The False Positive Tax: Why DISCOM Theft-Detection AI Must Be Budgeted in Inspection Man-Days

Executive Summary

India now has roughly 5.97 crore smart meters in the field, 4.55 crore of them under the RDSS, and national AT&C losses have fallen to 15.04% in FY25 from 21.91% in FY21. The ACS-ARR gap has narrowed to ₹0.06 per unit and the sector posted its first collective profit after tax. The meter data problem is largely solved. The decision problem is not.

Every DISCOM board I speak to now has an AI theft-detection initiative. Almost none of them have a line item for the thing that determines whether it works: inspection capacity. A model is not a revenue protection programme. A model plus a bounded number of field visits, scored and fed back, is a revenue protection programme. The difference between those two sentences is where most of the money leaks.

This post puts arithmetic behind that claim.

The Base Rate Nobody Puts in the Tender

A peer-reviewed study published this year on monthly billing data from an emerging energy market is the most useful public reference point I have seen. Its working population: 145,977 consumers, of whom 864 were inspection-confirmed theft cases. That is a base rate of 0.59%.

The best model in that study — a random forest — reached an F1 of 0.677 and an AUC of 0.748 on a balanced hold-out set, and missed about a third of theft cases. Reconstructed, that balanced performance is roughly 64% recall at 72% precision, which implies a false-positive rate near 25% on honest consumers.

Seventy-two percent precision reads well in a steering committee deck. Now move it to a real feeder.

Per 100,000 consumers at a 0.59% base rate, you have about 590 offenders and 99,410 honest consumers. At 64% recall the model correctly flags 378 of the offenders. At a 25% false-positive rate it also flags roughly 25,250 honest consumers. Precision in the field: 1.5%. Sixty-eight site visits per confirmed case.

That is the false positive tax. It is not a modelling failure. It is Bayes' theorem meeting a rare event, and it is entirely predictable before a single rupee of licence fee is paid.

The Inspection Capacity Equation

The fix is not a better model. It is choosing an operating point against a known field constraint and writing that choice into the contract.

Assume one inspection team completes 8 verification visits per day — generous for rural North Bihar terrain, thin for urban clusters. The table below is illustrative arithmetic from the published base rate and published scores, not measured field results from any utility.

Read the last column, not the first. The lab-tuned threshold needs 3,200 man-days per lakh consumers to realise 378 cases. A division with four inspection teams would need over two years to clear one lakh consumers — by which time the consumption patterns that generated the alerts are stale and the alert list has been quietly abandoned. I have watched exactly that happen, twice.

The precision-first threshold catches 45% fewer offenders and recovers more rupees, because it is executable inside a billing cycle. In revenue protection, a detected case that is never inspected has a realised value of zero.

Three Numbers That Belong in Every AMI Analytics Contract

AUC, accuracy and F1 are procurement theatre. They are measured on balanced sets that do not exist in your service area. Replace them with three operational metrics:

1. Precision at k — where k is your actual monthly inspection capacity, stated in visits. "Of the top 1,500 consumers you rank this month, how many did my JEs confirm?" This is the only accuracy number that maps to a man-day.
2. Confirmation-to-closure rate — how many confirmed cases converted into a provisional assessment, a reconnection on prepaid, or a recovered amount. Detection without assessment is a tidy report.
3. Realised recovery per inspection man-day — the denominator the CFO cares about. This single ratio will tell you more about vendor quality than any model card.

Pay against these. Hold a retention against the third.

What a North Bihar Geography Teaches You

The useful signal in a geography like ours is rarely the sophisticated one. SBPDCL's AI deployment with Bidgely, announced in February 2025, flagged over 136 potential theft cases in its first month of use. That is a sensible number — small enough to inspect, large enough to learn from. It is also a reminder that the first year of a theft analytics programme is a labelling exercise disguised as a detection exercise.

Three field realities that models trained elsewhere do not know:

• A zero-consumption series in a flood-affected block in August is migration, not tampering. Seasonal labour movement in North Bihar produces thousands of consumption collapses that are perfectly honest.
• A prepaid consumer who recharges in ₹50 increments and runs at the emergency credit floor is poor, not fraudulent. Flagging them burns field capacity and public trust at the same time.
• A tampered DT-level energy balance is worth more than a hundred consumer-level anomalies, because it localises loss to a transformer before anyone knocks on a door.

Which is why I sequence it that way: energy balance at the DT, then feeder-level outlier ranking, then consumer scoring. Narrow the search space physically before you narrow it statistically.

The Closed Loop Is the Actual Asset

Your model's value in year three comes from one thing: whether the inspection outcome got back into the training data with a structured verdict. Confirmed theft, meter defect, billing error, no fault found, access denied. Five codes, mandatory on the JE's mobile form, no free text.

Most DISCOMs lose this. The inspection happens, the assessment is raised in a separate module, and the analytics vendor never learns which of its 25,000 alerts were wrong. The model cannot improve, so next year the utility buys a different model and repeats the cycle. The defensible asset was never the algorithm. It was the labelled outcome ledger — and it only exists if someone insists on it in month one.

What This Means for Leaders

1. Size the programme in man-days before you evaluate models. Count your inspection teams, multiply by working days, and let that number set the alert volume. Capacity is the design input; the threshold is the output.
2. Rewrite the SLA around precision at k. Replace AUC and accuracy clauses with precision at your stated monthly inspection capacity, confirmed by your own JEs.
3. Mandate five-code outcome capture on day one. No analytics contract should go live without the field verdict flowing back into the training set.
4. Sequence DT energy balance ahead of consumer scoring. Localise loss to a transformer first; it cuts the consumer search space by an order of magnitude.
5. Tie 20–30% of vendor payment to realised recovery per inspection man-day. It aligns the vendor with your constraint rather than with their leaderboard.
6. Review the alert backlog monthly and kill stale alerts. An alert older than one billing cycle is a liability, not an asset.

Key Takeaways

• A model at 72% precision on a balanced test set can fall to roughly 1.5% precision at a 0.59% real-world theft base rate — about 68 visits per confirmed case.
• Inspection capacity, not model quality, is the binding constraint in DISCOM revenue protection.
• Precision at your actual inspection capacity is the only accuracy metric that converts into rupees.
• Structured inspection outcomes are the compounding asset; the model is a depreciating one.
• Physical narrowing (DT energy balance) before statistical narrowing (consumer scoring) is cheaper and faster.

Written from 14 years inside utility transformation programmes, where the alert list is always longer than the inspection roster.

Sources: India Hits 520 GW Capacity; RDSS and Smart Metering Cut Losses, Boost DISCOM Profits — SolarQuarter, March 2026 · India Installs 3.9 Crore Smart Meters under RDSS by Dec 2025; AT&C Losses Drop to 15% — ENERGDIVE · Comparing Machine and Deep Learning for Electricity Theft Detection from Monthly Billing Data in an Emerging Energy Market — Technologies, 2026 · Bihar Strengthens Digital Transformation with Bidgely's AI Driven Power Solutions — Business Wire, February 2025