Methodology

These are the principles common to all AI Analytics reports. They explain why the figures can be trusted and where the boundaries of trust lie.

Methodological principles

1. Mathematical reproducibility. Every numeric metric has a formula: archive coverage, data freshness, trailing gap, uncovered volume, Q/V delta, sensor state, passport completeness, incident priority.
2. Separation of facts, computations and assumptions. Measured data, computed estimates, heuristic indicators and hypotheses are shown separately. An uncovered-volume estimate is not a confirmed loss; a suspicious pattern is not proof of tampering.
3. Historical period separated from current state. Data of the selected period is assessed separately from current telemetry. A node can have a usable historical archive but degraded current monitoring — or the other way around.
4. Formal thresholds instead of arbitrary interpretation. Statuses (OK / Warning / Major / Critical / Ready / Not Ready / Aligned / Mismatch / Incomplete) are defined by numeric conditions; you see exactly which threshold fired.
5. Hard rules for critical conclusions. Some conclusions are impossible without mandatory data: no hourly archive → no under-metering estimate; no event log → no confirmed tampering; no passport → no full commercial verdict.
6. Metrological caution. A stuck sensor, zero flow, flat pressure or an archive gap are not automatically a violation — they are incidents that require verification.
7. Audit-readiness. Every status and recommendation carries its explanation: which data were used, which formula was applied, which threshold fired and which limits apply.

Data scope

Before running a report, you set the scope — which data enters the calculation.

A node-level report answers about one device — for example, how long its battery will last. A fleet-level report answers about many devices at once — for example, which hundred nodes lose the most gas.

If there is no data. When the selected company has no data for the period, the report honestly states the absence of data and does not substitute general “all companies” figures. An empty result is a signal to widen the scope or period, not a reason to trust a random number.

How to read a report

Start with the summary at the top — it gives the key figures in a few seconds. Further sections go in descending order of importance: what is critical, what is under observation, what is for reference. In economic reports, first look at what portion of the data the conclusion rests on (see confidence levels below).

A finished report is reproducible: with the same scope and period it yields the same result.

Three confidence levels for figures

Every figure is tagged with one of three levels — this guards against overstated promises:

• Confirmed — a measurable effect on real fleet data. Payback and ROI are built only on this.
• Scenario — a model estimate based on industry standards; requires verification against the company’s accounting policy. Shown separately, not included in payback.
• Possible — potential income upon meeting a condition (for example, selling CO₂ credits when a contract exists). Disabled by default.

How the totals stay honest

Filtering out “empty” nodes. Before calculating, nodes that would distort the total are removed: those long out of contact, with near-zero annual consumption (a dummy device), and without a monthly consumption archive. Only the “live” fleet is counted; how many nodes were filtered out and why is shown in the report.

Calculation from actual consumption, not “average × number of nodes”. Averages are deceptive: one factory equals a hundred apartments. Gas money is calculated from the actual sum of real annual consumption from each node’s monthly archive. If there is no data, the item is tagged “no data” rather than filled with an estimate.

Protection against double counting. Some benefit streams rely on the same physical gas (gas losses, billing accuracy, CO₂ credits). To avoid counting the same money twice, such streams are tagged as alternatives and enter the total only once.

Extrapolation “sample → fleet”. Link quality can be assessed on a sample of nodes (for example, 200) — enough for a large fleet. When a sample result is extended to the whole fleet, an explicit scaling factor and an “extrapolation” tag are shown, so sample figures are not passed off as fleet-wide ones.

Link availability. Availability is measured against a realistic expectation: the benchmark is one communication session per day, not per hour. This gives an honest “how many sessions arrived out of expected” metric.

A node’s personal norm. In control reports, the norm is determined by the node itself and its neighbors (similar nodes), rather than by a single threshold for all. Nodes that fall outside their own norm are highlighted — this is how metering bypass and anomalies are found.

Carbon methodology (CO₂e)

Emission reductions are calculated using the international methodology: the global warming potential of methane per IPCC AR6 — 28 over 100 years and 84 over 20 years. The result is presented as a business metric — tonnes of CO₂e per year; the monetary equivalent only with a confirmed contract on the carbon market.

Currencies

Money is shown in a single display currency; rates can be overridden manually. The conversion is given in one line, without burdening every cell with dual currency.

The role of AI

Optionally, a report can include a short management summary written by AI. It helps to quickly grasp the essence, but does not affect the calculations and can be turned off with a single parameter. All figures are calculated by deterministic formulas, not by AI.

The AI commentary can answer: what is the main problem of the node; which risks are visible from the computed metrics; which action to take first; why current monitoring status differs from historical archive status; which hypotheses require verification. It cannot: change a commercial metering verdict, raise or lower its legal weight, confirm tampering, or create evidence where there is none in the data.

Limitations

Reports are an assessment for prioritizing management and operational decisions, not a financial audit and not a metrological act. Confirmed figures rely on measured fleet data; scenario and possible figures rely on models and conditions the company must verify for itself. When data covers less than a third of the fleet, the conclusion is tagged premature: first build up coverage, then count money.

Glossary