By Pavan Kumar T V · 30 August 2026

A wellness coach was drowning in her own inbox.

She posts, something lands, and a stranger types four words into Instagram at eleven at night: I want to talk. That message is the whole business. The programme, the group class, the practice, all of it starts there, at an hour nobody chose.

She had an agency handling replies. They were not bad at their job. They just could not supply the one thing the message actually needed, which was not speed.

It was the feeling that somebody had heard you.

The mismatch that is nobody's fault

For her, that message is a helping instinct. Somebody reached out, something is going on in their life, you answer.

For an agency it is a task in a queue, worked when the queue gets worked. And to be fair to them, leads like this are reactive by nature. Someone asks at eleven at night and needs an answer at eleven at night. Nobody outside the practice, on a monthly retainer, is going to feel that at eleven at night.

I have seen this exact shape inside enterprises with a thousand people in them. The person who cares most becomes the constraint, and everything you bolt on to relieve them quietly removes the thing that was working.

So the requirement was never "automate her replies." It was: give a stranger, in the moment they reach out, something that feels the way she would feel if she were free.

In a wellness practice, warmth is not the packaging. It is the product.

First I called somebody else's assistant

Before building anything I rang a vendor's live voice agent, as a customer would.

It cut in on every pause. Every time I stopped to think, it took the turn. Did you say that? Did you mean this? And it had an agenda. I said plainly that I wanted to book an appointment, and it kept pushing me to explain what I wanted from my life.

Two minutes of that told me more than a week of reading would have.

An agent that interrupts is not a rough edge on an otherwise fine product. In this business it is the exact opposite of the service being sold. Somebody who has finally worked up to typing four words at eleven at night does not need to be talked over by a machine with a script to get through.

Where the warmth actually lives

My assumption was that this was a prompt problem. Write kinder instructions, get a kinder agent.

That assumption is wrong, and finding out why was the most useful hour of the project.

Whether an agent barges in is decided underneath the model, in the layer that judges whose turn it is. No wording ever reaches it. An agent configured to jump in at the first opportunity will jump in no matter how gracefully its instructions ask it not to.

The complaint I had after that phone call was actually four separate complaints, living in four different places:

• It cuts in when I pause. That is turn detection, below the model.
• It keeps re-confirming what I just said. That one is the prompt.
• Dead air before it answers. That is latency, and it needs a filler.
• It cut me off mid-sentence at the end. That is the call cap, with no closing line configured.

Only the second one is a prompt problem. Treating all four as a single "make it sound more natural" instruction is why so many of these things sound the way they do.

So the turn-taking went from eager to patient, and the wait before it assumes you are done went from five seconds to eight. Not an arbitrary bump. The questions are reflective, and somebody deciding whether to say the true thing out loud takes longer than somebody reading out a postcode.

Here is the number that stayed with me. These systems are around 99% accurate at recognising that a person has finished speaking, and around 87% at recognising that a person has not. Every miss in that second number is an interruption.

Knowing that someone is still gathering themselves is the harder problem. For machines, and if I am honest, for most people in most meetings I have ever sat in.

What you tell a voice that has to sound like it cares

The instructions that produced the biggest change are not technical, and none of them are clever.

One short sentence per turn. One question, then stop talking.

Acknowledge in two or three words. Vary them, and never use the same one twice in a row, because repeating a phrase is the single thing that most makes a machine sound like a machine.

Never repeat their own words back at them. They know what they said.

If an answer is vague or half-finished or rambling, take it and move on. One attempt per question, no asking them to be clearer. The reasoning is written into the file: an imperfect note costs the coach nothing, and interrupting costs the caller their train of thought.

While they are talking, you are not talking. The only reason to cut in is if they have described a crisis.

And the line I would put on a wall in most offices: if you are unsure whether they have finished, they have not.

Read as a list, that is not a configuration. It is a description of a person who is good at listening. The genuinely interesting thing is that it had to be written down at all, and that writing it down is now most of the work.

Warmth includes not lying

The vendor's own template, the one they hand to customers, has the agent say "thank you for that, I've noted it down" after taking your details. It has not noted anything down. It cannot.

That sentence would cost her a client the first time somebody turned up expecting to be remembered.

So the rule is blunt: an acknowledgement is not a promise. Never say you have written it down, saved it, noted it or added it anywhere. You have not. You cannot put anything in a calendar and you cannot take money, so never say a thing is booked, confirmed or paid. Hand it to her team instead.

Asked whether it is human, it says immediately that it is an AI assistant. No hesitation, no cleverness.

And if somebody describes self-harm or a crisis, everything stops. Warmly point them to a doctor or a helpline, and end the call gently. That is the one and only interruption permitted in the entire design.

One more rule survived from the first draft to the last: do not practise her technique on the person you are speaking to. An assistant that starts coaching is worse than no assistant at all.

The same vendor template also instructed the agent never to end the call until the customer says goodbye. That is engagement optimisation wearing empathy's clothes, and it went straight in the bin.

The voice is not a detail

People choose a coach roughly the way they choose a friend. The voice is a large part of why they chose this one.

So there is a cloned voice per purpose, rather than one voice for the account. Her own voice on the one-to-one call, where the person has already paid to speak with her and it should sound like her. A different voice for the group masterclass, because that is a different relationship and pretending otherwise is a small dishonesty.

The voice is pinned to the prompt it belongs with, so it travels with it and nobody can quietly swap it for a stock voice from a dashboard.

There is a gate in front of all of it. Her cloned voice goes nowhere near a real person without her recorded consent behind it, and the call opens by saying it is an AI assistant and asking whether recording is okay. That is the one line in the entire script that must never be missed.

Then there is the part nobody outside India would think to check. The agent opens in English and switches to Tanglish the instant the caller does, without announcing the switch and without ever asking which language you would prefer. Offering a language menu burns ten seconds of a two-minute call and makes a nervous person do admin.

The register is pinned too, and this is where the local knowledge earns its place. Always the respectful form of "you", never the familiar one, never the ceremonial one. An educated Chennai professional on a work call. Numbers, dates and prices stay in English, because nobody in this city says the twenty-ninth in Tamil. Get the register wrong and you either sound like a news reader or like a stranger being far too friendly, and both break it instantly, for completely different reasons.

Patience has a price

Listening properly costs seconds, and the call is capped at two minutes. Patience and the cap now pull against each other.

The answer is to ask one fewer question, not to wait less. A call that finishes beats a patient one that gets guillotined mid-sentence.

There is an assistant that sounds right in a rehearsal and an assistant that sounds right to somebody having a bad week at eleven at night. Only the second one is worth anybody's time, and everything above is an argument about which one you are building.

Where the human stays

Half of this engagement was not building anything. It was sitting with her and showing what is now within reach of a practice her size, because she had no way of knowing any of it existed. People running a good small practice rarely do. They are busy being good at the practice.

The other genuinely useful deliverable was a short list of the things she wanted that cannot be had. A private community where members engage but cannot see each other's numbers, which the platform simply does not offer. Reading her Instagram messages directly, which is a calendar problem rather than a technical one. A cloned voice without recorded consent. Each one written down, each one ending with a line saying this is your call and not mine.

When building is cheap, the expensive mistake is building the right thing on top of an assumption that was never true.

The one-to-one is still entirely her. She talks, she comes back, she writes the note, she decides by hand who joins the community. That boundary is deliberate and it is where the automation stops.

She is now hiring a junior intern to oversee the tools. Worth saying plainly, because the reflex assumption is that this removes a person. It changed the job from typing the same warm reply forty times a day into watching a system and stepping in where judgement is needed, and that is a better job than the one it replaced.

The assistant's entire purpose is to be genuinely decent for ninety seconds and then get out of the way.

Warmth was never the thing we automated. It was the thing the whole build existed to protect.