Azhar Khan · Utilities, the data layer
UTILITY BILL PAY · AI AGENTIC SYSTEM← Back

Check the account before the exception.

One bill carries 8 services, 9 meters and 48 charge lines. The pipeline processes 700,000 bills a month. It checks the account record before any exception is raised.

invoices a month
vendors
layouts
capture accuracy
accounts
01 · INTAKE Bills arrive. PDF · EDI · API · email 02 · CLASSIFY 300+ layouts. model routes each bill to its template 03 · EXTRACT 190 fields. 48 charge lines · 8 services · 9 meters 04 · VALIDATE Account check first. reconciled before flagged 05 · RECONCILE One vendor. One site. One meter. collated state view 06 · APPROVE 1 in 100 reaches a person 07 · POST Payable created. payment scheduled · three-way match 08 · PAY $2.5B a year. cleared · posted · reconciled
  • 01 · INTAKEBills arrive.PDF · EDI · API · email
  • 02 · CLASSIFY300+ layouts.model routes each bill to its template
  • 03 · EXTRACT190 fields.48 charge lines · 8 services · 9 meters
  • 04 · VALIDATEAccount check first.reconciled before flagged
  • 05 · RECONCILEOne vendor. One site. One meter.collated state view
  • 06 · APPROVE1 in 100 reaches a person
  • 07 · POSTPayable created.payment scheduled · three-way match
  • 08 · PAY$2.5B a year.cleared · posted · reconciled

Check the account record before raising an exception. That one change took exceptions from 50% to 1%.

← UtilitiesHome →Microsoft case study →Resume →

The back office becomes a buying desk.

One intake spine on the left. The reconciled model in the center, where a record that will not reconcile is held and never published. The data products on the right are cut from that one reconciled source, and the payment run is one of them.

LAYERS
03 AGENTIC ORCHESTRATION · UTILITY BILL PAY: DATA LAYER A BUYING DESK DATAone source of truth IDENTITYevery action has a name OBSERVABILITYthe layer watches itself EXCEPTIONSfailures arrive with instructions GOVERNANCEnothing ships without review Ontology underneath, the reconciled model in between, traces on top. One linear spine of intake feeds the model; the model makes the products and stays live off their loop. EXECUTION TRACES · OBSERVABILITY Every decision logged: path, outcome, error. The log feeds the evals and the error catalog. TRACES ON TOP ONTOLOGY the ground the model is built on BUSINESS ONTOLOGY what a bill, a vendor, a site, a meter, and a decision mean, and how they connect TECHNICAL ONTOLOGY which system holds what · one vendor, site, and meter reconciled across the source systems, so the model spans one operation end to end ONTOLOGY UNDERNEATH the ground the model is built on every job & model → traces the model resolves meaning here · one vendor, site & meter across systems INTAKE · CAPTURE → EXTRACTION EDI Feeds Portal Harvest Mail / Scan Meter Data Layout Classify Field Extraction THE RECONCILED MODEL the modeled operation: every bill, meter & site reconciled Normalizeruns autonomously Reconcileruns autonomously ML Forecastapplied ML RECONCILED SOURCE OF TRUTHone vendor · one site · one meter THE DATA MOAT: analytics no competitor can buy BENCHMARKING · savings ranked Manager Acts FORECASTS · ACCRUALS Controller Posted BILL AUDIT → PAYMENT the byproduct: the payments run this path, but they are not the point Payment CARBON · SCOPE 1/2 Report BUDGET vs ACTUAL To Owner INTELLIGENCE LOOP · actuals & outcomes sharpen the model: a live control loop sharper forecasts → the model actuals vs forecast variance signal model refit · eval gate forecast → model a record that will not reconcile is held, not published. Every failure is classified business, system, or model, in the typed error catalog OUTSIDE THE LAYER: the residual invoices no layout matches yet · spreadsheet reconciliation · the manual tail concentrates here THE DATA LAYER · every field, every meter, one source HOW TO READ IT One intake spine on the left: EDI, portal harvest, mail and scan, and meter data land on one queue, classified and extracted against 300+ layouts, capture accuracy 65% to 99%. The reconciled model in the center normalizes, reconciles and forecasts on one base; a record that will not reconcile is held, never published. The data products on the right are cut from that reconciled truth: benchmarking, forecasts and accruals, carbon, budget-versus-actual. Payment is the byproduct, not the product. 1 invoice in 100 reaches a person. 290 keyers became 65 analysts. runs autonomously human-in-the-loop the reconciled model applied ML intelligence loop exception feed Live deployment · governed data layer

HOVER OR TAP ANY NODE FOR WHAT IT DOES · DRAG THE MAP TO PAN · 1 INVOICE IN 100 REACHES A PERSON · 290 KEYERS BECAME 65 ANALYSTS