All storiesBullet Shorts

Customer story · Bullet Shorts

41% of real conversations became confirmed viewer feedback, at ₹3.93 a call

Bullet Shorts wanted to hear from people who tried the app and did not subscribe. A Hinglish voice agent asked them, and handed the answers back as data the team could query. The team is now building on the in-app agent too.

Sector
Microdrama app
Products
Voice agents, in-app agent
Language
Hinglish
Survey calls
June and July 2026
  • 3,115

    connected calls to 1,683 numbers, in seven days

  • 41%

    of real two-way conversations ended in confirmed viewer feedback (280 of 687)

  • ₹3.93

    per call (₹12,253 in all)

  • ₹40

    per confirmed viewer's feedback record

The job

Bullet Shorts needed to hear from trial users who did not convert, and to hear quickly. Its audience answers phone calls. It does not answer email surveys.

The first run, in June

Two agents ran first: one asked people why they had not subscribed, and the other asked what they like to watch. Reading the transcripts showed the conversations were useful, but the answers were not reaching the data fields, so there was nothing for the team to query. About a third of connected calls also ended on the opening line, within eight seconds.

Fixing the data came first, and by July the fields were being filled in.

The survey, in July

One survey ran as four campaigns, in Hinglish, with no one on the line but the agent. It checked that the person had watched the show, asked where they found the app and what should change, and recommended what to watch next.

Every call was scored against a five-field specification, so the team got fields to query rather than recordings to sit through: whether the person watched, where they found the app, what they would change, whether the recommendation landed, and why the call ended. Whether the person watched was filled in on 2,796 of the 3,115 calls.

What came back

The full funnel, so that 41% can be read against its base:

  • In 1,120 of the 3,115 connected calls (36%), the person spoke.
  • 687 calls (22%) became a real two-way conversation.
  • 280 of those 687 (41%) ended in confirmed viewer feedback, between 38% and 45% in each campaign.
  • The calls also brought in product feedback nobody had asked about, in viewers' own words.

What the pilot left to fix

A full survey was completed on 2.9% of all calls (90). Most people were busy when the call came, and this first script ended those calls when it could have offered a callback. The campaign that asked its key questions earliest converted best, at 45%, so the cheapest gain is asking sooner.

Now: the in-app agent

The team is now building on Graine's in-app agent, with the Kotlin Multiplatform SDK. The in-app agent lives inside an app: it chats and talks, remembers each person across conversations, can ring them like a phone call at a time they booked, and acts only through the actions the app declares.

That work is still under way, so there are no results to report yet.

See the in-app agent

See what it does on your calls.

Tell us who you need to reach and what you need to find out, and we'll show you how it runs on Graine. Or start building with your own account today.