Building the agent is the easy part

I’ve lost count of the posts I’ve scrolled past this year from people talking about how many agents they’ve built. Dozens, hundreds, whole teams of agents running across their organisation. I love data, I’ve built my own agents, and I’ve worked with companies rolling out every combination of skills, GPTs and agents.  So my immediate reaction is always: that’s great, but what value are they bringing? Oddly enough, the answer rarely makes an appearance in those posts.

The headline numbers are real enough. ++Microsoft’s 2026 Work Trend Index++ found that active agents in the Microsoft 365 ecosystem alone grew fifteenfold in twelve months. Creating a first version of an agent has become easy, and the tools make it easier every month. I think that’s a good thing. It does mean the number of agents created tells you little on its own.

Lots of agents, not much use

We’ve worked with a large fintech organisation that had rolled out a bottom-up AI approach, giving AI access to the vast majority of its workforce. After a few months, a member of the IT team started taking a look at how employees had been using their new AI capability. What he found was hundreds of skills and agents across the organisation, which is what you’d hope for when people are curious and excited. But when you look at the usage metrics, only a few of them were being used, and fewer still by anyone other than the person who created them. Quite often, even the person who created one hasn’t used it much. 

That doesn’t mean anyone did anything wrong. It’s what early experimentation looks like when building costs almost nothing: you automate whatever’s closest to hand, see what happens and move on. Where it gets harder is for senior leaders who want their organisations to see value from AI quickly. The difficulty is working out which agents are worth building at all, because an idea can sound great on paper and play out differently in practice.

A senior leader I’ve been working with found this out first-hand while rolling out AI across his organisation. One of the first agents he built sent him a daily summary of his activity for the day. It did what it was meant to do. Then he realised that glancing at his own diary was quicker than reading the email it sent him. In his words, “So what’s the point?”

That’s the right question, and it changed how he thought about what to build next.

What we look for is an agent that makes the work not only faster, but better. A marketing lead at one of our clients described what that looks like in practice: “It means spending more time looking at the content and asking whether it’s the right thing to put out, rather than spending that time on formatting and assembling it.”

The hard part is everything else

Creating the agent isn’t the hard part. The hard part is deciding what’s worth building in the first place, then how it’ll be used, how you roll it out, and how other people’s work changes so it has a place in their day. 

Context plays a big part too. An agent only knows what you give it, much like a new hire on their first day. The more it can draw on  your organisation’s way of working, the better its output, and (a nice bonus) the less it costs to run, because you’ve already done much of the hard thinking for it.

Inside Vaul, we begin with the job a person needs done and the difference the agent should make, work out what context it will need, and build it with the people who’ll use it. When we check whether an agent is earning its keep, we look at what’s changed. Personally, I’d much rather have a handful of agents aimed at real work than hundreds that nobody uses. 

If counting agents tells you little, the next step is deciding what to count instead, and that’s where this series will finish. For now, before you build the next one, ask who it’s for and what will change in their work.

Next in the series: People trust the agents they helped shape.