The economics of an AI helpdesk — what we learned from Lumen

The unit economics, honestly

A human tier-1 support agent in our client's market costs around ₹35,000 to ₹50,000 per month, fully loaded, and resolves 80 to 120 tickets per day depending on ticket complexity. That puts the human cost per ticket at roughly ₹15 to ₹30 — a wider range than people expect, driven mostly by ticket complexity and the average handle time the team can hit.

A model-resolved ticket on Lumen costs us about ₹1.20 in inference, ₹0.30 in vector-database queries, and ₹0.50 amortised infrastructure — call it ₹2 per resolved ticket. That looks like an order-of-magnitude saving, and it is — but only on the tickets the model actually resolves cleanly. The tickets it gets wrong, or that loop without resolution, are more expensive than human-handled tickets would have been — because they still end up at a human, who now has to undo the bot's framing of the issue.

The 62% deflection number, decomposed

Of every 100 tickets that hit the system, 62 are auto-resolved without escalation. The other 38 break down like this: 24 are escalated by the model itself (it recognises uncertainty and hands off with a clean summary), 9 are escalated by the customer (they're not satisfied with the bot's answer and ask for a human), and 5 are escalated post-resolution because the bot's answer was wrong and the issue came back later.

The 5% wrong-and-came-back rate is the one that matters most for trust. We spent more engineering time on reducing that number from 9% to 5% than we did on any other improvement. The mechanism: an eval harness that runs every model change against 500 real historical tickets with human-judged correct answers, and an automatic rollback if the answer-quality score drops below a threshold.

• 62% auto-resolved (the headline number)
• 24% model-escalated with clean context handoff
• 9% customer-escalated (asked for a human directly)
• 5% wrong-and-recurred (the metric we watch hardest)

Where AI helpdesks make things worse

We've seen three failure modes that turn AI helpdesks into a net negative. The first is over-confident wrong answers in regulated or high-stakes domains — a returns policy stated incorrectly to a customer with a refund claim, a medication query answered with non-medical advice. The model should refuse to answer questions outside its competence, and most off-the-shelf deployments don't.

The second is the doom-loop — when the customer asks a question, gets a slightly-wrong answer, rephrases, gets a slightly-different-but-still-wrong answer, and eventually rage-types ESCALATE TO HUMAN. By the time a human picks up that conversation, the customer is furious. The fix here is to escalate aggressively on signals of frustration (sentiment analysis is good enough now), not to optimise for deflection at all costs.

The third is brand-voice drift — the model says things your brand wouldn't say, in a tone your brand wouldn't use, sometimes about competitors or industry topics where you actually have a position. Style and policy guards on every response are essential, not optional.

What the support team actually thinks

We expected the human support team to be defensive about the AI. They weren't — once the deflection had been running for a month, they were the AI's strongest advocates. The reason is straightforward: every ticket that reaches them now is interesting. The repetitive password-reset and shipping-status work is gone. The work that's left is the harder, more diagnostic, more interesting kind — and the team enjoys it more.

Retention in the support team improved meaningfully after Lumen went live. We didn't predict that and we can't fully attribute it, but the correlation is strong enough to mention.