People trust the agents they helped shape
What makes people trust and adopt AI agents at work? Drawing on real examples, Regina Berengolts explores the importance of early involvement, ownership and feedback.
·Regina Berengolts, Director of Data & AI, Vaul Labs · 4 min read (updated )
In the first piece, I wrote about how easy it has become to build an agent, and how much harder it is to build one people use. This time, I want to focus on what happens next: the moment an agent leaves the person who built it, and the expectation on everyone else to use it.
That moment often falls flat. The agent turns up fully formed, with an expectation that people will start using it on Monday, and they don’t. In my experience, that’s rarely because people are against AI. Usually, they’re busy and don’t have the headspace to even think about it. A head of finance at a software company put it well when we spoke recently: “I’m sure there’s a hundred things that I could do to make my life easier and I don’t do it. But I’m not one for not wanting to try. It’s just about having the time.”
Involvement is where ownership comes from
When people don’t trust an agent, they either spend their time checking everything it does or they just don’t use it at all. The CEO of one consultancy we spoke to was holding his firm’s internal agent back from the clients who had asked to use it, for that reason: “we just don’t trust the answers yet.” Once someone is reading every output line by line, the time the agent was supposed to save has gone.
The instinct is to fix that by making the agent more accurate. That’s worth doing, and an agent that gets its work wrong won’t earn trust however involved people were. Accuracy on its own isn’t enough, though. Trust also depends on understanding: whether people know what the agent does, how it’s meant to work and whether it reflects the way they work. That’s hard to get across in a launch email.
What works much better is involving people from the start and taking them on the journey. People need some level of involvement, and ideally some ownership, while these agents are being built. They know the awkward exceptions, the steps that don’t need doing and the parts of their day they’d happily hand over, and they usually have strong opinions about all of it, which is what you want.
It takes longer at the start. But by the time the agent arrives, people already know what it’s for, because it was partly their idea.
From my agent to our agent
Building an agent for yourself is easy. It works the way you work, and you already know what good looks like. The moment you extend it to others, everybody has an opinion or a particular way of working, and they should, because they’re the ones who’ll use it.
The agent has to change to fit how other people work, and that changes the process around it too: who owns it, who decides what it should do next and how feedback reaches the person responsible for improving it.
Putting agents where people already work
We put our agents into Slack, where anybody at Vaul can ask them for help, and that lets me lead the way in showing others how I use them, so they can see what’s possible. People see the agents in use, see what they can do and can try them without an extensive training session. As a side note: this process has revealed that our marketing agent is comparatively rude to me and hugely polite to Brian, despite not having been given guidance on that.
Slack also gives us immediate feedback. People ask the agents to do things that their builder wouldn’t have known to incorporate into their tasks and tools. When that happens, we look at whether the agent should be able to do it properly, and if so , we enable it quickly while the person who asked is still interested. Every request tells us how people want to work.
As Brian put it in one of our catch-ups, “I don’t think this is a one-shot-and-done.” Agents a team relies on keep changing, because the team keeps finding new work for them, and that’s a good sign.
If you’re planning a rollout, spend less time on the launch and more time with the people on the receiving end, before there’s anything to launch at all. A little time spent early on, in short cycles, earns trust and saves frustration later.
The more people use an agent, the more they’ll ask of it and the more access it’ll need. That raises the question of how much access to give it in the first place.
Next in the series: Start narrow, then let your agent earn more.