I don't love predictions. They're usually wrong, and the confident ones are the most wrong of all. But after a year spent putting generative AI into real newsrooms at real scale, a few things have stopped feeling like guesses and started feeling like gravity. So here are the bets I'm willing to put my name on for 2026.
1. Agents move from demos to plumbing
For two years, AI agents have mostly been impressive on stage and brittle in production. In 2026 the interesting work gets boring: agents become quiet infrastructure that handles narrow, well-scoped tasks — tagging, packaging, monitoring, routing — inside workflows people already trust. The winners won't be the flashiest autonomous demos. They'll be the ones that fail gracefully and hand control back to a human at the right moment.
2. Retrieval beats raw model size for most real products
The gap between frontier models keeps narrowing for the tasks most businesses actually need. What separates a product that works from one that embarrasses you is increasingly the retrieval layer — how well you ground the model in your own trusted, current data. Expect more teams to spend 2026 investing in their content, metadata, and evaluation pipelines rather than chasing the next few points on a leaderboard.
3. "Human in the loop" becomes a design discipline, not a disclaimer
Everyone says it; few design for it. In 2026 the phrase starts to mean something concrete — clear handoffs, legible confidence, easy overrides, and interfaces that make the human's judgment faster instead of just adding an approval step. The organizations that treat oversight as a first-class part of the product, not a compliance checkbox, will ship AI their teams actually adopt.
4. Trust becomes the real moat
When anyone can generate infinite plausible content, the scarce thing is a reader's belief. I expect the publishers and platforms that invest in provenance, sourcing, and transparency to pull ahead — because it's the only durable differentiator left, not because it's virtuous. Trust is turning into a competitive advantage you can measure.
Ask me at the end of 2026 how many of these I got right. Some of them won't age well. But I'd rather be specific and occasionally wrong than vague and safely useless — that's usually where the useful conversations start.