Graaft

Designing Priya, an AI colleague you talk to

(Agent Design + Conversation Design + Interface Design)

Live · measurement pending

Priya, the studio’s SEO specialist, who is not a real person: a woman in a dark blazer standing in a home office, two monitors behind her showing traffic charts, a corkboard of printed reports on the wall, a desk lamp, a small plant and a linen curtain.

(Overview)Priya is one of our studio’s specialists and she is not a person. I designed what she does, how you talk to her, and the screens she works from.

At a glance

Client: Graaft
Sector: AI products
Year: 2026
Role: Design and build
Scale: 15 screens · 15 jobs
Status: Live · measuring

The problem

Ask an AI assistant to recommend a studio like ours. Nobody could tell us what it says.

A lot of people now ask an assistant for a recommendation instead of searching and scrolling. If it never names you, you have not slipped down a list. You are not in the conversation at all. Priya’s job is to find out whether we get named, by putting the questions a buyer would ask to the assistants they use, and recording what comes back. The tools everybody uses count visits to a website, which answers a different question. I built her as a colleague rather than a dashboard: a name, a job title and settled opinions of her own, modelled on the specialist you would hire to do this. A dashboard waits to be read. She has a view, and she tells you when it changes.

Set it up on a new machine, miss one setting, and it still runs, still charges you, and still hands back a confident answer with nothing behind it. Nothing goes red. A failure that costs money and looks like success is the worst shape a failure can take, so the setup check counts what it loaded and refuses to start when the answer is nothing.

Seen over Will’s shoulder at night: he sits at a timber desk in a dark studio, hand on the keyboard, talking to the Priya console glowing on a large display in front of him. — The console, being talked to rather than driven.

What I’d bring

I build the thing that can contradict me, then go looking for where it cannot.

Anything that measures you will eventually have to say something you would rather not hear, and the only version worth having is the one that still can. A failure that costs money and looks like success is the worst shape a failure can take, so the rule underneath all of this is plain: a zero means we looked and found nothing, and where nobody looked it says so in words instead. Then I went through my own build and found six places I had broken that rule, because testing whether a count is above zero feels like the right test and is not. Writing the rule is the part anybody can do. Auditing yourself against it, and then publishing the number, is the part that makes the rest of the screen worth believing.

She caught us out on the first day she ran, which is exactly the point of her. Every figure on the screen had quietly been about one country, because the build had assumed which market we sell into and nobody had ever checked. We sell into more than one. Nothing had looked broken, because a wrong number and a right one are the same shape.

A large display on the timber desk in the dark studio at night, the room lit by the screen. On it, the console’s Signal page: a table of markets with their impressions and clicks, a note that country was not collected until August 2026 so anything measured before that is marked absent rather than attributed to a market it was never assigned to, and Priya’s chat panel open over the page, answering a question about whether the zeros are zeros or failures with the line that the two zeros here mean different things. — The Signal page, and the months it refuses to guess at.

Outcomes

She is live, both founders use her, and every word she writes waits for a person.

She reads the brief you are standing on out loud, and answers the way you addressed her: speak and she speaks, type and she types. She runs the jobs, opens and closes things, and moves around the screens for you. She keeps a strategy, scores the work against it, and amends it herself when new evidence arrives, and tells you she has. When that evidence disagrees with something she said earlier, the evidence wins and she says so. She publishes nothing at all, and every change she recommends is carried out by a person. The findings themselves are the studio’s own commercial position and stay inside it. What I can show is her, and I would rather ship a screen that says nobody looked than one that says zero and lets you believe somebody did.

A phone held in one hand in the dark studio at night, its screen showing Priya’s chat panel with a question typed into it, ready to send. — The chat panel on a phone, mid-question.

Role

I designed her, built her and shipped her. The method came from elsewhere.

What I did

Designed and built her, alone
How she talks, and listens
The rules that keep her honest

What the team did

Dave, my co-founder, approves
The SEO method she runs on

Gallery

01 / 06

(The call)You hold the space bar to talk to Priya, and let go to send. I built the version that listens for her name first, then deleted all 25MB of it.

Listening for her name was the impressive version, and it worked. It also added 13.7MB to an application of 1.1MB, twelve times the whole thing, to save one keystroke at the start of a session. It misheard her name six ways, and since saying a phrase to send is a guess at what you meant, every mishearing was a send that silently never happened. Letting go of a key cannot be misheard.