AI models and tooling get more capable every quarter. Nearly every organisation uses them, and most people say AI makes them personally more productive. Yet, according to industry surveys, the share of companies that can point to a measurable improvement in their results has barely moved.
I believe the reason is not capability or adoption, but that most companies do not know what to do with AI on an operational level. So it stays where it was easiest to deploy first: individual tasks. An old idea from manufacturing explains what happens next: merely speeding up a step does not raise output, but merely piles up work further down the line.
AI has made individuals faster at writing, coding and summarising. That gain is significant but has mostly plateaued. The constraint for better output has now moved to shared context, coordination and decision-making at an organisational level. This is where AI is barely (or badly!) deployed. The usual steps of mandating more usage and adopting more tools only provide marginal, if any, further improvements. Companies need to go through the same step change at the organisational level that individuals have already been through.
I run two products at work, and I will share what I learned trying to deploy AI into the real bottleneck on my own team: one repo of shared context and organisational memory that any role could point any agent at. What I found echoes the wider industry pattern, and I will bring the data to show it. While I will cover some tooling, this talk is mostly about a few principles I collected over the past year. You will leave with a way to find your own team’s bottleneck and a short list of decisions that move AI onto it.
// on_stage
