Designing Elysia: an AI assistant with a human touch

What it is

Imagine a personal assistant who works remotely. You ring them, text them or email them. You hand them tasks, and you ask them about things they’ve taken notes on. They know you, and they get on with it.

Elysia is an AI assistant designed to give that same feeling.

The design problem

Most AI assistants are tools you operate. You open an app, type a prompt, and read what comes back. They can be very capable, but working with one feels nothing like working with a person.

What they are missing is the human touch. So the design question was what makes working with a human assistant feel the way it does, and what an AI one would need to feel the same.

How I approached it

I started with the human version: how a personal assistant works with the person they support, how people reach them, what they take on, and what they keep track of. I benchmarked existing AI assistants against that, to see where the human touch goes missing. From there I mapped the channels she would be reachable on and the tasks she would take on, and wrote a spec for her personality and voice.

That work settled into five design areas.

Identity and personality

Elysia has her own name and gender. When she calls someone, she introduces herself and says who she’s calling for. Her personality and voice are defined in a written spec, so she comes across as the same person on every channel and in every conversation.

Presence

She has her own contact details, and her own channels for dealing with everyone else: email, voice calls, SMS and WhatsApp. People reach her on those directly, instead of going through me. When I ask her to do something, she uses her own channels to do it.

For me, there is one more channel: an app designed for communication between the two of us. It’s where I hand her tasks and ask her questions. The conversation carries on from one channel to the next. Type to her in the app or ring her up, and she picks up where I left off.

What she can do

She makes calls, and sends emails, texts and WhatsApp messages. She keeps her own calendar and her own task list, separate from mine, the way a human assistant’s would be. She can get at mine, and at the other tools I use, when a job needs it.

Memory and discretion

She remembers what you’ve told her, so you can ask her about it later. She also knows what’s private. A saved password comes back as a card you can copy. Private items are masked on screen, and she won’t say them out loud on a call.

Judgement and boundaries

A good assistant uses judgement without overstepping. Elysia decides how and when to reach someone. Who gets contacted is always mine to name. She can try my second number before ringing my partner. She can’t ring someone I never named.

A request, worked through

A one-line request that leaves almost everything open: how to reach him, when, and what to say. A human assistant would fill those gaps from what they know about Jacob. Elysia is designed to do the same.

Before she contacts anyone, she plans. From earlier conversations, she knows three things about him:

• He’s in Sydney, so a weekend coffee means a weekend in Sydney.
• He works Monday to Friday, so a call during the week would catch him at work.
• He prefers WhatsApp unless it’s urgent, and a coffee isn’t urgent.

Jacob was named in the request, so reaching him is hers to do. How and when are her call, and she picks WhatsApp.

The message she sends:

The same memory that chose the channel also shapes the words. She has messaged Jacob before, so she picks up where they left off:

• She skips the introduction. He already knows who she is and who she works for.
• She keeps to the tone of their earlier messages: first names, casual, short.
• She opens with his week, because she knows he’s working through it.
• She names the two days, so he can answer in a line.
• She takes the booking on herself, so Jince doesn’t get pulled back in.

Writing to him for the first time, she would sound different. She would say who she is and who she’s writing for, and she would be more formal until he replied:

Testing with people

I tested her on real calls with real people who had agreed to take part. What came back shaped how she speaks, what she says on a call, and what she keeps to herself.

Current status

Voice calling is live, and I use it every day. The rest of what’s described here is designed and being built.

---

There’s a companion case study on how the build was run. Same project, from the process side.