Every January brings a fresh round of "AI predictions" pieces. This isn't one of those. Instead, here's a look back at what actually changed in artificial intelligence over the past year — the shifts that showed up in real products and real workflows, not just in demo videos.
Models got smaller, not just bigger
For a few years, the story of AI progress was simple: bigger model, better results. That story is now only half true. The more interesting trend in 2026 has been the rise of compact, specialised models that run efficiently on a laptop or a phone and still handle narrow tasks — drafting emails, summarising documents, tagging images — about as well as models many times their size. Teams building products started asking "what's the smallest model that solves this problem?" instead of reaching for the largest one available.
Agents that finish tasks, not just suggest them
The word "agent" was overused in 2025. This year, the gap between marketing and reality started closing. Tools that can plan a multi-step task, use software on your behalf, and check their own work became noticeably more reliable for bounded jobs: filling out a form across several tabs, reconciling a spreadsheet against an invoice, or triaging a support inbox. They still need supervision, and they still fail in ways that are obvious to a human but invisible to the system itself. The practical lesson for anyone adopting these tools: give them narrow, well-defined jobs with a human checkpoint, not open-ended authority.
Regulation stopped being theoretical
Several regions moved from draft frameworks to enforced rules this year, covering areas like disclosure of AI-generated content, data provenance, and higher-risk use cases such as hiring and lending. For businesses, this meant AI governance moved from a slide in a strategy deck to an actual line item — documentation, audit trails, and a named person responsible for a model's behaviour.
The interface changed more than the intelligence
Arguably the biggest shift wasn't in raw model capability at all. It was in how people interact with these systems. Chat windows became one option among several: voice interfaces got noticeably better at handling interruptions and follow-up questions, and "AI inside the app you already use" — a spreadsheet, an inbox, a design tool — became the default expectation rather than a novelty.
The most useful AI tools in 2026 aren't the ones that feel the most futuristic. They're the ones that quietly remove a step you used to do by hand.
What to actually watch next
- Cost per task, not cost per token. As agentic workflows spread, the meaningful metric is what it costs to reliably finish a task end-to-end.
- Verification tooling. As AI output enters more decisions, tools that check, cite, and flag uncertainty are becoming as important as the models themselves.
- On-device processing. Privacy expectations and smaller efficient models are pushing more AI work off the cloud and onto local hardware.
None of this is as dramatic as the headlines from a couple of years ago. But it's the kind of steady, unglamorous progress that tends to actually change how people work.
Have a correction or a follow-up question about this article? Get in touch — we read every message.