Orson
In 2018 I tried to build an AI that would keep a whole company on brand. I paused it. That was the mistake.
- Client
- Independent venture
- Role
- Founder
- Timeframe
- 2018–2019
- Sector
- Applied AI
Pausing it was a mistake
I should have pursued it, and I should have been more ambitious.
It was 2018. I was building a corporate tone-of-voice assistant, before I had heard of agents and before anyone had a slop problem to complain about. It was too early. I couldn't get the talent, and carrying on would have meant raising money and going in much bigger. I was coming off the back of a couple of other ventures, so I went back to university instead.
If I had the time over, I would do something different.
Nobody noticed
Different people across a business put out documents, emails, messages and presentations. Some were off brand, some slightly off brand, and nobody noticed. I kept seeing it in my own design practice, and nothing would tell you.
Orson would. It was for anyone in a company who isn't naturally a writer, which is most people, and was meant to go from small businesses up to enterprise, where you might tone the character down. It checked that work was on brand: the right products and services, the right tone of voice, perhaps even security and privacy. It had a mini secretary function.
Picture a Tuesday morning in 2019. A salesperson has mostly built a pitch. Orson is with them, reviewing it slide by slide. Or it sits in the background and pipes up: that service has been renamed, that price has changed.
A very specific form of Grammarly, in a way. Or Clippy, if Clippy had known your brand.
It looked like a spell checker
The first version was a layer on top of a basic spelling and grammar checker that could hold a tone of voice, like a language setting: UK English, American English, IBM English. It was good at ruling words out, such as swear words, or product names. It didn't add a great deal beyond the default checkers, but it worked. We built something and we tested it.
Going further was the hard part. I tried the University of Edinburgh's informatics department, its AI researchers and its entrepreneurial networks, looking for people who had worked in the area and could join for little or no pay. There were very few of them, and it was hard to persuade them to build what often looked like not much more than a spell checker.
I struggled to find people who could really build with AI.
Even today we don't have it
Tone of voice, positioning and pricing are critical, and every business would love to have everyone aligned on them. Claude and ChatGPT are very capable, but you still have to install the skills and you still have to run them.
They do a fantastic job of masking how complex they are. But most users can't begin to think which skill they need, let alone judge the document that has just come back. Trust is hard to build when the model changes underneath you. For something that is meant to be a companion and a guardian, it needs a stability that current models don't offer.
I believe in a dedicated tool, built for one thing. Somewhere you go with a single question. Is it on brand? Does it look right? Does it feel right? Should we be sending this out?
It was called Gonk
It was called Gonk first, after the little robots that scurry around in Star Wars. We wanted it to be friendly, and Orson Welles was in the air at the time, which suited a product about writing. The character was far less formidable than Orson Welles, but it had a charm to it.
I might rebuild it, just for fun, because it is so much easier now. There is still space in the market for it.