency before it goes live.
Unicall.ai builds and operates done-for-you AI voice agents that answer every call 24/7, take real actions in your systems, and hand over to your people, with full context, the moment they should.