Define the metrics ladder for the support copilot, rebuild the unit-economics model on tokens, and write a pricing one-pager — hybrid versus usage — with bill-shock safeguards and an India-margin scenario. The draft metrics sheet conflates deflection with containment, inflating the business case; the reconciled model is the graded artifact.
Kabir's support copilot is working, and now finance wants the business case while sales wants a price. An analyst has drafted the metrics sheet, and it tells a beautiful story: 62% of conversations "contained by AI," huge agent-cost savings, healthy margins. The story has a definitional rot at its center — silence counted as success — and every number downstream inherits it. Reconcile the metrics first; only then does the pricing page mean anything.
DRAFT METRICS SHEET (analyst v1) - monthly, 60,000 AI conversations
- "Containment rate: 62%" = conversations where no human agent was
involved AND the customer sent no follow-up within 24 hours.
62% x 60,000 = 37,200 conversations "resolved by AI".
- Agent handling cost: Rs 85 per conversation (loaded).
- Claimed monthly saving: 37,200 x Rs 85 = Rs 31.6 lakh.
- AI cost: Rs 12,500/month total -> "cost per AI resolution: Rs 0.34"
- Suggestion acceptance (agent-assist mode): 71% of drafts sent
with zero or minor edits.
- Complaint rate: 0.8% of AI conversations get a formal complaint.
OPS DATA EXTRACT - same month, same 60,000 conversations
- Of the 37,200 "no agent + silent 24h" conversations:
- 9,000 customers opened a NEW ticket on the same issue within
7 days (reopens - agent handled, avg cost Rs 110: repeat
contacts run long)
- 4,200 customers called the phone line within 7 days on the
same issue (shadow-channel escape, avg cost Rs 140)
- 24,000 stayed closed at 7 days; CSAT sample on these: 4.2/5
- Token usage: avg 6,000 input + 1,200 output tokens/conversation
- Model price (mid tier): $2.50 per 1M input, $10 per 1M output
- USD/INR: 84. Agent-assist minutes saved per accepted draft: 2.1
Attack the definition before the numbers. Write the two definitions side by side: deflection — the customer did not reach an agent; containment — the issue was actually resolved without an agent, verified by no reopen in a window and acceptable CSAT. Now classify the draft's "containment rate": which one is it really measuring? Note the tell: "no follow-up within 24 hours" measures silence, and silence has three causes — resolved, gave up, or went to another channel. Only one of the three is success.
Reconcile containment with the ops extract. True containment = (37,200 − reopens − shadow-channel escapes) ÷ 60,000. Compute it. Then recompute the savings honestly: contained conversations save Rs 85 each, but reopens cost more than baseline (Rs 110 repeat contact) and shadow-channel calls cost Rs 140 — count the excess over Rs 85 as negative savings the draft hid. Produce the corrected monthly saving and put it next to the claimed Rs 31.6 lakh.
Rebuild cost per resolved conversation. First verify the AI bill itself from tokens: 60,000 × (6,000 × $2.50 + 1,200 × $10)/1M in USD, then into INR at 84. Compare with the draft's Rs 12,500 — is the draft's AI cost even right? Then compute the metric that matters: AI cost ÷ truly contained conversations — not ÷ deflected. Both errors move the same direction; note how the draft's Rs 0.34 was manufactured.
Build the metrics ladder. Four rungs, each with its definition, data source, and what it gates: (1) suggestion acceptance (agent-assist quality — leading indicator); (2) true containment (with the reopen window and CSAT floor in the definition); (3) complaint rate and safety incidents (guardrails — can veto everything above); (4) cost per resolved conversation (the economics rung, computed on true containment). Write the ladder rule: a rung may only be reported alongside the rungs below it — acceptance without containment, or containment without complaints, is how AI features lie in QBRs.
Price the feature — two structures. The copilot sells as an add-on to Kabir's field-service customers. Model both: pure usage at Rs X per AI-handled conversation, and hybrid at a base fee including an allowance, plus per-conversation above it. Anchor X to value and cost: true cost per resolved conversation (step 3) is the floor; a fraction of the Rs 85 agent cost the customer avoids is the value anchor. Propose numbers and show margin at 1,000 / 5,000 / 20,000 conversations per customer per month.
Add the bill-shock safeguards. Usage pricing for AI dies on the first surprise invoice. Specify: usage alerts at 80% and 100% of allowance; a monthly hard cap with graceful degradation (copilot falls back to agent-assist drafts rather than going dark); anomaly detection for volume spikes (a bot loop at a client should page you, not bill them); and one-click usage reports. Each safeguard gets one line on the one-pager — customers do not buy usage pricing, they buy bounded usage pricing.
Run the India-margin scenario. Domestic customers pay in INR at roughly 40% of the US price point, but token costs stay in USD. Rebuild the step-5 margin table at the India price. Find the point where the hybrid base fee stops covering the allowance's token cost, and state the design response: smaller included allowance, routing India traffic to the economy tier where the eval allows it (Lab 12's frontier work earns its keep here), or a floor price. A pricing page that ignores currency asymmetry ships negative margins to your fastest-growing market.
Write the one-pager and the confession. One page: the metrics ladder with this month's honest numbers; recommended pricing structure with the margin table; safeguards; India scenario. Then the confession box: what the draft claimed, what reconciliation showed, and the one-line definitional rule that prevents recurrence. Numbers a CFO can check beat numbers a CFO likes.
Deflection counts who never reached an agent; containment counts whose problem was actually solved — every AI support business case that confuses the two is overstated by exactly the customers it failed twice.
A complete lab has all of the following:
AI feature economics stand on definitions before arithmetic — containment verified by reopens and CSAT, costs rebuilt from tokens, savings that charge failures as negatives. Price on the honest floor with bill-shock bounds, and check the India margin before the sales deck does.