Insights.
Opinions, not brochures.
What we think about AI, data and business software, with the reasoning in plain sight. Written by people who build the systems, so not neutral: we say so where it matters.
- Who answers for an AI agent’s mistake?People talk about AI agents as if all that is missing is a better model. But what holds back fully automated business processes is not technical: it is that it isn’t clear who answers when an agent gets it wrong.
- Software without a screenIn September 2026 Salesforce introduced AIforce, a layer that makes its data and functions reachable by applications and agents without going through its interface. More than news about Salesforce, it is a signal: business software is giving up its own screen.
- Write software as if it were already publicThere is a test I use when designing a system: if tomorrow the code, the database schema and a copy of the data ended up online, what would be lost? If the answer depends on nobody looking at them, the design already has a problem.
- MCP is an interface, not an add-onWhen a system already has a web interface and APIs, adding an MCP server feels like the last lap: a few functions exposed to an assistant. That is a mistake of perspective. If an agent can do something the interface doesn’t allow, or the other way round, you have two systems with different rules.
- AI is already in use at work, but nobody chose itMany companies have chosen to wait before adopting AI, and often they are right: data first, tools second. But there is a consequence that goes largely unnoticed. If people adopt new tools with little friction and the company hasn’t named one, AI isn’t postponed: it gets used anyway.