We’ve had a run of success giving agents well-defined work: drafting, checking, summarising, monitoring. The pattern behind those wins is no mystery. Where we can specify inputs, outputs and what good looks like, agents perform.
The harder lesson came when we tried to automate work our best people do, and discovered nobody could say what that work actually is.
Watch a truly experienced operator, engineer or analyst and much of their value sits in judgement they can’t articulate. The number that’s technically in range but feels wrong for this time of year. The customer complaint phrased mildly that is, to a practised ear, a serious warning. Knowing which upstream event makes today’s reading meaningless, which rule to bend and which never. Twenty years of pattern exposure, compressed into instinct. Ask them to write it down and you get the official process, which is real but is not the job. The job is the thousand corrections applied to the official process, mostly without conscious thought.
This is tacit knowledge, and it’s the hardest thing to hand an agent, precisely because handing anything to an agent starts with articulation.
Two conclusions have reshaped how I think about our automation roadmap. First, sequencing: the work most worth automating is often the work least able to be specified, so the path runs through patient knowledge-harvesting, sitting with experts, mining their exceptions and war stories, turning instinct into rules where possible and into documented judgement calls where not. Slow, human work that no platform purchase can skip.
Second, urgency, and this is the one that keeps me up. That tacit layer currently walks out the door with every retirement and resignation, and in utilities the demographic maths on this is not kind. The AI program has made visible a knowledge-loss problem we should have been panicked about anyway.
So ask this in your organisation: whose judgement, if it left next month, could nobody reconstruct? That work is simultaneously your automation frontier and your biggest single point of failure. Both facts deserve a plan.