AIVS Platform: AI Invoice Verification System for Power Sector Billing and Dispute Intelligence

About the Project

The AIVS Platform is an AI Invoice Verification System built for organisations within India's power sector billing chain: state Discoms, Independent Power Producers, captive power producers, large open-access industrial consumers, and power trading licensees. Every month, these organisations exchange invoices worth hundreds of crores under two-part tariff Power Purchase Agreements governed by CERC and SERC regulations.

Kreate Technologies designed, built, and now maintains AIVS as an end-to-end AI product: a tariff -aware rule engine, machine learning anomaly detection, cross-period reconciliation, metering data integration, and automated dispute evidence generation, all connected directly into the organisation's ERP and payment approval workflow.

This is the story of what happens when an organisation stops assuming its power bills are correct, and starts verifying them.

The Challenge

A single monthly invoice from a thermal generator to a Discom can run to Rs. 47.3 crore. That figure is not one number. It is the sum of a fixed charge component tied to approved capacity norms and working capital costs, a variable charge component calculated against actual heat rate and coal cost, a fuel and power purchase adjustment surcharge that changes every month, and a set of transmission, SLDC, RLDC and open access charges, each governed by its own regulatory formula. Before AIVS, the finance team's honest position was straightforward. They checked the total. They confirmed the generation quantum against the SCADA report. They approved the bill for payment. What they did not do, consistently, was recompute each line item against the tariff order that governed it. Not because the team lacked the skill, but because doing that for hundreds of bills a month required more regulatory expertise and more hours than any accounts payable function could realistically deploy invoice by invoice. The working assumption was that the bills were probably right. What the organisation discovered, once it started checking properly, was that this assumption was quietly costing crores every year.

Problem Statement

Formula complexity outpaced manual capacity. The fuel and power purchase adjustment surcharge alone requires comparing actual and normative fuel costs, applying transmission loss factors, and adjusting for a lagged billing period. An error in any one input produces a result that still looks plausible on the page.
Tariff orders kept moving. CERC and SERC tariff orders are amended, revised through true-up petitions, and supplemented by orders that change individual parameters. Knowing which version applied to which billing period required a regulatory monitoring function that most accounts payable teams do not maintain.
Billing arrived fragmented across many counterparties. A Discom's monthly liability comes from long-term PPA generators, bilateral suppliers, exchange settlements through IEX and PXIL, the state transmission licensee, SLDC scheduling charges, RLDC fees, and open access claims, each on a different format and timeline. Reconciling all of this in spreadsheets could not scale.
Dispute evidence took too long to assemble. Supporting a billing challenge meant manually pulling together metering records, tariff order excerpts, SCADA data, and prior correspondence. That process took days, while the payment clock kept running under the Late Payment Surcharge Rules 2021, weakening the organisation's negotiating position with every day that passed.
Duplicate and retrospective billing went undetected. Supplementary bills for prior period adjustments could duplicate a charge already recovered in an earlier invoice. Spotting this required comparing a claim against the full billing history for that counterparty, something single-invoice review was never built to do.
The organisations most exposed to this gap were not the ones with weak finance teams. They were the ones whose finance teams were generalists by design, facing counterparties (generators and transmission licensees) that maintain dedicated, specialist billing departments. That expertise asymmetry, more than any individual error, was the real structural problem AIVS was built to close.

Our Solution

Kreate's approach was to embed regulatory economist-level verification into the invoice workflow itself, so that by the time a bill reached the finance team for approval, the arithmetic, the tariff compliance, and the metering consistency had already been checked.

The guiding design question for AIVS was never "does this number look reasonable?" It was "is this number correct, given the applicable tariff order, the metering data for this period, and the regulatory formula that governs this charge category?" Answering that question for one complex invoice might take a specialist three to four hours. AIVS answers it for every line item, across hundreds of invoices a month, without a human being the primary verification step. The system does not replace the finance team's judgement on whether to pursue or settle a dispute. It ensures that judgement is exercised with complete, verified information already in hand.

AIVS is built around five integrated verification engines.

Implementation Phases

Phase 1

AIVS was rolled out through a structured, twelve-week Billing Architecture Discovery engagement before a single invoice was verified in production.
During this phase, Kreate mapped every billing counterparty the organisation dealt with, catalogued invoice formats across PDF, Excel, and portal exports, and built the version-controlled tariff database encoding CERC and SERC tariff orders, PPA schedules, and amendment history specific to each counterparty. In parallel, the team mapped the organisation's metering data sources (SCADA, AMI, POSOCO, and SLDC) and configured the rule engine and anomaly detection models around the client's specific counterparty portfolio.
From that foundation, the platform's engines went live together: the tariff compliance checker and metering reconciliation running as the first layer of every invoice, cross-period reconciliation building a longitudinal record as billing history accumulated, and the dispute intelligence engine ready to generate evidence packages from day one of flagged exceptions.
The engagement continues through an AMC arrangement: the tariff database is updated as new SERC orders, amendment orders, and true-up determinations are issued, ML model performance is monitored and recalibrated, and dispute outcomes feed back into the system, all under SLA-backed operational support.

Results

Business Outcomes

2 to 4% average overbilling detected, as a share of total verified invoice value across deployments
90% invoice line-item coverage, compared with under 10% under manual spot-checking
70% reduction in the time to identify and initiate a billing dispute, from weeks down to hours
60% reduction in AP team time spent on routine bill processing, redirected to exception resolution
100% audit-ready verification trail across every processed invoice and billing period
So what did that actually change? — A 2 to 4% detection rate sounds modest until it is applied at scale. For a Discom processing Rs. 500 crore in monthly billing, a two percent detection rate represents roughly Rs. 10 crore of billing discrepancies identified for recovery every month, discrepancies that, under the previous manual verification regime, were simply being paid without challenge.
The shift from under 10% manual line-item coverage to 90% automated coverage is not just a productivity gain. It means the organisation is now checking charge categories it previously could not check at all, not because the team was unwilling, but because the volume made full coverage impractical by hand.
The move from weeks to hours in dispute initiation matters specifically because the Late Payment Surcharge Rules 2021 keep the payment clock running regardless of whether a bill is under review. Every week a dispute takes to assemble is a week the organisation's negotiating position weakens. Compressing that to hours changes the commercial leverage available to the finance team, not just the paperwork. As one Finance Controller at a state power distribution company put it, the platform functions as a tariff economist embedded in the invoice processing workflow, checking every line of every bill, every month, automatically. In the first year of deployment, the organisation identified and recovered billing overcharges that substantially exceeded the cost of the platform.

Operational Impact

For the accounts payable team, bill processing shifted from a full-invoice manual review to an exception-handling role: verified invoices move straight to payment, and attention concentrates on the invoices AIVS has actually flagged, each one already accompanied by a documented reason.
For finance leadership, the Financial Recovery Dashboard turns the dispute pipeline into a controlled, visible financial process, tracking identified discrepancies, submitted disputes, counterparty responses, and recovered amounts as a single measurable line of return on investment.
For regulatory and legal teams, decisions about whether to pursue, negotiate, or settle a dispute are now made against complete, AI-verified information rather than a partial manual check, while the judgement itself, which involves commercial relationship management and regulatory strategy, remains entirely theirs.
For the organisation's audit function, every processed invoice carries an immutable, timestamped verification record, making statutory and regulatory audit a matter of retrieval rather than reconstruction.

Key Takeaways

Verify quantity before verifying price. AIVS checks billed quantities against SCADA, AMI, POSOCO, and SLDC metering data as the first stage of every invoice review, before any tariff formula is applied. A correctly calculated charge built on an incorrect quantity is still wrong, so the platform was built to catch the foundational error first.

The costliest errors are the ones that look correct. Manual review was already catching gross, obviously anomalous mistakes. What it consistently missed were formula-correct-looking calculations built on an outdated parameter or a duplicated adjustment, the errors that pass a cursory glance but fail a systematic recomputation. Designing specifically for that category, rather than for obvious errors, is where most of the recovered value came from.

Automation and human judgement need a clear boundary. AIVS identifies discrepancies and assembles evidence. It does not decide whether to dispute a bill, negotiate a settlement, or accept a counterparty's revised calculation. Keeping that boundary explicit, rather than blurring it, is what let the organisation trust the system's output without displacing the people accountable for the commercial decision.

Cross-period visibility catches what single-invoice review cannot. Duplicate and retrospective billing errors are only detectable by comparing a new claim against the organisation's full billing history for that counterparty. A platform that treats every invoice in isolation, however thorough, will still miss this category of leakage.

Regulatory currency has to be maintained, not just built. Tariff orders change through amendments and true-up petitions on an ongoing basis. The AMC structure, with continuous tariff database updates and model recalibration, is what keeps the verification accurate over time rather than accurate only at launch.