Finding an available partner at a workable price used to take 30 minutes of phone calls. It now takes three. The bot costs 25% more per call than the person it replaced — and that was still the right decision.
Allocating non-standard work means calling partner after partner: are you available, will you do it at this price.
Messaging was too slow and too quiet for this work — around 60% of these cases needed a phone call anyway.
The calling was a dedicated team's job. Repetitive, and identical on every call.
The team had to be staffed for peak demand, because the work arrives in spikes.
Most of the time they were idle or doing filler work. The cost was there whether calls were needed or not.
Ten partners could not be called at once without ten people calling at once.
Every call follows the same shape, and resolves into one of four outcomes.
The bot states there is a new project, gives the commercials, and reads a short scope brief. This is the part that was identical on every human call — the same information, delivered the same way, dozens of times a day.
The partner's answer lands in one of four buckets: accepts at the quoted price, counter-offers higher, doesn't understand the job and wants a human, or isn't interested and may want calling back later. Everything is captured as structured data, not as a note someone has to remember to write.
Outcomes feed a collection queue. Straight acceptances are allocated directly in the platform. Anything needing judgement — a price negotiation, a confused partner — routes to a person with the conversation already on file.
Escalation rate: ~30% of calls reach a human. Two people retained to handle them.
Language was the deciding factor in vendor choice. The bot runs on a model trained on Indian data, which handles Indian accents and Indian English materially better than the general-purpose alternatives — and switches language mid-call on the partner's cue. In practice it opens in English in the south and Hindi in the north, and follows comfortably if a partner switches to Punjabi or Bengali. Translation quality is not best-in-class. Conversations complete and partners do not drop off, which is what the job needs.
| Metric | Before | After |
|---|---|---|
| Time to find a partner | ~30 minutes of sequential calls | ~3 minutes |
| Call capacity | One call per person, per moment | ~10 simultaneous from a single number |
| Peak throughput | Bounded by team size | 500 calls/day |
| Cost per call | Lower | ~25% higher — and not the point |
| Cost structure | Fixed, sized for peak, idle in between | Low fixed + variable, paid only when calls happen |
| Standing cost of the function | Full team, always on the books | ~55% lower |
| Record of the conversation | Whatever the caller wrote down | Structured outcome on every call |
"When a cost is fixed and sized for a peak that rarely arrives, the saving is in the shape of the cost, not the price of the unit."
Per call, the bot runs about 25% more expensive than a person making that same call. Judged on unit price, it is a worse deal, and a unit-price comparison would have killed it.
But the calling team was never a per-call cost. It was a standing cost that existed whether or not there was calling to do — staffed for the busiest day, idle on the rest. Replacing it converts a large fixed cost into a small fixed cost plus a variable one that is only incurred when a call actually happens. That is roughly 55% off the standing cost of the function, paid for by a unit price that went up.
The honest limit: about 30% of calls still reach a human, mostly price negotiation, where the location- and context-dependent judgement is not in the bot yet. That is the next increment, not a hidden failure.
A clinic needing cover calls down a list until someone says yes. The same brief every time, the same four answers. The bot works the list in parallel and hands back a filled shift or a shortlist to negotiate.
Chasing quotes and lead times across a vendor list before a bid. Rate and timeline captured as structured data rather than a dozen inbox threads to reconcile by hand.
A cancellation at short notice needs the waitlist called fast, in order, until someone takes it. Parallel calling turns a slot that would have gone empty into a booked one.
Any network of independent contractors where a job has to be offered, priced and accepted before work can start — and where the offering is identical every time.
90-minute AI Opportunity Audit. No pitch — just findings you can act on.