The decisions we give away

Early on, I tried to build an insight agent. I gave it the problem I wanted to solve, connected it to data, asked it to find insights and waited for brilliance.

What I got back were gems such as “With almost 90% of UK adults now shopping online, brands must rethink their digital-first strategy to meet consumers where they are” or “Authenticity has become the defining currency of brand trust for Gen Z.” Statements delivered with polished confidence, the kind that might trick you into thinking they weren’t utterly useless.

It took me a minute to realise the problem wasn’t the agent – it was me.

I’d asked for insights without defining what an insight actually is. I hadn’t explained the difference between an observation and a truth, a trend and an actionable finding. I’d outsourced the task – and in doing so, I’d outsourced the judgement. The agent defaulted to what the internet rewards as an insight: a data point wrapped in narrative.

That experience changed how I think about agents entirely. It taught me the starting question isn’t “What task do I want my agent to do?” – it’s “Which decisions am I giving away?”

The hidden decisions

The word “agent” comes from the Latin agere, meaning to act. But agency isn’t just about acting – it’s about deciding and then acting.

When we give AI agency, we aren’t just handing over tasks – we’re handing over decisions. Not just the big high-stakes calls, but the small buried ones: interpreting a word, applying ambiguous instructions, determining what ‘good’ looks like without guidance.

The risk is that AI is extraordinarily good at sounding intelligent, so those hidden decisions rarely get questioned.

My insight agent wasn’t broken – it was making a reasonable call about what an insight is, based on its training data. The problem is that the internet is wrong about many things worth being right about, and I hadn’t told it otherwise.

Strategy is a chain of decisions

Once I started seeing agents as decision-makers, I couldn’t unsee it.

Take a launch campaign for a brand entering a new market. The team maps it as steps: target audience, behaviour change, insight, creative idea, execution, media allocation.

But look deeper – it’s actually dozens of decisions, each shaping the next.

Some have known answers. Some can be predicted through data and expertise. And some have no right answer – only better or worse options, chosen by a human who must live with the consequences.

That difference matters. Not all decisions should be delegated equally.

Where the answer is knowable, agents can be tireless partners.
Where expertise is required, agents can help only when that expertise has been genuinely baked in.
Where taste, conviction and context are needed – that’s the point of the human.

Agency without accountability

Decision-making rights matter because every decision has consequences. AI agents may have agency, but they lack accountability.

They don’t care if strategy is brilliant or mediocre. They don’t feel the weight of a recommendation that shifts a client’s direction.

This is why I say “humans at the helm” – not “humans in the loop”.
Being in the loop is passive: approving outputs, checking homework, taking a backseat.

Being at the helm means:

• Setting intention – problem, purpose and success criteria
• Making the calls where no right answer exists
• Owning the outcome, regardless of the automation involved

The best strategy teams I’ve worked with over twenty years in strategy have this clarity – not as a slogan on a wall, but in how they operate.

Human–AI teams need even more discipline – because when AI gets it wrong, it gets it wrong confidently and at scale.

Knowing where to stand

We often say taste and judgement are irreplaceable human contributions – and that’s true.

But the harder skill is knowing when and where to apply them. You can’t apply judgement everywhere – there are too many decisions. Some should be left to agents. The real skill is mapping the chain of decisions and identifying which are yours, which are theirs and which are shared.

What I got wrong with my insight agent wasn’t poor judgement on what makes a good insight – it was failing to apply that judgement early, in the design, in the criteria, in the definition of the ask.

The decision I gave away wasn’t “find me an insight”.
It was “decide what an insight is”.
And that one was mine to make.