Platform cost by user segment

Free acquisition is healthy and steady at ~4,400 a month. Paid conversion has fallen 70% — from 33 new paid accounts in January to 10 in August. Premium is the worst hit: 26 in January, 2 in August. The free-to-paid ratio has widened from 145:1 to 461:1. Nothing about the free tier changed; what changed is the rate at which we convert off it.

Free users take 97.2% of AI calls and 93.3% of the AI spend, and the free share has climbed every month since February. But keep the magnitude honest: the entire eight-month AI voice bill is about $5,580 — roughly ₹4.9 lakh, against ₹94.5L of sales in the same period. This is a ~5% cost line, not an existential one.

There is no abuse problem. The heaviest free user in the last 30 days took 29 calls and cost $1.26. The single most expensive free user cost $4.20. Cost per active user is flat at roughly $0.65–0.89 across every usage band, so no gating threshold would recover meaningful money. All 25 of the top consumers are on the FREE plan — but that is unremarkable when 99.8% of accounts are free.

Note the direction of causation: these are outbound calls we place to them, not a service they pull on demand. Heavy AI usage is us dialling repeatedly, not a user consuming a free resource. A usage cap would be capping our own outreach.

509,768 assistant messages since January, and the trend is falling — down 49% from the April peak. Whatever the free tier is costing in LLM inference, it is shrinking, not growing. Per-message cost is not stored, so this cannot be converted to rupees from the database. That number has to come from the provider bill.

Plan segment from entitlements.activePlan.planName; individual-service buyers are entitlement holders with no plan. Free is the remainder of platform_subscriptions. A user is counted in their highest tier only.

An inbound call lasts 4.4× longer than an outbound one — 145 seconds against 33. Inbound is 5.6% of call volume but 16.8% of the cost, because when someone rings us they actually talk. Outbound is cheap per call precisely because most of it never connects.

The direction of the call tells you which segment it is. Free users take 98.5% of outbound — we chase them. Paying users generate 41.8% of inbound from 0.34% of the accounts — they ring us.

A Premium user calls in 678 times more often than a free user. That is the clearest engagement signal in the entire dataset — and it costs $1.21 per user across eight months against ₹34,999 of revenue. Inbound volume is a retention asset, not a cost problem. Premium alone accounts for 26% of all inbound calls from 200 accounts.

1,009 counsellor hours in eight months — this is the real cost line, and it dwarfs the AI. At a fully loaded ₹400/hour that is roughly ₹4.0 lakh of human time, against ₹4.9 lakh of AI voice. Comparable in rupees, but the AI number covers 80,600 calls and the human number covers 34,829. Per connected conversation, a counsellor costs roughly 20× what the AI does.

Only 45% of coach dials connect. Inbound coach calls connect at 30% — worse than outbound at 48% — which suggests missed inbound calls are being logged. That is a service-level problem worth its own look. Per-segment attribution of coach hours is not computed here: argonaut_counsellors_call_history keys on lead_id while segments key on gamePlayerId, and the bridge is not indexed for a platform-wide join.

Anika wrote 187,651 replies in 2026 — 28% of everything in the Spaces. Volume peaked in April and has roughly halved since; Anika's share of the conversation dropped sharply after April, which looks like a deliberate change worth confirming with engineering.

Voice cost is a rounding error against every paid tier — under 0.5% of revenue in all four. The heaviest users of the platform are also the most profitable ones by a wide margin. Nothing here supports throttling a paying segment, and the free segment costs $0.017 a head over eight months.

The one genuine exposure is free outbound: $4,204 spent dialling people who have never paid, producing 78% of total voice spend for the segment with no revenue attached. That is the line to manage — not by closing the platform, but by dialling fewer, better-qualified people, which is exactly what the ₹299 cohort analysis argues for.