AI can now do an extraordinary amount of work inside a business, and do it well. By June 2026 the tools are genuinely capable — Anthropic’s Claude Opus 4.8 landed on 28 May with sharper judgment and the ability to run hundreds of tasks in parallel, and its more powerful Mythos model is expected to reach everyone in the coming weeks. The capability question is largely settled. The open question is no longer what can AI do — it’s who steers it.
That’s the part most small businesses haven’t thought about yet. And it’s the part that’s going to matter most.
AI is brilliant, but it can be blinkered
Give a capable AI a goal and it will pursue that goal efficiently. The problem is the things it wasn’t told to consider — the customer who’ll react badly, the supplier relationship that matters more than this quarter’s number, the unusual situation that doesn’t match anything in the data. AI will drive hard toward the target it was given and, in doing so, can quietly trample things a human would have flagged on instinct.
This isn’t a flaw that gets patched away. It’s structural. An AI optimises for what it can see. A person who has lived the work sees what isn’t in front of them.
I’ve watched this in my own field. An automated feed will happily push products that are technically compliant but commercially wrong. An ad system will spend a budget perfectly and still point traffic at a page that was never going to convert. The machine did exactly what it was asked. Nobody asked the right question. That gap — between doing the task and understanding the situation — is where the damage happens.
This isn’t just my hunch — it’s the emerging consensus
I’ll be honest: this started as a feeling, and I went looking to see whether the people building and studying AI actually agree. They do.
Nvidia’s Jensen Huang put it bluntly at GTC this year: AI won’t replace your job, it will replace everything in your job that doesn’t require judgment. What’s left is direction, taste, ethics, accountability — deciding what’s worth doing and asking the right question. Sam Altman has walked back his own earlier predictions of rapid white-collar wipeout, and the current consensus among the major AI leaders treats AI as a productivity layer that still depends heavily on human judgment and supervision, not a replacement for it.
The research community is saying the same thing. Across high-stakes fields, “human-in-the-loop” oversight has moved from nice-to-have to structural requirement — and crucially, the experts stress that having a human in the loop isn’t enough. McKinsey found that 51% of organisations using AI have already had at least one negative consequence from it, a third of those tied to AI getting something wrong. The reviewer has to have genuine domain expertise, or the checkpoint is worse than useless — a poorly briefed person rubber-stamping flawed AI output is a liability dressed up as governance.
There’s even regulation arriving on this. The EU AI Act’s Article 14, which requires demonstrable human oversight of high-risk AI systems, takes effect in August 2026 — and it reaches any business whose AI touches EU customers, which includes plenty of UK firms.
The role now has a name
When I started writing this I thought I was describing something I’d noticed. It turns out the role has already been formally named. In February 2026, Harvard Business Review defined the “agent manager” — the person responsible for making sure AI agents actually deliver the right business results, working safely alongside people. It’s now being described as one of the fastest-growing job titles of the year.
What an agent manager actually does is telling. It isn’t a deeply technical role — you don’t need to be a data scientist. You need to understand the workflow: where the pieces connect, where things break, what the numbers are quietly telling you, and when to step in. It’s an operations-and-judgment role with enough AI fluency to know what the machine can and can’t be trusted with.
In a large enterprise this becomes a job title. In a small business, it becomes a function someone has to own — and most small businesses don’t have anyone who fits.
Who should be steering it?
Here’s the crux. The best person to steer AI in a given business is someone who has both:
- Deep knowledge of the actual subject — the lived experience, the instinct for when something’s about to go wrong, the feel for the customer and the context that no model has.
- A real understanding of how AI works — what it’s good at, where it’s blinkered, how to brief it, and when to overrule it.
Most people have one or the other. The domain expert who’s frightened of the tools. Or the AI enthusiast who doesn’t know your industry well enough to catch the costly mistake. The value sits with the rare person who holds both at once — and that’s exactly the person businesses are going to start searching for over the coming months and years.
What this means if you run a small business
Someone who knows your kind of business and knows the machine.
You don’t need to hire a full-time AI department. You need someone — even part-time — who knows the business or trade that they work in very well and also knows how these increasingly powerful tools work, point them at the right targets, and catch the things they’d otherwise miss.
That’s the work I do. I bring senior-level expertise in e-commerce, Google Ads, Merchant Centre and AI search, combined with a working understanding of how to make these AI tools serve a real business without the disasters. The foundations-first approach I’ve always taken — audit what’s actually happening, fix what’s broken, then let the machine run on solid ground — is exactly what steering AI well requires.
The capability is here now. The steering is what’s missing. Get that part right before you scale anything, and AI becomes the best thing that ever happened to your business. Get it wrong, and it’ll burn budget and goodwill faster than any human ever could.
Who should be steering AI in your business — and how to make sure it’s pointed at the right targets — that’s a conversation worth having now, not after something’s gone wrong.
