AI Agents with water-treatment engineer Mr Lam KC: handling site problems faster and more neatly

This afternoon I sat down with Mr Lam KC to talk about AI Agents. He is a water-treatment engineer whose working life revolves around treatment plants, pipe networks and process performance. The question was not whether AI will replace engineers, but a more practical one: can a tool take over part of the repetitive work that still has to be done right every day?

One: what actually eats the hours is rarely the calculation

With an engineering problem, the time goes into searching: for the code, the equipment manual, the work order from three years ago, the name of whoever changed that parameter. Once found, it has to become a report, minutes, a handover note — and then someone has to chase who will finish it, and when.

Water treatment is especially like this. Water-quality data arrives daily, and every process adjustment has a reason — but those reasons mostly live in people’s memory and in scattered spreadsheets. When the person leaves, the experience scatters.

Two: three things an AI Agent can take

• Find: put codes, manuals and past work orders somewhere you can simply ask — “when was this pump last overhauled, and what was replaced?” — and get an answer in seconds.
• Write: daily and weekly reports, minutes, handover notes — drafted by the tool, then corrected in two sentences by the engineer before sending.
• Track: to-dos, inspections, anomaly alerts — who owns it, how far along, who it is stuck with — visible at a glance.

Three: what must not be handed over

Judgement and accountability. Whether an abnormal condition justifies a shutdown, whether this fluctuation is a process problem, whose name goes on the report — the engineer decides. A tool can offer options and evidence; it cannot carry the responsibility.

And relationships. Working with contractors, owners and operators runs on trust built over years, not on an automatically generated email.

Four: how I would start

Do not start with a grand system. Pick one task that repeats weekly and is most error-prone — the weekly report, or the equipment register — let the tool draft, let a person review, and keep the record. Run it for a month, then extend.

The key is the trail: who changed it, when, and on what basis. Engineering runs on traceability, and that must not be lost when AI steps in.

Five: one thing I took away

Water treatment is an invisible system: turn on the tap and water appears; few think of the processes, the pipework and the people on night shift behind it. The value of an AI Agent here is not letting a machine think for people, but turning experience scattered across individual heads into something the whole team can look up.

Tools carry the repetition; engineers keep the judgement.