Solutions · by industry · energy & utilities

Utilities: reliability, losses and collections, read from the warehouse every morning

CIO (the warehouse already holds OMS, AMI and billing extracts; the runtime stays inside the network) · distribution operations director (reliability, restoration, repeat outages) · losses-reduction and metering lead · CFO and collections director (billing, arrears, cash). The buyer is the CIO with the operations director; the regulator's reporting team reads the same records.

Questions

The three questions a utility board asks every month

Did reliability worsen because of more interruptions or longer ones, and where?
Are losses technical or commercial, and in which substations?
Is the cash coming in as fast as the bills go out, and which customer class is slipping?

In conversation

Sample questions people type into the conversation

deepen path — the finding's evidence is carried frozen; a new number needs a new query, and the agent says so

"Break the central region's outage minutes down by cause code."
"Which feeders had three or more unplanned outages in August?"
"Show the loss ratio by substation for the last three months."
"Estimated bills by meter type and district."
"How current is the outage data?"

Personas

Three example agents

The agents below are examples for this industry, not a fixed set: Distribution Operations Director Agent · Metering & Losses Agent · Billing & Collections Agent. Each one’s scheduled run is shown in the example sets that follow.

Set 1

Distribution Operations Director Agent

Audience: the operations director, regional operations managers, the regulatory reporting team.

Set 2

Metering & Losses Agent

Audience: losses-reduction unit, metering operations manager, regulatory reporting.

Set 3

Billing & Collections Agent

Audience: CFO, collections director, district billing offices.

An agent's name says which measures it reads and for whom. It holds none of the role's authority: it does not decide, approve or act.

Agents for the topics you choose

An agent is a job description, a bound semantic model and a schedule — not code — so the topics it watches are yours to set: a payer, a service line, a campaign, a supplier. We write the first ones with you.

Build your agents with us →

Example sets

Three example sets, one run each

Each set is one agent persona and one scheduled run’s output — the brief, the recommendations and the findings they rest on — in the product’s own record envelope. The first set is open; the other two are collapsed. The model and the measures the sets assume are listed at the foot of the page.

IllustrativeSet 1Distribution Operations Director Agent

Persona: Reads reliability for an electricity distribution company serving 1,412,000 customers across six regions; separates "more interruptions" from "longer interruptions", and a network story from a single-feeder story. · Audience: the operations director, regional operations managers, the regulatory reporting team. · Tone: number-led, plain; where a cause code is missing it says so rather than guessing. · Output language: en
Signals it watches: (1) SAIDI composition — customer-minutes split into interruption count (SAIFI) and duration per interruption (CAIDI). (2) Concentration — which regions and feeders carry the customer-minutes against their share of customers. (3) Repeat-outage feeders — feeders with three or more unplanned outages in the month, and their share of customer-minutes. (4) Restoration time drift — mean restoration minutes, with the dispatch-timestamp coverage stated. (5) Planned vs unplanned — the two lines never added into one sentence.
Thresholds it was given: finding 15 % · critical 40 % · ratio metrics (SAIDI, SAIFI, CAIDI, shares) 8 % · minimum share 3 % of network customer-minutes · severity rule: medium when a threshold is crossed, high when a ratio metric moves more than 1.25× its threshold or a count metric more than 1.5× the finding threshold, critical above the critical threshold; an ongoing topic keeps the severity it was given

Run header

Scheduled weekly, Monday 05:00, Asia/Baku
Period: August 2026 (31 days) vs August 2025 (31 days) — equal calendar days; major-event days excluded by the model's own flag in both years
Data complete through 2026-09-20, measured via max(OUTAGE_END_TS) (interruption end, not the daily SCADA landing time), 7 hours ago
Scope: the agent runs under the network operations scope (all regions, no billing amounts)
Model: the workspace's approved model, written onto the run
Cost: recorded on the run

Brief

August 2026: Interruptions Fewer but Longer — SAIDI +23.75 % on Duration Alone, Central Region 41.20 % of Customer-Minutes, Repeat-Outage Feeders 3.90 %

Assumptions strip

You did not state a period; last full month (August 2026) was assumed. Major-event days are excluded in both years by the model's own flag. Billing amounts are outside this agent's scope and do not appear.

August's reliability worsened on duration, not on frequency: customers were interrupted less often (SAIFI 0.182 → 0.171, −6.04 %) but for longer, so SAIDI rose from 11.80 to 14.60 minutes per customer (+23.75 %) and the average interruption lengthened from 64.83 to 85.38 minutes. The minutes are concentrated: the Central region carried 41.20 % of them on 27.50 % of the customers, and 46 feeders (3.90 % of the network, against 2.46 % a year earlier) had three or more unplanned outages and carried 22.70 % of all customer-minutes. Mean restoration time rose from 118 to 142 minutes (+20.34 %), a figure the agent publishes with a gap: 17.31 % of interruptions carry no dispatch timestamp, so the split between dispatch and on-site time is known only for the rest. Planned outage minutes fell 12.94 % and are reported separately.

covers: Finding 1, Finding 2, Finding 3

— Run: weekly, the workspace's approved model, data complete through 2026-09-20. Guard Ledger: held 1 — candidate f-04, reason unevidenced_causality: "storm activity" names a cause with no weather source in the model; one coached rewrite, then held; claim-check rewrite 1 — candidate f-01's title carried the superlative "worst month on record" with no supporting series, rewritten once and published as Finding 1 without it; duplicate 0; not_selected 0.

Recommendation 1

Break the Central region's August customer-minutes down by feeder and by cause code, and put the asset-age band beside each feeder

based_on: Finding 1, Finding 2 · owner: regional operations manager, Central, with the asset-management lead

The region carries 41.20 % of the network's customer-minutes on 27.50 % of its customers, and 31 of the 46 repeat-outage feeders. A feeder-by-cause cut will show whether the minutes sit on a few feeders with a named cause or are spread with unknown causes; the asset-age band is in this model and costs nothing extra. If the "other/unknown" cause share stays near 38.00 % after the cut, the next measurement is cause-code coverage, not an asset programme — the data cannot support the second without the first.

The findings this rests on

Finding 1

Reliability Worsened on Duration, Not on Frequency: SAIDI 11.80 → 14.60 Minutes per Customer (+23.75 %) While SAIFI 0.182 → 0.171 (−6.04 %) — CAIDI 64.83 → 85.38 Minutes; Central Region Carries 41.20 % of Customer-Minutes on 27.50 % of Customers

severity highdirection upnovelty newtrust Abreakdown REGION = Central

Unplanned customer-minutes interrupted, excluding major-event days, reached 20,616,000 in August 2026 against 16,660,000 in August 2025, so SAIDI moved from 11.80 to 14.60 minutes per customer (+23.75 %), 2.97× the 8 % ratio threshold — severity high. Interruptions per customer fell: SAIFI 0.182 → 0.171 (−6.04 %). The average interruption therefore lasted longer — CAIDI (SAIDI ÷ SAIFI) 64.83 → 85.38 minutes. The Central region carried 8,494,000 customer-minutes (41.20 % of the network) with 388,300 of the 1,412,000 customers (27.50 %): its own SAIDI moved from 16.22 to 21.87 minutes (+34.83 %), while the other five regions together moved from 10.12 to 11.84 (+17.01 %, on 1,023,700 customers) — both above threshold, the Central move twice the size. Planned outage minutes moved the other way, 9,120,000 → 7,940,000 (−12.94 %), and are not added into this figure. This data cannot separate a change in fault type from a change in crew response: the cause-code dimension carries "other/unknown" on 1,861 of the Central region's 4,897 interruptions (38.00 %), and the weather is not in this model.

Finding 2

Repeat-Outage Feeders: 46 of 1,180 Feeders (3.90 %) Had Three or More Unplanned Outages in August, Against 29 (2.46 %) a Year Earlier (+58.62 %) — They Carry 22.70 % of Customer-Minutes

severity criticaldirection upnovelty newtrust Abreakdown FEEDER_CLASS = Repeat (≥3 outages)

Of the 1,180 feeders in service, 46 recorded three or more unplanned interruptions in August 2026 (3.90 %), against 29 of 1,180 in August 2025 (2.46 %); the ratio rose +58.62 %, above the 40 % critical threshold — severity critical. The group carries 4,680,000 of the 20,616,000 customer-minutes (22.70 %). Thirty-one of the 46 feeders (67.39 %) sit in the Central region. The asset-age band is in this model: 29 of the 46 feeders (63.04 %) carry equipment in the oldest band, against 217 of 1,180 feeders network-wide (18.39 %). This data cannot say whether the repeat outages share a cause — the cause code is "other/unknown" on most of them, and the loading data that would show overload is not in this model.

Recommendation 2

Put the dispatch-timestamp coverage on a weekly watch beside mean restoration time

based_on: Finding 3 · owner: control-centre manager

A restoration-time trend read on 82.69 % of the records is a partial trend. The coverage ratio is the metric to watch first; when it returns above 95 %, the split into dispatch time and time on site becomes reliable and the agent will report it. Until then the agent stamps the restoration topic ongoing and does not alert again.

The findings this rests on

Finding 3

Restoration Time Drift With Incomplete Dispatch Records: Mean Restoration 118 → 142 Minutes (+20.34 %) — 17.31 % of August Interruptions Lack a Crew-Dispatch Timestamp

severity mediumdirection upnovelty newtrust Abreakdown — (network)

Across 12,790 unplanned interruptions in August 2026, mean restoration time was 142 minutes against 118 in August 2025 (+20.34 %), 1.36× the 15 % finding threshold — severity medium. The finding is published with a stated gap: 2,214 of the 12,790 records (17.31 %) carry no crew-dispatch timestamp, so the restoration time cannot be split into "time to dispatch" and "time on site" for that share. On the 82.69 % of records that carry the timestamp, time to dispatch rose from 31 to 44 minutes (+41.94 %) and time on site from 87 to 98 (+12.64 %) — both moved, the first more. Whether the missing timestamps are a recording fault or a process change cannot be read from this data.

What this run could not see.

Weather and network loading are not in this model; the cause code is unknown on 38.00 % of the Central region's interruptions; 17.31 % of records lack a dispatch timestamp. The agent reads the warehouse at its data edge, which lands the day's SCADA extract once a day — it is not a control room and does not see the network in real time.

IllustrativeSet 2Metering & Losses Agent

Persona: Reads the energy balance and the meter book: input against billed energy by substation, read success by meter type, estimated bills and zero-consumption meters; tells a technical loss from a commercial one only where the data can, and otherwise says "pattern requiring review". · Audience: losses-reduction unit, metering operations manager, regulatory reporting. · Tone: plain, careful with labels. · Output language: en
Signals it watches: (1) Loss ratio — (input − billed) ÷ input by substation and region, month against the same month last year. (2) Estimated-bill share by meter type and district. (3) Zero-consumption active meters for three consecutive months — a pattern requiring review, never a determination. (4) Read success by meter type. (5) Input growth against billed growth — the two lines side by side.
Thresholds it was given: finding 15 % · critical 40 % · ratio metrics 8 % · minimum share 2 % of energy input · consecutive-month rule 3 months · severity rule: medium when a threshold is crossed, high when a ratio metric moves more than 1.25× its threshold or a count metric more than 1.5× the finding threshold, critical above the critical threshold; an ongoing topic keeps the severity it was given

Run header

Scheduled monthly on the 6th, 06:00, Asia/Baku
Period: August 2026 billing month vs August 2025 billing month — equal calendar days; a billing month is closed on the 5th of the following month
Data complete through 2026-09-05, measured via max(READ_DATE) (read date, not posting date), 17 days ago at run time
Scope: no row-level rule is bound to this agent — the run was produced on every row
Model: the workspace's approved model
Cost: recorded on the run

Brief

August 2026: Losses 8.26 % → 9.88 % of Input While Billed Energy Held Flat — Estimated Bills 21.70 % of Accounts, 31,240 Zero-Consumption Active Accounts Flagged for Review

Assumptions strip

You did not state a period; the last closed billing month (August 2026) was assumed. The loss ratio is a residual of input and billed energy and is not split into technical and commercial components — this model cannot do that. No row-level rule is bound to this agent; the run was produced on every row.

August's energy balance opened a gap: input grew 2.39 % to 612.4 GWh while billed energy grew 0.58 % to 551.9 GWh, so losses rose from 49.4 to 60.5 GWh and the loss ratio from 8.26 % to 9.88 % — a +19.61 % change. The South-East Industrial substation carries 15.54 % of the losses on 8.10 % of the input. In the same month estimated bills rose from 14.20 % to 21.70 % of accounts, with manually read meters read at 71.40 %, and 31,240 active accounts (2.21 %) showed zero consumption for a third consecutive month, 58.29 % of them commercial — a pattern the agent flags for review and does not adjudicate. A sentence comparing the loss ratio with a national average was held: the average is not in this model.

covers: Finding 1, Finding 2, Finding 3

— Run: monthly, the workspace's approved model, data complete through 2026-09-05. Guard Ledger: held 1 — candidate f-05, reason numeric_provenance: token national average (loss ratio) has no address — no external series is in the model; duplicate 1 — candidate f-04 (estimated bills restated per district) merged into Finding 2 on a fingerprint match (metric + period + meter type); not_selected 0.

Recommendation 1

Break the South-East Industrial substation's losses down by feeder and by tariff class, and put the estimated-bill share of each feeder beside it

based_on: Finding 1, Finding 2 · owner: losses-reduction unit lead, with the metering operations manager

The substation carries losses out of proportion to its input, and the network's estimated-bill share rose in the same month; whether the two overlap on the same feeders is the first thing to know, and both cuts are in this model. If the feeders with the highest loss share are also the feeders with the highest estimated-bill share, the losses figure is partly a billing artefact and the metering team owns the next step; if they are different feeders, the losses team does. Transformer loading is not in this model; adding it is a model change.

The findings this rests on

Finding 1

Loss Ratio 8.26 % → 9.88 % of Energy Input (+1.62 Points, +19.61 %) While Input Grew Only +2.39 % — the South-East Industrial Substation Carries 15.54 % of Losses on 8.10 % of Input

severity highdirection upnovelty newtrust Abreakdown SUBSTATION = South-East Industrial

Energy input was 612.4 GWh in August 2026 against 598.1 GWh in August 2025 (+2.39 %); energy billed was 551.9 GWh against 548.7 GWh (+0.58 %). Losses therefore rose from 49.4 GWh (8.26 % of input) to 60.5 GWh (9.88 %), a +19.61 % change in the ratio, 2.45× the 8 % ratio threshold — severity high. Of the 41 substations, the South-East Industrial substation accounts for 9.4 GWh of the 60.5 GWh of losses (15.54 %) on 49.6 GWh of the 612.4 GWh input (8.10 %); the next three substations together hold 12.9 GWh (21.32 %). The loss ratio is a residual of two meters — input and billed — and this data cannot separate a technical loss on the line from unbilled consumption: the substation's transformer loading, which would bound the technical share, is not in this model.

Finding 2

Estimated Bills 14.20 % → 21.70 % of Accounts (+52.82 %) — Manual-Read Meters Read at 71.40 % Against 96.10 % for Smart Meters; the Rise Sits in Two Districts

severity criticaldirection upnovelty newtrust Abreakdown METER_TYPE = Manual

Of 1,412,000 accounts, 306,400 received an estimated bill in August 2026 (21.70 %), against 200,500 in August 2025 (14.20 %); the ratio rose +52.82 %, above the 40 % critical threshold — severity critical. Read success on smart meters stood at 96.10 % (812,050 of 845,000); on manually read meters at 71.40 % (404,840 of 567,000), against 84.70 % (525,140 of 620,000) a year earlier. Two of the six regions' districts carry 187,520 of the 306,400 estimated bills (61.20 %) while holding 166,700 of the 567,000 manually read meters (29.40 %). An estimated bill is a billing state, not a consumption fact: whether these accounts are over- or under-billed cannot be read until an actual read replaces the estimate, and that reconciliation is not a metric in this model.

Recommendation 2

Put the zero-consumption commercial accounts on a monthly watch, by district, and record the inspection outcome as a dimension so the pattern can be closed

based_on: Finding 3 · owner: metering operations manager, with the field-inspection lead

The pattern is worth a human's review, not a label. A monthly watch by district will show whether the count grows or clears; the inspection outcome — vacant, faulty meter, closed account, other — is the dimension that would let the agent report what the pattern turned out to be. Without it the agent will keep reporting the count and stamping the topic ongoing.

The findings this rests on

Finding 3

Pattern Requiring Review: 31,240 Active Accounts (2.21 %) Show Zero Consumption for Three Consecutive Months, Against 18,900 (1.34 %) a Year Earlier (+65.29 %) — 58.29 % of Them in the Commercial Tariff Class, Which Holds 9.80 % of Accounts

severity criticaldirection upnovelty newtrust Abreakdown TARIFF_CLASS = Commercial

Accounts marked active whose meter recorded zero consumption in June, July and August 2026 number 31,240 (2.21 % of 1,412,000), against 18,900 (1.34 %) for the same three months of 2025 — a +65.29 % change in the ratio, above the 40 % critical threshold — severity critical. Of these, 18,210 (58.29 %) are in the commercial tariff class, which holds 138,380 of the 1,412,000 accounts (9.80 %). The agent labels this a pattern requiring review: an active account with no consumption may be a vacant premises, a faulty meter, a disconnected customer not yet closed in the system, or unmetered consumption, and this data cannot tell these apart. The site-inspection outcome that would settle it is not in this model, and the agent makes no determination about any account.

What this run could not see.

Transformer loading, inspection outcomes and the reconciliation of estimated bills against later actual reads are not in this model. The agent does not see a meter in real time; it reads the closed billing month. No account is named; the finest cut in this model is district and tariff class.

IllustrativeSet 3Billing & Collections Agent

Persona: Reads the cash side of the utility: what was billed, what was collected, what is ageing, and through which channel the money came; tells a billing rise from a collection fall. · Audience: CFO, collections director, district billing offices. · Tone: formal, short. · Output language: en · Currency: Azerbaijani manat (₼)
Signals it watches: (1) Collection efficiency — collected ÷ billed by customer class and district, month against the same month last year. (2) Arrears ageing — the 90-plus-day share of receivables. (3) Billed growth against collected growth — the scissors. (4) Payment channel mix. (5) Disconnection and reconnection orders where the metric returns rows.
Thresholds it was given: finding 15 % · critical 40 % · ratio metrics 8 % · minimum share 2 % of billed amount · scissors 4 points · severity rule: medium when a threshold is crossed, high when a ratio metric moves more than 1.25× its threshold or a count metric more than 1.5× the finding threshold, critical above the critical threshold; an ongoing topic keeps the severity it was given

Run header

Scheduled monthly on the 8th, 06:30, Asia/Baku
Period: August 2026 vs August 2025 (31 days each; a billing month closes on the 5th of the following month)
Data complete through 2026-09-06, measured via max(PAYMENT_DATE) (payment value date, not the bank-file posting date), 3 days ago at run time
Scope: the agent runs under the finance scope (all districts; no meter-level rows)
Model: the workspace's approved model
Cost: recorded on the run

Brief

August 2026: Bills +8.46 % but Cash +3.55 % — Collection Efficiency Down to 89.28 %, Commercial Customers at 84.91 %, Arrears Over 90 Days Now 56.82 % of Receivables

Assumptions strip

You did not state a period; the last closed billing month (August 2026) was assumed. Amounts are in Azerbaijani manat. The disconnection and reconnection metrics returned no rows for July and August; this is recorded as a data state, not a finding.

August's bills grew faster than the cash: ₼ 71.8 M billed (+8.46 %) against ₼ 64.1 M collected (+3.55 %), taking collection efficiency from 93.50 % to 89.28 %. Commercial customers, 31.48 % of the billed amount, collected at 84.91 %; residential at 91.61 %; public bodies at 89.28 %. Receivables reached ₼ 84.3 M, and the part older than 90 days grew from ₼ 38.4 M to ₼ 47.9 M (+24.74 %), now 56.82 % of the total; the recovery-status dimension is empty this year, so the agent cannot say which of those balances are already in recovery. The disconnection-order line could not be reported.

covers: Finding 1, Finding 2

— Run: monthly, the workspace's approved model, data complete through 2026-09-06. Guard Ledger: held 0; data state 1 — disconnection_orders and reconnection_orders returned 0 rows for July–August 2026 (order-system extract stopped in June); duplicate 0; not_selected 0.

Recommendation 1

Break commercial collections down by district and by billed-amount band for August, and put the estimated-bill flag beside each row

based_on: Finding 1 · owner: collections director, with the district billing offices

Commercial customers carry the lowest collection ratio on a third of the billed amount. If the shortfall is spread across districts and sits in estimated bills, the reconciliation belongs to the billing office; if it sits in a few large actual bills, the next measurement belongs to the collections team. Both cuts are in this model. The dispute flag is not — until it is, disputed and unpaid cannot be told apart from this data.

The findings this rests on

Finding 1

Billing–Collection Scissors: Billed ₼ 71.8 M (+8.46 %) While Collected ₼ 64.1 M (+3.55 %) — Collection Efficiency 93.50 % → 89.28 % (−4.23 Points); Commercial Customers at 84.91 %

severity mediumdirection downnovelty newtrust Abreakdown CUSTOMER_CLASS = Commercial

August 2026 bills totalled ₼ 71.8 M against ₼ 66.2 M in August 2025 (+8.46 %); collections in the month totalled ₼ 64.1 M against ₼ 61.9 M (+3.55 %). The gap between the two growth rates is 4.91 points, 1.23× the 4-point scissors threshold — severity medium; collection efficiency fell from 93.50 % to 89.28 % (−4.51 % relative, below the ratio threshold). By class: residential ₼ 38.75 M collected of ₼ 42.3 M billed (91.61 %); commercial ₼ 19.19 M of ₼ 22.6 M (84.91 %); public bodies ₼ 6.16 M of ₼ 6.9 M (89.28 %). Commercial customers hold ₼ 22.6 M of the ₼ 71.8 M billed (31.48 %) and the lowest ratio. The estimated-bill flag is in this model but was not opened in this run, so this record does not say how much of the billed growth is estimates; and it cannot say how much of the uncollected amount is disputed rather than unpaid — the dispute flag is not in this model.

Recommendation 2

Put the 90-plus arrears share on a monthly watch by customer class, and add the recovery-status dimension for 2026 before the next run

based_on: Finding 2 · owner: CFO, with the legal-recovery unit and the data-model owner

The ageing shift is the record that matters for the year-end provision; the recovery status is what turns a balance into a decision. The dimension exists in the model and is empty for this year; filling it is a source-system task, not an agent change. The disconnection-order extract should resume for the same reason — the two metrics belong side by side.

The findings this rests on

Finding 2

Arrears Older Than 90 Days ₼ 38.4 M → ₼ 47.9 M (+24.74 %) — Now 56.82 % of Receivables, Against 51.20 % a Year Earlier, While Total Receivables Grew +12.40 %

severity highdirection upnovelty newtrust Abreakdown AGEING_BAND = 90+ days

Receivables stood at ₼ 84.3 M at the end of August 2026 against ₼ 75.0 M a year earlier (+12.40 %). The 90-plus-day band grew faster: ₼ 47.9 M against ₼ 38.4 M (+24.74 %), 1.65× the 15 % finding threshold — severity high, so its share of receivables rose from 51.20 % to 56.82 %. The younger bands together fell −0.55 % (₼ 36.6 M → ₼ 36.4 M). The 90-plus band is ₼ 29.75 M commercial and public-body debt (62.11 % of ₼ 47.9 M), against ₼ 29.5 M of ₼ 71.8 M billed (41.09 %). This data cannot say which of these balances are in a payment plan or in legal recovery — the recovery-status dimension is empty for 2026 in this model.

Data state

Data state (not a finding): the disconnection orders and reconnection orders metrics returned no rows for July and August 2026 (the order system moved to a new platform in June and its extract has not resumed). No severity; no recommendation rests on it; the model owner is notified.

What this run could not see.

The dispute flag and the 2026 recovery status are not populated; disconnection and reconnection orders returned no rows. The agent reads value dates from the closed billing month and does not see today's payments. No customer is named; the finest cut is district and customer class.

Every number carries an address

In the product each figure binds to an evidence address — envelope, row, cell, check digit — and a number without one is held, not published. The anatomy of a record →

Derived figures show their operation

A change, a share or a gap is computed by the engine and carries its operands; the model never divides. How every number is proved →

AI governance is the publication layer

Eight deterministic gates, the trust tier, badged interpretation, an assumptions strip that cannot be switched off, and the Guard Ledger. Where the model can and cannot reach →

FAQ

Four questions this page is usually asked

01Does the agent connect to SCADA or the outage system directly?

No. It reads the warehouse extracts your team already lands, through a read-only gateway, and states the data edge it found.

02Can it act on the network or send disconnection orders?

No. It reads data and publishes records; a human owns every action.

03Does it see individual customers or meters?

Customer name, meter serial and street address are marked personal in the model; the finest cut is district and tariff class.

04What happens when an extract stops, as the disconnection orders did in this example?

The metric returns no rows; the agent records a data state with no severity and builds nothing on it; the model owner is notified.

Free data discovery study

See which of these signals your own warehouse can carry.

Send us the name of your data platform and the domain you argue about most. We read your semantic model with you and answer in writing: which of the signals on this page your model carries today, which need a change in Axoria Data Studio, and what a four-week pilot would measure. No cost, no sequence — a person replies within two working days.

Request a free data discovery study → Start a 4-week pilot

What the examples assume

The data, the model and the measures behind the three sets

Nothing in a set rests on a source outside the model described here. Open what you want to reconcile.

01Assumed data sources — data a distribution company already has

Outage management system (interruption start and end, customers affected, cause code, feeder, crew dispatch), advanced-metering and meter-data management (interval reads, read success, estimated flags), customer information and billing (bills, payments, arrears ageing, tariff class, district), GIS asset registry (feeder, substation, asset age), enterprise asset management (work orders, backlog), SCADA extracts landed daily. Typical warehouse grain: one row per interruption, per meter-read, per bill line.

02Assumed semantic model — the minimum for the three personas

Measures: customers_served · customer_minutes_interrupted · interruption_count · customers_interrupted · saidi and saifi (engine-derived: customer-minutes ÷ customers; interruptions ÷ customers) · caidi (engine-derived) · planned_outage_minutes · unplanned_outage_minutes · mean_restoration_minutes · repeat_outage_feeder_count · energy_input_gwh · energy_billed_gwh · loss_gwh and loss_ratio (engine-derived) · meters_read · meters_estimated · zero_consumption_active_meters · amount_billed · amount_collected · collection_ratio (engine-derived) · receivables_total · receivables_90_plus · complaints_count · work_orders_open.
Dimensions: date (interruption start; read date; bill date) · region · district · substation · feeder · cause code · planned/unplanned · customer class (residential, commercial, public body, industrial) · tariff class · meter type (smart, manual) · payment channel · asset age band. Marked personal and never broken down by: customer name, meter serial, street address (the finest allowed cut is district).
Data edge: measured by the engine via max(outage_end_ts) for the reliability star, max(read_date) for metering, max(bill_date) for billing; each star declares its own.

03Assumed KPIs — what utility executives track

KPI	Definition	Unit	Typical band	Usual breakdown
SAIDI	customer-minutes interrupted ÷ customers served	min per customer per period	US ≈125.7 min/yr excl. major events [assoc: EIA via secondary]	feeder · region
SAIFI	interruptions ÷ customers served	count per customer per period	US ≈1.1/yr excl. major events [assoc: EIA]	feeder
CAIDI	SAIDI ÷ SAIFI	minutes	none universal	region
Mean restoration time	interruption end − start, mean or median	minutes	none universal	cause · region
Repeat-outage feeders	feeders with ≥3 unplanned outages in the period ÷ feeders	%	none	region · asset age
Transmission & distribution losses	(energy input − energy billed) ÷ input	%	US ≈5–6 %; world ≈8.3 % [assoc: World Bank / EIA]	substation · region
Collection efficiency	amount collected ÷ amount billed	%	none universal	customer class · district
AT&C loss	1 − (billed × collection ratio ÷ input)	%	none universal	region
Meter read rate / estimated-bill share	meters read ÷ meters due; estimated bills ÷ bills	%	none universal	meter type · district
Lost-and-unaccounted-for gas	receipts − deliveries ÷ receipts	%	2–5 % typical; best <1 % [assoc: NARUC/ICF]	district
Arrears 90+ days	receivables older than 90 days ÷ total receivables	%	none	customer class
Complaints per 1,000 customers	complaints ÷ customers × 1,000	count	none	region · type
Capex delivery vs plan	capex spent ÷ planned	%	none	programme
Capacity factor (generation assets)	actual ÷ maximum generation	%	US 2024: wind 34 %, solar 23 % [assoc: EIA]	plant · month

Utilities: reliability, losses and collections, read from the warehouse every morning

The three questions a utility board asks every month

Sample questions people type into the conversation

Three example agents

Distribution Operations Director Agent

Metering & Losses Agent

Billing & Collections Agent

Agents for the topics you choose

Three example sets, one run each

August 2026: Interruptions Fewer but Longer — SAIDI +23.75 % on Duration Alone, Central Region 41.20 % of Customer-Minutes, Repeat-Outage Feeders 3.90 %

Break the Central region's August customer-minutes down by feeder and by cause code, and put the asset-age band beside each feeder

Reliability Worsened on Duration, Not on Frequency: SAIDI 11.80 → 14.60 Minutes per Customer (+23.75 %) While SAIFI 0.182 → 0.171 (−6.04 %) — CAIDI 64.83 → 85.38 Minutes; Central Region Carries 41.20 % of Customer-Minutes on 27.50 % of Customers

Repeat-Outage Feeders: 46 of 1,180 Feeders (3.90 %) Had Three or More Unplanned Outages in August, Against 29 (2.46 %) a Year Earlier (+58.62 %) — They Carry 22.70 % of Customer-Minutes

Put the dispatch-timestamp coverage on a weekly watch beside mean restoration time

Restoration Time Drift With Incomplete Dispatch Records: Mean Restoration 118 → 142 Minutes (+20.34 %) — 17.31 % of August Interruptions Lack a Crew-Dispatch Timestamp

August 2026: Losses 8.26 % → 9.88 % of Input While Billed Energy Held Flat — Estimated Bills 21.70 % of Accounts, 31,240 Zero-Consumption Active Accounts Flagged for Review

Break the South-East Industrial substation's losses down by feeder and by tariff class, and put the estimated-bill share of each feeder beside it

Loss Ratio 8.26 % → 9.88 % of Energy Input (+1.62 Points, +19.61 %) While Input Grew Only +2.39 % — the South-East Industrial Substation Carries 15.54 % of Losses on 8.10 % of Input

Estimated Bills 14.20 % → 21.70 % of Accounts (+52.82 %) — Manual-Read Meters Read at 71.40 % Against 96.10 % for Smart Meters; the Rise Sits in Two Districts

Put the zero-consumption commercial accounts on a monthly watch, by district, and record the inspection outcome as a dimension so the pattern can be closed

Pattern Requiring Review: 31,240 Active Accounts (2.21 %) Show Zero Consumption for Three Consecutive Months, Against 18,900 (1.34 %) a Year Earlier (+65.29 %) — 58.29 % of Them in the Commercial Tariff Class, Which Holds 9.80 % of Accounts

August 2026: Bills +8.46 % but Cash +3.55 % — Collection Efficiency Down to 89.28 %, Commercial Customers at 84.91 %, Arrears Over 90 Days Now 56.82 % of Receivables

Break commercial collections down by district and by billed-amount band for August, and put the estimated-bill flag beside each row

Billing–Collection Scissors: Billed ₼ 71.8 M (+8.46 %) While Collected ₼ 64.1 M (+3.55 %) — Collection Efficiency 93.50 % → 89.28 % (−4.23 Points); Commercial Customers at 84.91 %

Put the 90-plus arrears share on a monthly watch by customer class, and add the recovery-status dimension for 2026 before the next run

Arrears Older Than 90 Days ₼ 38.4 M → ₼ 47.9 M (+24.74 %) — Now 56.82 % of Receivables, Against 51.20 % a Year Earlier, While Total Receivables Grew +12.40 %

Every number carries an address

Derived figures show their operation

AI governance is the publication layer

Four questions this page is usually asked

See which of these signals your own warehouse can carry.

The data, the model and the measures behind the three sets

Where these pages lead

On this site