Building AI Workflows That Survive Platform Risk
I’ve been on the road again this week, travelling to Portland, Oregon for the Independent Book Publishers Association Publishing University, where I’m speaking today. One quick observation, informed by email exchanges with my friend and his AI agent while I was travelling, and several conversations at the convention.
For the last couple of years, most of the discussions about AI have assumed one of two futures: a bubble-bursting collapse if you believe the critical view, or an up-and-to-the-right trajectory with models improving and costs falling. Either of those scenarios could prove broadly true, but they are not guaranteed. It’s more likely that we’re going to see something between them that’s a little messier.
The experience of recent weeks highlights some of the potential issues. More and more friends and colleagues are using AI models—particularly Claude—and seeing significant results. But over the same period of time, I’ve seen outages, changes to usage limits, subscription tiers and token pricing. This volatility is not unprecedented: startups built businesses on the Twitter API, only to see the rules of the road changed. Many publishing brands became over-reliant on social networks for their audience, only for algorithms to be more aggressively monetised. As consumers, we’ve all benefitted from inexpensive prices because startups use investor money to subsidise pricing in the expectation of raising prices once they have built market share. For three weeks in a row, my newsletter has featured stories about rising token prices, suggesting we’re leaving the Uber phase of AI pricing. All of this volatility makes it hard to build businesses and workflows around AI tools.
How can individuals and businesses mitigate against that risk? Many of the smartest people I know are converging on an interesting pattern of AI use. They are not building workflows inside Claude, ChatGPT or other specific models. Instead, the model is providing a harness or interface, through which they access their own systems. Personal information, where the real value exists, sits elsewhere in their stack. For example, my notes and writing exist as Markdown files saved in the cloud, and my personal data in Google Workspace, accessed through MCP. Claude Cowork and Claude Code access, interact with and update that data. But if Claude became unavailable, prohibitively expensive or just outpaced by a competitor, I can switch out the harness without changing the underlying system.