What a utility AI agent actually does.
Most things sold as AI agents answer a question and hand off. The useful distinction is whether it can act in the systems the utility already runs. Here is one interaction, step by step.
01
Understand
The caller says why they rang, in their own words, without choosing from a menu. There is no phrase list to match and no wrong way to say it. A caller who opens with “my bill is double what it normally is” has stated a billing enquiry, a usage question and a complaint at once, and the agent has to hold all three.
02
Authenticate
Before anything account-specific is said aloud, the caller is identified. At Kingsport 67% were verified from the number they called on, with no account number keyed in. The rest are taken through a challenge. Nothing about an account is disclosed until this step has passed.1
03
Retrieve
The agent reads the real record: balance, due date, consumption history, open cases, service history. From the system of record itself, not a nightly copy, so what the caller is told is what a CSR looking at the same screen would tell them.
04
Decide
The step vendors skip. Given this caller, this account and this request, what is actually allowed? A payment arrangement has eligibility rules. A reconnection has a sequence. Deciding correctly requires the policy, not just the data.
05
Execute
The part that separates an agent from a chatbot. It takes the payment, books the appointment, starts or stops the service, opens the case. A system that can only describe what the customer should do next has moved the work, not done it.
06
Document
The interaction is written back: what was asked, what was checked, what was done. A call that resolves but leaves no record creates a second call later, from a customer who assumes nothing happened.
07
Escalate
When the request falls outside what it should handle, it transfers with the context attached: who the caller is, what they asked, what was already checked. The customer never repeats themselves. Restraint at this step is a feature, not a shortfall.
Steps one, two, three, six and seven are table stakes: most vendors in the category do them. Four and five are where the category divides. Deciding what is permitted requires the utility's policy, and executing requires write access to systems that bill real people. Ask any vendor to demonstrate both on your own data.
Common questions
What is the difference between a utility chatbot and a utility AI agent?
A chatbot answers questions. An agent completes the request. The practical test is step five: can it take the payment, book the appointment, start the service and write the result back to the system of record, or can it only explain how the customer could do that themselves? A system that can only describe the next step has moved the work to the customer rather than doing it.
What systems does a utility AI agent need to connect to?
At minimum the customer information system, since balance, due date and usage all live there. In practice also the payment provider, the CRM or case system so interactions are recorded, and the outage management system if outage calls are in scope. Grace integrates directly with existing telephony, CRM and customer systems, reading account state, balances, open cases and service history.
How does a utility AI agent authenticate a caller?
Usually in two layers. First a match on the number the caller is ringing from, which at the City of Kingsport verified 67% of callers with nothing keyed in. Callers who do not match are taken through a challenge on account details. No account-specific information is disclosed before authentication passes.1
How many calls can a utility AI agent handle at once?
Hundreds of concurrent conversations, scaling with demand. This matters more than total volume for a utility: a contact centre's worst day is defined by how many people call in the same ten minutes, which is what makes storm and billing-cycle peaks expensive to staff for.
What happens when the agent cannot help?
It transfers to staff with the full context attached: who the caller is, what they asked, and what was already checked. The customer does not repeat themselves. A request outside the defined scope should go to a person untouched rather than receive a partial answer.
Can a utility AI agent take a payment?
Yes. The card capture is handled by the telephony provider's PCI flow, so the card number never passes through the agent, and the result is written back to the account. Where a payment arrangement is in scope, the agent applies the utility's own eligibility rules rather than improvising.
Does the AI agent keep a copy of our customer data?
No. Grace reads from and writes to the customer information system directly and holds no shadow copy, so there is one source of truth and no second database to secure or explain to an auditor. Each organisation's data is isolated from every other's by the platform rather than by configuration.
How is a utility AI agent stopped from making things up?
By calling functions instead of generating answers. Grace runs on models fine-tuned for calling functions correctly and following instructions, so asked for a balance it reads the balance the system returns. Guardrails keep it inside its configured scope, and a review agent assesses every conversation afterwards and flags anything that looks wrong for a person to check.
What languages does a utility AI agent speak?
Grace speaks 45+ languages with no additional build. Instructions can be written in English while the conversation runs in the caller's language, and the City of Kingsport runs English and Spanish on the same line.
How do you measure whether a utility AI agent is working?
Start with the handle rate, then check it against the stricter measures. Handle rate is the share of inbound calls Grace completed itself, without transferring the caller to staff. Grace reached ~70% by the end of the first week in live production. Containment alone flatters, so also watch time on hold, the share resolved end to end, and how many callers came back within a week.2
Ready to deliver faster, more consistent service?
See how Grace handles a real call on your systems: voice, chat, email and beyond.