Safeguard Global hires and pays people in countries where its customers have no legal entity: 187 of them, for 1,500+ organisations, supported by 400+ in-country experts.
The expertise is the product. Employment law, tax, benefits, notice periods: different everywhere, changing constantly, and living in the heads of those experts.
Which means every customer question waits for a human, and growth means hiring more humans.
We took two bets on changing that, both early. I ran them with three designers.
Before either, I mapped the service front stage and back stage, you can't remove a wait until you know where the waiting actually happens, and set up a research practice: five customers on a standing rhythm, giving us feedback every few weeks rather than at the end.
ChatSG put Safeguard's HR knowledge base behind a chat interface. Employment law, benefits, hiring, policies: asked in your own words instead of hunted through documentation.
This was 2022. There were no patterns to copy and the models were confidently, plausibly wrong in exactly the place the product couldn't afford it.
The example that reset our thinking: ask about public holidays in Spain, get a fluent answer blending national holidays with Catalan ones. Reads right. Isn't. In an employer-of-record product that isn't a bad answer, it's a payroll error.
The fix wasn't better prompting, it was narrower ground. Answers came from Safeguard's own maintained knowledge base rather than the model's general knowledge of the world. And the interface opened with a default prompt that steered people toward what that knowledge base could actually answer. An assistant that's reliable about your jurisdiction and honest about its edges beats one that's fluent about everything.
Global Workforce Analytics let customers compare employment costs across markets, the kind of question that previously meant emailing someone.
Every instinct we had pointed at an interactive map. Salary ranges by country, colour-coded, explorable.
Then we put it in front of the five customers. They wanted a search box and a table.
They weren't exploring, they were checking a number to put in a budget. We built the table. The research practice was young and had already stopped us shipping the wrong thing, which was the more valuable outcome than the feature.
Both are live in the Safeguard platform today, available with a free account. app.safeguardglobal.com ↗
I also sat down with our CTO to talk through ChatSG on camera, worth watching if you want the unpolished version.
ChatSG, explained · with Safeguard Global's CTOIn 2022 the hard part of AI wasn't making it clever. It was making it narrow enough to be right.