Voice of customer analytics that reads every call, email, chat and vote, and names each recurring problem with an owner and a fix
Every voice read, and each recurring problem named with an owner and a fix.
See one case, screen by screen ↓A utility hears from its customers through several doors: inbound calls, outbound feedback calls, the customer care inbox, the virtual assistant and the happiness votes in the app and on the website. Each sits in its own system, from the telephony platform to the happiness-meter tool and the complaint system, so nobody sees one customer's voice across them. An insights analyst codes a sample by hand each month, and a manual QA review hears about 1–2 % of calls. A failure four customers describe in two weeks is noise in a sample and a pattern in the full set.
Sentiment does not name the cause either. “Unhappy, billing” does not tell Billing Ops which rule to change, and a good rating can hide a failure: a customer taps “Happy” for the app and writes that the bill still ignores her reading. A theme in a slide pack has no owner, no measurable outcome and no date, so it is in the next quarter's pack too, while the same customers call back about the same estimated bill.
Estimated split of an insights analyst's monthly pack, in working days, by hand and with the solution.
Seven agents read every voice and fold the failures into named challenges; the analyst confirms and the service owner accepts the fix.
Lumina Utilities serves electricity and water customers in districts from Saltmarsh to Old Quay. Between 1 and 12 Oct 2026, 50 customers spoke to it through five doors. Here is what the 06:00 run made of them, screen by screen, in the working solution.
The run of 06:00 read 50 voices from 1–12 Oct in 4 min 12 s: 12 inbound calls, 3 outbound feedback calls, 10 emails, 10 virtual-assistant chats and 15 happiness votes. The voice reader kept 43 and set 7 aside, among them a silent call, two auto-replies and votes with no remark. The problem finder found 41 problems in 36 voices, five of them with two each, and they fold into 13 named challenges. Eleven have a recommendation; two wait for a second voice.
“all sources · 1–12 Oct 2026 · 4 min 12 s · every voice read, none sampled”
The heatmap sets six services against seven service-quality lenses. The darkest cell is Electricity × Timeliness: 9 problems in 2 challenges, energisation not scheduled and reconnection waiting for the overnight batch. Sara clicks Billing & Payments × Information accuracy: 4 problems, 1 challenge, 4 voices. The panel shows the challenge, the sources its voices came through and the strongest quote.
CH-4F2A · Customer meter readings ignored, bills stay on estimates: “They show as received, but not used.”
CH-4F2A has a three-word key, Mobile App · meter reading · not used, and every problem with those three standard terms joins it. The problem statement: customers submit meter readings in the app or by email, are told the reading was received, and are still billed on an estimate. Behind it: readings that arrive after the reading window, which closes five days before the bill run, are kept but not used, and the app gives no warning. Electricity and water accounts are both affected. Sara presses play on the first voice.
00:44 · Mariam, agent: “I can see readings from you on 28 July, 28 August and 28 September. They show as received, but not used.”
Helen sends her electricity reading in the app on the 28th of every month. Her September bill is $412.80 on an estimate of 1,940 kWh; her own reading puts the period at about 1,415 kWh, 525 less. Mariam raises a re-bill and promises a billing callback by Thursday, 8 Oct. Seven minutes later Helen votes Unhappy in the app, and the customer linker joins the call and the vote into one voice. No callback is logged by 12 Oct, so the voice also feeds CH-7C19, promised callbacks never made.
“Contact 1 of 2 · Inbound call · 6 Oct 09:14 · 1:52 · recorded” · two contacts joined into one voice by the customer linker
An anonymous customer rated the app Happy and wrote one line. The classifier reads the words as Unhappy and gives its reason, and the customer's own rating stays as they gave it, side by side. It is one of three Happy ratings in the run that describe a failure: an estimated bill, an idle fee charged again and a night without power after paying.
“App is quick. Shame the bill still ignores the reading I typed in.” · Customer's rating: Happy · AI reading: Unhappy
The run log shows how the challenge grew. On 2 Oct an email from Old Quay: the third estimated water bill, with a photo of the meter at 3,418 m³. No open challenge matched, so the challenge builder named CH-4F2A. When Helen's call joined on 6 Oct, the challenge reached two voices, and at 09:24 the recommendation writer wrote REC-4F2A. A chat on 8 Oct and the Happy vote on 10 Oct made four.
“Waiting for Farah Siddiqui (Billing Ops) to accept · Recommendations are proposals; the owner accepts or declines”
REC-4F2A: accept app and portal readings up to 1 day before the bill run, not 5; re-bill on actual reads where a valid customer reading was set aside in the last 3 cycles; show in the app which bill a reading will be used on; replace the outright rejection with a ±40 % check against the same month last year. Owner Billing Ops with Metering Ops, risk High, short-term, within 3 months. Sara sends it with the 4 voices and the recording; Farah accepts it for Billing Ops, due 15 Jan 2027, and outcome tracking switches on.
Expected outcome: “95% of valid customer readings used on the next bill, measured monthly”
Recommendations line up by owner: Billing Ops with Farah Siddiqui, Power Ops with Victor Lang, Digital Services with Tom Ashby, Customer Operations with Ines Varga and Customer Support with Ruth Okafor. Each carries its risk, its timeline and the outcome it will be measured on. Deposit refunds, already in progress, show the measure moving.
REC-3A61 · “78% refunds paid within 30 days · week of 5 Oct (was 52% in September)”
The customer linker builds 32 customers from account and premise numbers, phone and email, and never links an anonymous vote to a person. Fifteen need attention: supply affected, a repeat contact, or a promise not kept. Helen Marsh, Irene Castell and Diane Foster were each promised a callback, and none is in the contact log.
“Callback promised, not made” · CH-7C19, owner Customer Support
Happiness is happy voices over valid voices. By the AI's reading of all 43 valid voices the index for 1–12 Oct is 9 %: 4 happy, 8 neutral, 31 unhappy. On the 12 votes, customers' own ratings say 42 % happy and the AI's reading of the same remarks says 17 %. By channel, the Mobile App shows 57 % by rating and 14 % by reading, and each hidden failure links to its challenge.
“People contact a utility when something is wrong, so contacts score lower than surveys.”
The same estimated bill, seen by the four people around it: what their month looked like, and what it looks like now.
Seven agents read every voice and fold the failures into named challenges; the analyst confirms and the service owner accepts the fix.
Reads every voice, whether a transcript, an email, a chat or a vote remark, decides whether it is valid and useful, writes a two-line summary, and sets the customer category and language.
Sets the main and sub service, the service-quality lens, the journey stage, the voice type, the aspect and the district, and reads the sentiment from the words with its reason.
Writes one problem for each distinct failure a customer describes, with a one-line statement and a three-word key: channel · object · failure.
Folds spelling variants into the standard terms, such as “web portal” and “website” into Customer Portal, so counts compare across channels and months.
Attaches each ready problem to an open challenge with the same key, or creates a new challenge with a plain, specific name, so the list stays free of duplicates.
When a challenge reaches 2 voices, writes a recommendation: a title, a numbered action plan, the business owner, a risk, a timeline and an expected outcome that can be measured.
Builds the customer from account and premise numbers, phone and email, and joins contacts about the same incident, such as a call, its vote and the callback, into one journey.
The analyst confirms merges; the owner accepts or declines the fix. The agents propose; a named person decides.
Ask in plain words about a challenge, a voice or a score. Every answer cites the voices behind it.
Why are bills staying on estimates when customers send readings?
Because readings that arrive after the reading window are kept but not used. Helen Marsh sent hers on the 28th of July, August and September; the agent confirms they show as “received, but not used”. The window closes 5 days before the bill run. Four voices say the same, including a water customer in Old Quay. The recommendation is to accept readings up to 1 day before the run, with 95 % used as the measure.
Which happy ratings hide a real problem?
Three of the twelve votes. Each was rated Happy, and each remark describes a failure: a reading ignored on the bill, an idle fee charged again, and a night without power after paying. On the votes alone happiness is 42 %; read from the customers' words it is 17 %.
What should Power Ops fix first?
Energisation that is never booked: 5 voices, the most of any challenge, from two contractors, a business park and two households, all with the inspection passed and fees paid. Victor Lang accepted the fix on 9 Oct. Next is reconnection waiting for the overnight batch, 4 voices; its recommendation is waiting for Ines Varga.
Who was promised a callback that never came?
Three customers: Helen Marsh (by Thu 8 Oct, about her estimated bills), Irene Castell (2 Oct, a high bill) and Diane Foster (compensation after the Saltmarsh shutdown). None is in the contact log. They are on the Needs attention list.
13 screens from the working solution, on its sample data. Pick one to see it large.
Sources, the schedule, the milestone, the owners and the vocabulary are settings, chosen by Customer Experience. Changes are versioned and need a second person.
| Setting | Default | Choose from |
|---|---|---|
| Sources | Inbound calls, outbound feedback calls, contact-centre emails, virtual-assistant chats, happiness votes | Each source on or off · social posts on request, with approval from Communications |
| New voices read | Every 10 minutes | Every source is read on the same schedule |
| Full re-fold of challenges | Nightly at 06:00 | The run the analyst replays on Voice Insights |
| Voices before a recommendation is written | 2 | A number set by Customer Experience; challenges below it wait for another voice |
| Merge two challenges without the analyst | Never above 3 voices | Above it, the analyst confirms every merge or split |
| Term match the standardiser may apply alone | 0.90 or higher | Below it, the analyst decides |
| Time-sensitive services | Power and water supply · new connections | Flagged on the challenge list |
| Owners | Billing Ops, Power Ops, Customer Operations, Digital Services, Customer Support, Metering Ops, Water Ops | A named person for each, who accepts or declines its recommendations |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample data · “target” = the design goal, measured in a pilot (95 % = the expected outcome of the sample recommendation REC-4F2A, measured monthly) · “estimated” = our arithmetic: $191K = 20,000 repeat estimated-bill calls a year × $9.54 a call, and the customer's own call volume replaces ours; the split of an analyst's month is also our estimate · “typical” = published figures: Genesys and CallMiner (manual quality review hears about 1–2 % of calls) · E Source ($9.54 average cost per utility call; repeat contacts over $2.3M a year at large utilities) · J.D. Power 2025 (record-low 499/1,000 residential electric satisfaction). Lumina Utilities, its districts, customers and people named on this page are fictional: characters and sample data in the working solution.
Software that reads what customers say to a utility, in calls, emails, chats, surveys and votes, and turns it into named problems the business can fix. Voice of Customer Analytics reads every voice, folds the failures into de-duplicated challenges with the calls that prove them, and writes a recommendation with an owner and a measurable outcome.
Seven agents work in stages. The voice reader checks each voice is real, the classifier sets the service, service quality, journey stage and sentiment, the problem finder writes one problem per failure, the term standardiser and challenge builder fold them into named challenges, and the recommendation writer drafts the fix at the second voice. The customer linker joins a customer's contacts into one journey.
Every voice. New voices are read every 10 minutes and the challenges are re-folded nightly; in the sample run, 50 voices were read in 4 min 12 s, none sampled. A manual quality review typically hears about 1–2 % of calls.
The customer's rating is what they tapped; the AI's reading comes from their words, with its reason. Both are shown side by side and the rating is never overwritten. In the sample run, 12 votes were 42 % happy by rating and 17 % by reading, because three Happy ratings described a failure.
When a challenge reaches two voices, the recommendation writer drafts an action plan, the business owner, a risk, a timeline and an expected outcome. The insights analyst sends it with the voices and the recording, and the service owner accepts or declines it. Once accepted, its outcome is measured every month.
Yes. The analyst confirms merges of larger challenges, restores set-aside voices and corrects fields; the service owner accepts or declines every recommendation; the Head of Customer Experience signs off the monthly pack. Agents never assign work, contact a customer or promise compensation.
Card numbers and national IDs are masked in transcripts, anonymous votes are never linked to a customer, customers are never linked on a name alone, and recordings are kept for 24 months. These are settings, and changing them needs a second person from Customer Experience.
The Agentic Solution Engine builds and deploys it from your requirements and documents, such as your service list, owners, call and survey sources and a sample of past feedback, and it goes live once every quality gate has passed. We will walk you through it on your own customers' voices first.
No unhappy customer left without a call back, and every recovery measured.
1-min demoComplaint Root Cause AnalysisEach complaint trend proved with system data, owned, and fixed at the cause.
1-min demoCustomer Satisfaction Index ReportingEvery move in the score explained by name, with the customers’ voices behind each line.
1-min demoWe'll run Voice of Customer Analytics on a sample of your own calls, emails, chats and votes.