We spent three days at London Tech Week with a single filter on every conversation: what's actually shipping in production AI, and what's still keynote theatre? Here are the patterns we walked away with, and the bets we're making in 2026 because of them.
Agents are eating workflows, not jobs
Every credible production AI story we heard was about agents absorbing workflows, supplier follow-ups, KYC reviews, contract redlining, not replacing a job description. The teams winning are the ones with a precise definition of the workflow, not the ones with the boldest vision of the role.
Evaluation is the moat, again
Two years in, the moat in AI products isn't the model. It's the eval suite. The teams shipping confidently have a golden dataset, a scoring rubric, and a discipline of running evals on every prompt change. The teams not shipping have demos and a backlog of complaints from sales.
If a team can't show me their eval set, I assume they don't have one. They almost never do.
Cost discipline is back in style
Token costs got a lot of conference time. The credible operators talked about retrieval caching, smaller models for routine traffic, and per-tenant budget guardrails. The talkers talked about scaling laws. The doers were doing finance.
Three bets we're making in 2026
- Agentic workflows for back-office operations, manufacturing, logistics, finance, will be the highest-volume AI deployment of the year
- Eval tooling will become the most important software category nobody is talking about
- The market will reward AI partners who can ship end-to-end, strategy, design, engineering, ops, far more than partners who specialise in only one of those
What we're not betting on
We're not betting on bespoke fine-tuned models for problems frontier models will solve in 12 months, on UI-less agent chat for users who actually wanted a button, or on conference-stage architectures that nobody has run in production. The discipline is not new, it's just being tested by a new generation of tools.