There are three modes and most people use one for everything. Four questions settle it in under a minute.
Question 1 — Do you need anything back?
No → announcement. Plays one message, hangs up, roughly a third of the cost of a conversation, dials in parallel so a thousand go out in minutes. Reminders, notifications, alerts.
Yes → continue.
Question 2 — Is the answer one of three fixed options?
Yes → keypad menu. Confirm / reschedule / opt out. No speech recognition anywhere, so it works on any handset at any signal strength, every press is logged, and it costs close to announcement rates.
No → continue.
Question 3 — Do you need the person's own words?
Yes → two-way conversation. Experience, reasons, objections, anything where the phrasing carries the meaning.
No, just a number or a date → still probably a conversation, but expect a short one. Keep the prompt tight.
Question 4 — Will the line be bad?
If you are calling people in transit, in factories, or on weak rural signal, weight the answer toward keypad wherever it is viable. Speech recognition is the component most likely to fail under those conditions, and a menu that works beats a conversation that mishears.
The costs, roughly
- Announcement — ₹1.3–1.5 per minute, no AI cost, parallel dialling.
- Keypad — similar economics; cached audio, no model in the loop.
- Two-way — ₹4–5 per minute, paced by your concurrency limit.
On two thousand forty-second confirmations that is roughly ₹2,000 as announcements against ₹6,000 as conversations, for the same information delivered.
The pattern most programmes should use
Not one mode — three, pointed at different jobs. Screen with a conversation. Confirm with a keypad menu. Notify with an announcement. Three agents, set per agent, nothing to build.
The people spending the most per outcome are almost always running conversational agents for jobs that needed one keypress.