Many organisations want to "do something with AI." They see the possibilities, feel the pressure to move with the times, but wrestle with the same question: where do you start?
Often that question gets immediately translated into technology. Into models, data, or tooling. While the most important step usually hasn't been taken yet. Not looking at what AI can do, but at how the work is actually done today.
Technology rarely fails
When AI applications don't deliver what was expected, it's rarely down to the technology itself. In most cases, the model functions exactly as designed. The problem lies elsewhere: in an unclear process, an implicit expectation, or a task that was never made explicit.
AI amplifies what's already there. If the process isn't sharp, that lack of clarity gets magnified.
Two ways to make AI better
There are broadly two ways to improve AI systems.
The first is well-known: have the model learn from more or better data. That takes time, infrastructure, and scale.
The second gets far less attention: making explicit what the model should do, who the output is for, and how that output will be used within the existing process.
It's precisely this second step that doesn't require new technology—it requires better understanding.
A 'good' AI output doesn't exist
During a demo, you see an automatically generated summary. The first impression is often: "That's a good summary." But that impression presupposes something that doesn't exist.
A good summary isn't universal. What counts as "good" depends on context:
- who reads the summary?
- how long can the summary be?
- which information is crucial?
- should it include follow-up actions?
Without answers to those questions, a model can function perfectly from a technical standpoint and still produce unusable output.
The problem isn't in the AI. It's in the absence of clear process agreements.
Understanding what really happens
Good software solutions therefore don't start with the model, but with peeling back the task.
That means returning to questions like:
- what's the goal of this task?
- who is the output for?
- which information is actually relevant?
- what happens next with this information?
These aren't technical questions, they're process questions. And that's exactly where the key to usable technology lies. Those who start at the work floor can deploy AI as support for real work. For the occupational health context, we've worked that principle out on the definitional page AI triage for occupational health — what it is, how it works, and how it sits alongside your existing case management system.
First the process. Then AI.
Stefan Samba — CTO, Triagen
