DO you ever get that slightly anxious feeling when you jump in a taxi? The meter starts ticking and you don’t know whether it’s going to be a five-minute hop or a 20-minute crawl through traffic? You watch the numbers climb, wondering whether you are nearly there.
It’s a familiar kind of uncertainty. You’re committed to the ride, you just don’t know yet what it will cost you.
I demonstrated an AI agent last month that read and analysed 100 customer emails, drafted personalised replies, built a sentiment-tracking dashboard for my team and wrote a board report on the trends, work that would have taken me the best part of a day. The agent did it in just over 12 minutes, checking in when it wasn’t sure and asking before it acted.
This wasn’t a bespoke build. It was Copilot Cowork, Microsoft’s newest personal productivity agent.
The shift is real.
Instead of spending hours doing the work, I reviewed work that had already been done. A few years ago, delegating a complex multi-step task like that to AI would have been difficult to imagine. Today the output is good, and it arrives in minutes.
Last week, I ran the demo again. This time I wasn’t watching the output. I was watching the cost.
For two years, most Jersey businesses have experienced AI as a subscription: a flat monthly fee per user, used as much or as little as people liked. Cowork usage changes that. On top of your Copilot subscription, you pay a metered rate for what the agent produces. The more work you ask it to do, the more it costs.
That demo cost just under £10. To complete the task, the AI model used around 1,170 Copilot Credits, Microsoft’s new pay-as-you-go pricing for agentic work.
Every Cowork task is now a small commercial decision: is this work worth £10?
And the meter doesn’t care whether the output is any good. A vague prompt, a half-formed brief, a task the agent wasn’t really suited to. It doesn’t matter. You pay the same £10 either way. The skill isn’t just knowing when to use Cowork, it’s knowing how to ask it for the right thing, the first time.
For half a day of skilled work delivered in around 12 minutes, £10 feels like a bargain. Most professionals would pay it without thinking twice. But the economics of one task are easy. What happens when those decisions are made several times a day, across a whole team?
Twenty people running two or three tasks each can quietly generate a significant bill.
That is where this stops being a technology story and starts being a management one.
Finance will want visibility of spend. IT will need the ability to monitor usage and enforce limits. Operational teams will need to explain which work is being handed to Cowork, why it is worth doing and what value it delivers.
And the people running the tasks will need to know how to get maximum value for minimum cost because blind experimentation, at metered rates, will be expensive.
The era of subsidised AI is ending. The organisations that benefit most won’t be the ones using it most heavily. They will be the ones that understand where it creates value, where it doesn’t and what it should cost when it does.
We’ve spent two years asking whether AI can do the work. The next two will be about making sure the ride is worth the fare, getting value without losing control of the bill.
We’re helping organisations, including our own, to understand when to accelerate, when to slow down and how to keep the meter under control.
