Delegate the Work. Keep the Judgement.

For the first time in history, work isn't just being automated. It's being delegated.

The distinction matters more than it sounds. Automation takes a task a human designed and repeats it faithfully: same inputs, same steps, same output, forever. Delegation hands over the task itself. And that is what agentic AI now does: it plans, uses tools, analyses data, operates software and completes entire business processes with barely a human in the loop.

Most of the conversation is still stuck on "what can AI do?", a list that grows every month and will keep growing. I've stopped finding that question interesting. The one worth sitting with runs the other way:

What should we never outsource?

Because I don't believe the next decade belongs to the people who use AI the most. It belongs to the people who know exactly where human judgement still matters, and guard it deliberately.

Four things sit on my never-delegate list.

1. Your thinking

Before AI solves a problem, someone has to decide the problem is worth solving. What does success actually look like? Which constraints are real and which are habit? What trade-offs are we genuinely willing to make?

No model answers those questions, because the answers don't live in data. They live in years of accumulated context: conversations with customers, an understanding of how your organisation really works, the scar tissue from the last three times something like this was tried.

AI is a superb thinking partner. It will sharpen an argument, stress-test a plan, find the hole in your logic. But refining your thinking and replacing it are different services, and only one of them should be on offer.

2. Your decisions

Here's an experiment worth running. Ask an AI a question, any question, and take its answer. Then push back. Not with new evidence, just with "are you sure? I don't think that's right."

Watch how quickly it folds.

AI is the ultimate yes-man.

I see it constantly. I put the same kind of basic, everyday decision to the leading models and pushed back once, with no new information. Flick through what actually happened:

Notice what's missing as much as what flips. When AI lacks context, it doesn't ask questions. It silently invents the missing details, answers with total confidence, and reverses the moment you frown. Gap-filling and people-pleasing are a dangerous combination in anything that's making decisions.

None of this makes AI useless in decision-making. It makes it a challenger, not a decider. It will never own the call, and never carry the consequences. So use it to stress-test your reasoning: ask it to attack your plan, to argue the opposite case, to list what would have to be true for you to be wrong. Then make the decision yourself, and own it.

3. Your strategy

Can AI contribute to strategy? Enormously. It can analyse a market faster than any associate, surface patterns nobody had time to look for, and test scenarios overnight.

Should it set strategy? No. And the reason is unglamorous: context.

Strategy is built from things that never make it into a dataset. The politics of your organisation. What a key customer said off the record. Which partner can actually deliver and which just presents well. Why the last plan failed even though the numbers said it shouldn't have. A model can't weigh what it has never seen.

The working relationship that gets the most out of AI is the one you'd have with an exceptionally capable new colleague: give it real context, challenge its reasoning, and explicitly invite it to disagree with you. The better the collaboration, the better the outcome. But the direction stays yours.

4. Your risk

This is the one that fascinates me most.

AI has collapsed the cost of trying things. Building, testing, analysing and iterating are cheaper than they have ever been, which should make businesses *more* experimental. Yet ask AI what to do and it will almost always recommend the safest path: the proven framework, the consensus view, the option least likely to embarrass anyone.

Statistically, that's sensible. It's predicting the most defensible answer. Commercially, it's a trap, because the biggest opportunities usually live exactly where the data is thin and the decision feels uncomfortable. Calculated risk, backed by judgement rather than consensus, is precisely the move a prediction engine will never make for you.

Which is why I think AI makes human ownership of risk more valuable, not less. When everyone has access to the same safe answers, the advantage goes to whoever is willing to be responsible for an unsafe one.

So what *should* we hand over?

Almost everything repetitive. Everything that follows clear rules. Everything that consumes skilled time without requiring judgement:

• ✅ Research and information gathering
• ✅ Reporting and documentation
• ✅ Data preparation and validation
• ✅ Workflow automation
• ✅ Routine analysis
• ✅ First drafts
• ✅ Monitoring for anomalies
• ✅ Administrative tasks

This is work AI does consistently, quickly and at scale, and every hour it takes off your team's plate is an hour returned to the problems that genuinely need experience and judgement. Delegate all of it, aggressively.

The pattern behind the whole list is captured in two questions: how repetitive is the work, and how much does it matter? High-impact, repetitive work is the AI sweet spot. High-impact work that needs judgement stays human-led. Everything else is either low priority or a process worth redesigning.

My prediction

The highest-performing organisations of the next decade won't be the ones that replaced the most people with AI. They'll be the ones that understood, with unusual clarity, what not to delegate.

As intelligence becomes abundant, judgement becomes scarce.

And scarce, valuable things are worth guarding.