Building your team's capability

The most durable thing you can build is not a deployment. It is thirty people who use these tools well and know where they fail — because the tools will change, and that capability transfers to whatever comes next.

Three levels, and who needs each

Everyone: competent use. Briefing a tool properly, iterating rather than accepting the first output, recognising the failure shapes, knowing the traffic light. This is a day of training and a month of practice, and it produces most of the value your team will ever get. It is also the level almost nobody invests in, because it is not a project.

A few: building. Two or three people who can configure assistants, build automations, and assemble tools without engineering. This is where the long tail of small improvements comes from — the fifteen small annoyances no vendor will ever address because each is worth four hours a month to one team.

One: going deeper. Someone who understands the mechanism well enough to evaluate claims, spot a bad architecture in a vendor pitch, and tell you when a proposal is unrealistic. Not necessarily technical by background. This person saves you more money in avoided mistakes than they cost.

What to actually give them

For everyone, AI for Everyday Work covers competent use in a work context — briefing, iterating, verification, and what not to paste. It is the right thing to put in front of thirty people.

For the two or three builders, Build AI Tools Without Code is the practical course: configured assistants, automations with the guards that stop them misfiring, document chatbots, and the honest signals that a problem has outgrown no-code.

For the one going deeper, How AI Actually Works first, then the technical shelf as their role demands.

For yourself, the audit and the pilot design in this course are the parts to actually use. You do not need more than that.

The internal prompt library

The single highest-return, lowest-effort thing on this list.

When someone works out a genuinely good way to get a task done — the briefing that produces a usable acknowledgement letter first time — that is institutional knowledge, and by default it stays on one laptop. A shared document with the prompt, what it is for, and an example of good output turns one person's fifteen minutes of figuring-out into thirty people's.

Keep it plain. A document with headings, not a system. The ones that get built as systems get maintained for a month and abandoned.

Kavita's has nineteen entries after five months. The most-used is a six-line briefing for acknowledgement letters, written by someone on the complaints desk, which is used forty times a week by people who have never met her.

Where AI capability should sit

Not in a central AI team, for an organisation your size. The knowledge of which tasks matter is in the operating team, and central teams reliably build things that are technically sound and operationally beside the point.

What central functions should own: the approved tool list, the data policy, the vendor and legal review, and the sharing mechanism between teams. What the operating team should own: which tasks, which process changes, and whether it worked.

The thing worth protecting

Guard against the tool becoming the way people avoid thinking about the work.

The most valuable outcome of Kavita's five months was not the summarisation deployment. It was the work audit — the first time in years that anyone had looked at what her team actually does, task by task, with the people who do it. That produced a form redesign that saved more hours than the AI did, and it produced a shared vocabulary for talking about the work.

The audit is the capability. The tools are what you do with it. Teams that get this right keep getting value as the technology changes; teams that bought a product and stopped thinking get one improvement and then a renewal invoice.

Do this today: start the prompt library. One document, one entry — whatever the most useful thing anyone on your team has worked out. It takes five minutes and it is the cheapest institutional memory you will ever build.