Voice AI Meter Reading Capture: The Utilities Guide

A customer rings to give a meter reading. They read five or six digits off a display, sometimes a second set for a night rate, sometimes with the red decimal digits mixed in. Get the number right and the next bill is accurate. Get it wrong and you have either an angry customer disputing an inflated bill or a silent under-charge that surfaces months later as a large correction. For a utility or any metered-service provider, this small spoken number is one of the highest-volume, highest-consequence data-capture events in the whole contact centre.

At the end of March 2026, according to the Department for Energy Security and Net Zero, over 41 million smart and advanced meters were in homes and small businesses across Great Britain, and 72% of all meters are now smart or advanced. That is a large majority, but it also means more than one meter in four is still not smart, and even smart meters that have temporarily lost their connection revert to being read by hand. Millions of readings a year still arrive as a person reading a dial down a phone line. A voice AI agent that captures those readings has to solve a specific accuracy problem, and it is not the same problem as reading back a reference number.

Across the wider economy the pattern is familiar. McKinsey's State of AI research found that around 88% of organisations now use AI in at least one function, yet only about 6% capture material earnings impact from it. The gap is rarely the model. It is the unglamorous work of getting a captured value correct, confirmed and written back to the right system, which is exactly where a meter reading lives or dies.

This guide is shipped by the team behind Dilr Voice, enterprise voice AI built for regulated, high-volume deployments. Or see DATS, our five-stage AI consulting system.

What does a voice AI agent do when it captures a meter reading?

A voice AI agent capturing a meter reading does more than transcribe digits. Dilr Voice identifies which supply and meter the reading belongs to, asks which register the number is for, reads the value back digit by digit, then checks that the reading is possible against the last one held. A spoken meter reading is a claim about consumption, and the agent's job is to confirm that claim before it reaches billing.

That last check is what sets meter readings apart from other field capture. A reference number is validated with a check digit, a self-contained test that needs nothing else. A meter reading has no check digit. Instead it has an expected range, set by the previous reading and the customer's consumption pattern, so the agent can run a plausibility check the moment the number is spoken. We cover the checksum approach separately in our guide to reference number capture accuracy; here the validation is about what the meter physically can and cannot show.

Which meter and which rate is the reading for?

Before a voice agent captures a single digit, it must know which meter and which register the reading is for. A single-rate electricity meter shows one number. An Economy 7 or two-rate meter shows a day reading and a night reading. A gas meter can be metric or imperial. Capture the wrong register, or miss the second one, and the number is accurate but useless, because billing splits consumption on the register the agent forgot to ask for.

This is why the first job is disambiguation, not transcription. The table below maps the common meter and tariff types to what the agent must ask and what breaks if it does not. It is the same logic a good human agent applies without thinking, made explicit so a voice agent applies it every time.

The register question also decides how the agent labels the value it writes back. An accurate day reading filed against the night register is still a billing error, so the label travels with the number. This is close cousin to the record discipline we describe in address capture, where the confirmed value has to land in the right field of the right record, not just be heard correctly.

How does the agent check a spoken meter reading is plausible?

The plausibility check is the heart of meter reading capture. On a standard cumulative meter the number only goes up, so a reading below the last one held is a red flag: usually a misheard digit, sometimes a meter exchange. A reading far above the expected consumption band is the other. Dilr Voice compares the spoken value to the previous reading, and if it falls outside the plausible range it re-confirms with the caller before accepting it.

Note that this is a range check, not an identity check. The agent is not asking whether the reading matches a value we already hold, the way a date of birth capture flow compares a spoken date against a record to verify who is calling. A meter reading is a new fact each time. The test is whether that new fact is physically possible given the meter's history, and the response to an implausible value is to re-read and re-confirm, then escalate to a human only if it still looks wrong.

The same diagnostic logic underpins our AI operating model consulting, where each automated decision has an explicit threshold and an explicit fallback rather than a hopeful guess.

The full capture path runs in a fixed order, so nothing is skipped under time pressure.

Reading the value back digit by digit is now a well-worn habit across field capture, and we treat it the same way here as in vehicle registration capture and name capture: confirm before you commit. What meter reading adds on top is the possibility test, which no checksum-based field can offer.

Why does an inaccurate meter reading cost the business money?

An inaccurate meter reading is expensive twice over. First, it produces a wrong bill, which generates a complaint, an adjustment and a hit to trust. Second, if the customer was not billed correctly for a long period, the supplier may not recover the money at all. Ofgem's back-billing rules cap how far a supplier can charge when it was at fault, so a reading that drifts wrong for a year becomes a write-off, not a correction.

Ofgem states the rule plainly in its guidance for households:

Our back billing rules mean you do not have to pay for energy you used more than 12 months ago

That duty binds the supplier, not the customer and not the voice vendor. It is the supplier who loses the revenue when a meter reading is wrong and goes unnoticed, which is why a plausibility check at the moment of capture is a commercial control, not just a nicety. Getting more customers to submit an accurate reading, and catching the implausible ones on the call, directly protects billing revenue. That is a clearer return than most voice AI business cases, and it is the kind of outcome a DATS engagement is built to surface before a deployment starts.

How much of the meter estate still needs a manual reading?

Most meters in Great Britain are now smart, but a large minority still depend on someone reading the dial by hand. The Department for Energy Security and Net Zero reported that 72% of all meters were smart or advanced by March 2026. The rest are traditional meters, and smart meters that lose their connection revert to manual reads, so the volume of spoken readings a voice agent handles for a large supplier runs into millions a year.

Of that smart estate, over 38 million meters were operating in smart mode by March 2026, and those working smart meters send their reads automatically through the Data Communications Company, so the voice channel is not competing with them. It is picking up the readings the automated network cannot get: traditional meters, meters in hard-to-reach properties, and smart meters that have dropped offline. That is a durable workload, not a shrinking one, and it is worth building voice AI agents to handle it properly rather than treating it as a rounding error.

Who is accountable for the reading a voice agent captures?

Accountability does not move to the vendor because a voice agent took the call. The energy supplier remains the data controller and the regulated party: it answers to Ofgem for billing accuracy and carries the back-billing cost when a reading is wrong. The voice agent is a tool the supplier operates, so the supplier owns the plausibility thresholds, the escalation rules and the write-back. Dilr Voice makes those rules explicit and auditable rather than buried in a prompt.

Writing the confirmed reading into the system of record cleanly is its own discipline, and we treat it with the same care described in our email capture work: the value that was confirmed on the call must be the exact value that lands in billing, with the register label attached and a timestamp. A supplier running this at scale usually wants it inside a defined operating model, with clear ownership of thresholds and a named team maintaining them, which is where an AI execution office earns its place. You can see the rest of our thinking on captured-value accuracy across the voice AI guides.

What is the best way to capture meter readings with voice AI in 2026?

The best way to capture meter readings with voice AI in 2026 depends on volume and consequence. Self-serve builders such as Vapi, Retell AI, Bland AI and Synthflow can stand up a reading-capture flow cheaply, and for a low-volume, single-rate line that never touches a regulated bill, that is a reasonable choice. Governed platforms such as PolyAI and Dilr Voice earn their place when volumes are high, tariffs are mixed and an inaccurate reading carries a real back-billing cost.

The deciding factors are the plausibility logic, the write-back integrity and the audit trail, not the demo. A self-serve agent that transcribes a number and files it is easy to build; an agent that knows a reading below the last one is impossible, asks for the second register on an Economy 7 meter, and files the confirmed value against the right billing field is harder, and that is the difference between a novelty and a control. Read about our approach to placing AI inside enterprise systems, and where a governed platform is worth the extra work.

Can a voice agent read a smart meter automatically?

No. A working smart meter sends its readings automatically over the Data Communications Company network, so there is no phone call to capture. A voice AI agent is for the manual side of the estate: traditional meters, meters in properties the network cannot reach, and smart meters that have temporarily lost their connection. Dilr Voice is built to pick up exactly those readings accurately, not to duplicate what a healthy smart meter already does on its own.

What happens if the caller reads the meter wrong?

If the caller misreads the meter, the plausibility check is the safety net. When the spoken value is below the last reading held, or spikes beyond the expected band, Dilr Voice reads the number back and asks the caller to check the display before accepting it. If the second attempt still looks impossible, the agent flags it and routes to a human rather than pushing a suspect reading into billing. Confirm and escalate protects every captured field.

Does a voice agent work for both gas and electricity meters?

Yes, with a different question set for each. A gas reading needs the unit confirmed, cubic metres or hundreds of cubic feet, because the conversion into kWh depends on it. An electricity reading needs the register confirmed, single-rate or a day-and-night split. Dilr Voice branches on the supply type so the right questions are asked, and applies the same plausibility check to both, because a gas reading below the last one held is as impossible as an electricity one.

Want to see this in production? Try Dilr Voice live, book an AI placement diagnostic, see our DATS methodology, or read about our approach to placing AI inside enterprise systems.

Written by the Dilr.ai engineering team, practitioners who ship enterprise AI in production. We also run an about page if you want the team behind the guides. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.