Inbound or outbound: where an AI voice agent earns its keep

A real-estate office in Hyderabad loses site-visit enquiries because the phone rings during lunch. A D2C brand shipping from Surat has hundreds of customers waiting for delivery updates. Both think "AI calling", but the first problem is about answering and the second is about reaching out. They need different set-ups, carry different risks and should not be started together.

Inbound: the caller chose to ring you

On an inbound call the person wants something from you. They are usually patient enough to listen, they expect to be asked questions, and they already know whose number they dialled. That makes inbound the gentler place to start.

Typical inbound jobs:

• Reception. Greet, find out why they are calling, answer common questions, book or reschedule an appointment.
• After-hours cover. Take a message, book a slot for the morning, or flag a true emergency to the on-call person.
• Overflow. Pick up when every line is busy or nobody answers within a few rings, so the call is not lost.
• Routing. Work out what the caller needs and transfer to the right person or department, with a note so they need not repeat themselves.

Outbound: you chose to ring them

On an outbound call you are interrupting someone. The first ten seconds decide whether they stay on the line, and the rules about whom you may call and when are stricter. In exchange, outbound work is where volume is: far more calls than a team can place by hand.

Typical outbound jobs:

• Reminders. Appointment confirmations, fee due dates, renewal dates, delivery slots.
• Follow-ups. "You enquired last week; do you still want a site visit?"
• Lead calls. Ring a new enquiry soon after the form is filled, ask two or three qualifying questions, and pass warm leads to a salesperson. See calling new enquiries while they still remember asking.
• Collections-style reminders. A polite, factual reminder that a payment is pending, with an easy way to pay or ask for time. The agent should never threaten, shame or imply consequences it cannot name, and should hand over to a person when the customer disputes or is in difficulty.

What each one needs to work

Outbound calls, especially promotional ones, are covered by Indian telecom rules on registered numbers, preference registers and consent. This post is general information, not legal advice; read the rules in plain words in the rules for automated and AI calls in India and check with your adviser before you start.

What goes wrong in each direction

Inbound failures are usually about knowledge. A caller from a Nashik clinic asks whether the doctor accepts a particular insurance, the agent has no answer, and it guesses. The fix is to write the answer down, or to tell the agent to say "I will have the clinic confirm that" and take a number. A second common failure is a dead end: the caller asks for a person and the agent cannot arrange one. Always build the transfer path, and a message-taking fallback for when nobody is free.

Outbound failures are usually about context. The call arrives at a bad time, reaches the wrong person (a relative picks up), or opens so slowly that the listener hangs up. Keep the opening to one sentence that says who is calling and why, ask whether it is a good time, and make leaving easy. Store opt-outs the moment they are spoken, and never ring that number again. A person who says "stop calling" and gets called the next day will remember it.

Both directions benefit from a disclosure that the caller is speaking to an AI. We discuss the options in should your AI agent say it is an AI?

The hand-off is the real design question

In both directions, the quality of the exit matters more than the quality of the happy path. On an inbound call the exit is a transfer to the right person with a short note of what was said. On an outbound call the exit is a clean ending: confirm what was agreed, say what happens next, and stop. A caller who has to repeat their name and problem to a human after the agent has already collected them will feel the agent wasted their time. Decide in advance which phrases and situations trigger a transfer, and who answers when it happens. Our post on handing an AI call to a person covers this in detail.

It also helps to decide what the agent does when it simply does not understand. Two attempts to clarify, then an offer to take a message or transfer, is a far better pattern than a third repetition of the same question. Callers forgive an agent that admits its limits; they do not forgive one that loops.

Can one agent do both?

Technically yes, and many businesses end up with both. A booking made by an inbound call can trigger an outbound reminder the day before, and a missed reminder can trigger an inbound-style rebooking offer. But the two need different prompts, different openings and different success measures, so treat them as two agents sharing the same calendar rather than one agent with two moods.

How to choose the first one

Pick the job where a mistake is cheap and the benefit is easy to see. A few questions help.

• Where do you lose the most? If calls go unanswered, start inbound. If your team spends hours on repetitive reminder calls, start outbound.
• How bad is a wrong answer? Booking and opening hours are forgiving. Medical advice and fee disputes are not.
• Is the data ready? Outbound on a messy list produces messy calls.
• Can you measure it? Choose something with a number you already track, such as missed calls per week or no-shows.

Where to start

1. Choose one call type and write its ideal five-line version by hand.
2. Decide the exits: when the agent must transfer, take a message or end the call.
3. Run it on a small share of calls, or on staff and friends first.
4. Read the transcripts and fix the prompt before widening it.
5. Add the second direction only once the first is steady.

You can hear an inbound agent in the live demo, and estimate cost per call on the pricing page.