Nine lessons from scaling AI.

There are no shortcuts. I learnt that the hard way — across hype cycles, early reads on what AI would mean for my teams, tool churn, and the slow work of getting AI to earn its place in production.

Two years in, here is what I wish someone had told me at the start. Nine lessons, in the order I learnt them. None of them clever. All of them earned.

01The hype distorts both ways.

Every leader walks the same arc — the peak of inflated expectations, the trough of disillusionment, the slope of actually getting somewhere. My early read was that AI would shrink the work; in practice, demand grew and I ended up holding the team closer. The hype distorts in both directions: it makes you over-invest at the peak and under-invest at the trough. Awareness of where you are on the curve is the first useful thing.

02Fluency is earned, not delegated.

There is no shortcut for roughly three months of active, hands-on prompting across multiple tools — Copilot, OpenAI, Claude, and the rest. When a colleague tells you Claude is remarkable, the tool is only part of the story. What they have actually built is three months of accumulated prompting skill. Hand the same tool to someone on day one and they will not know what to do with it. Familiarity compounds. You have to earn it.

03Match the tool to the work.

Many requests in a normal business day are deterministic — a fixed sequence of steps that always produces the same answer. For these, AI is the wrong tool. An Excel macro, a conditional rule, or a simple automation does the job faster, cheaper, and more reliably. AI earns its place where the work needs judgement, synthesis, or generation — places where the answer cannot be fully written down in advance. Before you put AI on any workflow, ask the deciding question: is this task deterministic or not?

04Treat the toolchain as unstable.

Every AI tool is moving faster than any organisation can standardise around. Claude may lead this quarter; Gemini, OpenAI, or Copilot may lead the next. Locking into a single vendor is a strategic risk. Invest in your team's underlying fluency, not in proficiency with one tool. Reassess your stack at least every quarter. Make experimentation a normal part of the week, not a side project. Skills transfer. Tool lock-in does not.

05There are no shortcuts between the levels.

Think of AI use in four levels. Level 1 is 80% human, 20% AI — you write the document; AI refines and checks it. Level 2 is 50/50 — you start; AI extends. Level 3 is 20% human, 80% AI — you set the problem; AI composes the answer; you approve. Level 4 is autonomous — agents identify problems and resolve them end-to-end, with humans involved by exception. Leaders who try to skip levels consistently underdeliver. Each level has to be lived through.

There are no shortcuts between the levels.

06Look for your business technologists.

Gartner calls them business technologists. I just see them in my teams — people with no formal engineering background who are quietly building working solutions with AI, the same way an earlier generation built whole businesses on top of Excel. Think of AI as a platform that is now complete enough to run real parts of the company in the right hands. Find these people. Give them runway. The barrier to entry is curiosity and prompting skill, not a CS degree.

07Beware AI washing.

AI washing is to AI what greenwashing is to sustainability — a surface commitment that quietly damages the organisation underneath. It usually starts with applying AI everywhere at once, and then making strong assumptions about what the work no longer needs. Six months later, capability has thinned out, the AI was never as ready as the demo suggested, and the rebuild begins. The cure is patience: tell the difference between what AI can do in a demo and what it reliably does in production. Validate before you commit. Build in review cycles.

08Hire your digital coworker, properly.

The most practical question I ask myself now: when are you hiring your digital coworker? This is not a figure of speech. It needs the same care as a human hire. Write the job description — what tasks, what outputs, what quality bar. Onboard properly — context, examples, tone, constraints. Review performance — refine the prompts, update the instructions, raise the bar. Pair it with a human lead, the same way you would a new joiner. The only real difference: this one is available twenty-four hours a day.

09Start with your highest-value work.

The instinct is to pilot AI on something safe and low-stakes. The outcome is usually a polite, inconclusive result and no real mandate to do more. Instead, point AI at the highest-value problem your team owns. The stakes create accountability, the signal is strongest, and the win — if it lands — earns you the licence to scale. It is the same advice Gartner gives. Our own version, learnt the harder way, says the same thing.

Lead from experience, not from policy.

Underneath these nine, there is one ask. AI is not really a tool deployment; it is a mindset shift and a process change. There is no skipping the levels for you either. Spend the three to six months yourself before you write the policy. Lead by personal example. Your team will believe your story before they believe your strategy.

I will keep adding to the field notes as the picture changes. This is what I know so far.