SmartSupply AI

Billing Dispute v3.1 · Steering Section 3.1

Source

HITL escalation esc-4471 (Sarah Thompson) — operator resolved a verbal tariff agreement claim by issuing goodwill credit. Pattern matches 3 similar escalations in the last 30 days.

Proposed change

Add decision pathway to Section 3.1: when customer claims verbal agreement, agent should autonomously query call recording API. If verbal confirmation found, auto-approve goodwill credit up to tariff differential value. Escalate only if no recording evidence or value exceeds £150.

A/C Optimiser v3.0 · Decision Tree Section 2.3

Source

Pattern detection — agent set contradictory temperatures in connected zones (e.g. cooling one room while heating an adjacent open-plan area) in 5 contacts over 2 weeks. Each required HITL override.

Proposed change

Before applying any setpoint change, agent checks all connected zones (adjacency map from BMS). If the new setpoint conflicts with an adjacent zone's active target by more than 4°C, warn the customer and offer to harmonise both zones. Eval suite: 4 new scenarios added, 3 passing.

Retention Intervention v1.5 · Authority Rules Section 5.2

Source

Repeated HITL escalations (4/week) where the operator always approved rate matching when the competitor quote was within 15% and the customer's retention value exceeded the annual cost difference. 100% approval rate indicated unnecessary escalation.

Change deployed

Agent may now autonomously match competitor rates where: (a) competitor quote is verified, (b) rate difference is within 15%, (c) customer retention value exceeds annual cost delta. Logged with full calculation. Deployed 10 Jul, canary passed 12 Jul, full production since 13 Jul.

Tariff Modeller v1.8 · Decision Tree Section 4.1

Source

Eval suite failure — tariff comparison scenario for customer with solar PV export failed after knowledge base update added new SEG (Smart Export Guarantee) rates. Agent modelled gross consumption instead of net, overstating annual cost by £400-800.

Proposed change

When customer has generation meter (G-MPAN) on account, agent must use net consumption (import minus export) for tariff modelling. Add check: if generation meter present and export > 0, apply net calculation. Flag if export data is unavailable rather than defaulting to gross.

Vulnerability Detector v1.2 · System Prompt Section 3

Source

Complaint root-cause — customer (Maria Santos) was in financial hardship for 2 months before being identified. Payment patterns showed: missed payment, partial payment, late payment in consecutive months. Agent only flagged when customer explicitly mentioned hardship in conversation.

Proposed change

Add payment pattern monitoring to detection triggers: if customer shows 2+ of (missed payment, partial payment where previously full, DD cancellation followed by manual attempt, payment date shifting later each month) within 60 days, proactively flag as potential hardship and offer early support before the situation escalates. Eval suite: 6 scenarios added covering true positives and 3 false-positive edge cases.

Energy Concierge v5.2 · Tone Guide Section 2.4

Source

AI pattern detection — 73% of the HITL spike (4.1% rate vs 2.3% baseline) traces to customers querying the new "green levy" line item on July invoices. The billing agent's current steering covers the levy but provides a policy reference rather than a plain-English explanation. CSAT on levy-related contacts: 2.1/5 vs 4.3/5 baseline.

Proposed change

Add conversational explanation to Tone Guide Section 2.4: "The green levy is a small charge that funds renewable energy projects across the UK. It's set by the government, not by us, and applies to all energy suppliers equally. For your account it works out to about £X/month." Replace the current response which quotes the policy document verbatim.