Predictions made at the start of a year rarely survive full contact with events. The ones made about AI in 2025 were no exception. Reviewing what the forecasts said and what actually happened is an exercise that teaches more than the forecasts themselves.
The prediction that proved substantially right was around agentic AI. The framing that AI would move from question-answering to task-completing was accurate. By the end of 2025, genuine agent deployments in production were common enough that they stopped being news. Customer support agents handling multi-step queries without human escalation. Code review agents running as part of CI pipelines. Research agents aggregating information from multiple sources to produce briefings. The pattern was real, and the scale of deployment was larger than the more sceptical forecasters expected.
What got wrong was the nature of the breakthrough. Many predictions assumed the capability jump would come primarily from model intelligence. The systems that actually worked in production relied as much on tool design, workflow integration, and failure handling as they did on model capability. The agent success stories of 2025 were systems that were carefully constrained, that knew their own limits, and that handed off gracefully when they reached those limits. The dangerous agent operating autonomously at large scale was also real, but the organisations that shipped something that worked had invested heavily in the structure around the model rather than betting on the model alone.
Regulation was more measured than the most dramatic predictions. Several significant frameworks were proposed. Implementation moved more slowly than critics argued it should and more quickly than the industry preferred. The actual effect on what teams could build in most jurisdictions was less than the debates suggested it would be.
The multimodal shift was faster than most expected and the applications were different from what most imagined. The use cases that got the most attention in predictions were the obvious ones: image generation, video, voice assistants. The ones that created the most business value were less glamorous: document processing, technical drawing interpretation, recorded meeting analysis. The pattern of the useful being less exciting than the impressive held across the year.
What I carry into 2026 from watching this year is a strong prior toward things that are obviously useful in the current state of the technology over things that would be impressive if the technology were slightly better. The gap between current capability and impressive-if-slightly-better is regularly smaller than it looks from the outside. But the gap between current capability and obviously useful is also smaller than the hype suggests. Both gaps are worth understanding before deciding what to build.