PLATFORM
From treatment to customer care, see how utility data turns into decisions — with analytics first, and models only where useful.
Not every question needs AI. Each example shows the method it relies on.
150+ digests every month. Two real examples below.
Two digests a water utility's teams received from the platform, redrawn with their original figures and translated. Names, logos and links removed.
Calculated from the data of two operators — a water utility and an urban services operator — aggregated by month and by category. No names, places or accounts.
Invented figures, unless marked Real data.
Most use cases rely on rules, KPIs and statistical analysis. When a model adds value, it is called for a defined task, under the same Sparkoper governance.
This digest covers the requests created between 31 August 2026 and 6 September 2026.
The 10 main entries are shown.
The 10 main entries are shown. All actions on requests are counted.
Number of requests created for the assigned teams.
Period from 1 July 2026 to 31 July 2026.
Out of a total of 233 pieces of equipment.
Preventive and corrective interventions carried out during the month in question.
Preventive and corrective interventions carried out during the period.
Corrective interventions created during the period.
Preventive and corrective interventions carried out during the period.
Share of customer requests by nature. Free-text entries that fit no category are grouped under Other.
The 406 billing requests by nature, from April 2025 to September 2026. Billing without consumption alone accounts for more than half.
Share of requests whose customer account raised the same issue category more than once over the period, by category. Only requests with a usable account number are counted. A repeat contact does not prove an issue was left unresolved: it shows where first-contact resolution and proactive follow-up deserve a closer look.
Interventions created each month: planned work orders and corrective requests. September 2026, a partial month, is not shown.
Planned and corrective interventions per system, December 2025 to September 2026. The four clarifiers and the sand filters are each grouped.
Corrective requests per asset, December 2025 to September 2026. 42 assets generated at least one corrective record; the 16 that came back more than once account for 51 of the 77 records. They are the first candidates for root-cause analysis.
Requests by category since January 2023. Smaller categories are grouped under Other.
Current status of requests by year of creation. 2023 and 2024, with only 15 requests, are grouped. 2026 runs to 12 September, so recent requests are still in progress.
Monthly requests since November 2025. Projection: the average of the last three complete months; the band covers 80% of past month-to-month deviations. A deliberately simple statistical method, not a model.
Real, anonymised data. Corrective requests per month, January to August 2026. Projection: a Poisson count model on the average daily rate from 1 January to 12 September 2026 (77 requests in 255 days), with an 80% predictive range. Rather than predicting individual failures from too few variables, it sizes the corrective workload to plan for.
Invented figures. The score combines running hours, vibration and temperature trends, and failure history. Above 70, the pump is inspected before its next planned maintenance.
Invented figures. The probability combines pipe age and material, pressure variations, night flow and repair history. The top sectors are the first candidates for acoustic leak detection.
Invented figures. Average daily demand per month, in thousand m³, projected three months ahead with an 80% range. The top of the range crosses capacity in November: time to plan storage and pressure management.
Real, anonymised data. The structured category says Billing, but 85 of the 406 comments also describe a service interruption: billing during a period without supply, bills without consumption, a prolonged lack of water. Across the whole dataset, 132 requests reveal a water-supply issue that is not visible in their primary category, about 13% of requests outside Water shortage. Only counts are published; no comment is shown.
Real, anonymised data. A model rated the tone of 1,861 customer request descriptions, from April 2025 to September 2026: negative, neutral or positive. Only the shares are published. A complaint is negative by nature; this share is the baseline to track over time.
Recommendation for maintenance review — not an automated decision.
Real, anonymised data, December 2025 to September 2026. After removing three exact duplicates (same date, asset and type), the treated-water tank system has 12 distinct corrective events; 11 concern three pumps of the same family, all hydraulic-mechanical. Analytics compute these facts; a model can then draft a recommendation from them, such as the one above. It does not detect the figures and does not decide in the engineer's place.
Invented figures. A model proposes a category for each request written in free text. Below a confidence threshold, an agent checks the proposal before the request is routed.
Invented figures. For each request, a model recommends the next action. Nothing happens until an agent accepts, edits or rejects it, and every call to the model is logged.