Almost every conversation I have with a business leader at the moment ends up in the same place. They’ve been to enough AI seminars and read enough articles to know what an AI agent is. Where they get stuck is knowing what to get their first agent to do.
And that’s a fair thing to get stuck on. The first job you give an agent is more or less the whole experiment. If it goes badly, nobody in the building wants a second one, and the whole idea quietly dies without anyone having to say so out loud.
I’ve got a team of AI agents running this website, so I’ve had to make that call a few times now. The early ones were rougher than they needed to be. What I’ve ended up with is a handful of questions I ask before I hand a job over, and none of them are about which software you use. That part matters far less than people think.
What an AI Agent Actually Does
An AI agent is software you give a job to, rather than a question. That’s really the whole difference.
You’re not opening a chat window and asking it something. You give it a role, you tell it how you want the work done, and you give it access to whatever it needs to get on with it. Then it goes away and does the job and comes back with the result. Sometimes it comes back with a question instead, which is usually a good sign.
The chat assistant you’ve already got open in a tab is sitting there waiting on you for every step of it. An agent keeps going while you’re in a meeting, and that changes what you can sensibly give it.
Why the First Job Usually Goes Wrong
Ask someone which job they’d hand over first. Most of the time they’ll name whichever one everybody in the office hates.
Makes complete sense on paper. The trouble is that the hated jobs are usually hated because nobody has ever been able to explain them properly. They’re full of exceptions and judgement calls and somebody’s twelve years of knowing which client gets handled differently and why, and none of that is written down anywhere, it just lives in their head. Hand that over first and you’ll spend longer correcting the work than you ever spent doing it.
The other answer I get is something customer facing. Sales replies, the support inbox, the enquiry form. I understand the appeal. You’d see the win straight away. You’d also be putting the first few weeks of it getting things wrong in front of the people who pay you, and you only really get to do that once.
The jobs that actually work first are boring and internal and nobody outside the business would ever notice them. Not much of a story to tell at a conference, I’ll admit. But you get to a second job with your credibility intact, and inside a business that counts for a lot more than being right.
The Test I Run Before Handing Anything Over
So here’s what I run a job through now. Four questions, and if it can’t get through all four, it isn’t the one to start with.
Can you tell in about a minute whether it was done properly? If checking the work takes as long as doing the work, you’ve moved the bottleneck rather than removed it.
Could you hand it to a new starter with a page of written instructions? If you can’t write those instructions down, an agent has nothing to follow either.
Does the work already leave a trail somewhere, in a system or a shared drive or an inbox? Agents need something to read, and a job that lives entirely inside somebody’s head has to come out of their head first.
If it goes wrong, what does that cost you? You want the answer to be an afternoon of redoing it. If the answer is a customer, or a compliance problem, or a number that ends up in front of your board, go and pick something else.
None of that is complicated, but almost nobody does it. I’ve watched good businesses go straight past all four because someone saw something on social media and felt like they were getting left behind. And I get that, nobody wants to be the last one in their industry to work this out. Just don’t let it be the thing that picks the job for you.
Process and Readiness Come First
Process and Readiness are the first two letters of P.R.I.M.E.™, and they’re the two people skip. Which is a shame, because everything after them rests on them.
Process means going and having a proper look at how the work actually runs day to day. It’s hardly ever what the process document says it is. Half the time you’ll find a step in there that shouldn’t be automated at all, it should just be deleted. Automating something you never needed to do means you now do the wrong thing faster and at a bigger scale, which is an expensive way to find out it was pointless.
Readiness is the stack underneath it. Where does the agent get its information from? Can it actually reach that system? Who has to approve the access? And is the data in there any good? That last one matters more than people expect, because the quality of your information sets the ceiling on everything an agent can do for you. If your CRM is half empty, or half of what’s in it is wrong, and you’ve got an agent making decisions out of it, that agent will be confidently wrong. These things are built to produce a result. They’ll produce one whatever you feed them, so wrong information going in means a wrong answer coming back. And you might not find that out for months, until it causes you a problem.
None of that work is glamorous and none of it demos well. It’s also the reason some of these things pay for themselves inside a quarter, and others turn into the pilot nobody wants to give an update on.
What Changes When You Get It Right
You get time back, obviously. The bigger change for me was that the work started going straight from one agent to the next, instead of coming back through me every time.
On my site team, the search agent finds a gap in what we cover and writes it up as a brief. The writer picks that brief up, drafts the article, checks it against my language rules and gets it ready to go. The developer handles anything that touches the structure of the site. All of that happens between them. I read the thread and say yes or no.
I’m still the one in charge of it, and I approve every change before it goes live. What I’m not doing any more is standing in the middle of every handover, which in most businesses is where the time goes. Or the job just never gets done. Or you end up paying somebody else to do it.
SEO is a good example. It used to be something I looked at every six months, if I remembered, and by the time I got to it whatever I found was months out of date. Now it runs every morning without me. I look at a dashboard the agents built and keep updated themselves. Plenty of businesses are paying an agency hundreds or thousands a month for that.
Building agents that are capable and useful forces you to not only understand the task but to write it down as well. You can’t brief an agent without describing how the work actually runs. So the writing down happens first, before anything else does.
I spent seven years in the Air Force before any of this and everything we did ran on written procedures. Detailed, reviewed, kept up to date, because lives depended on them being right. Most businesses know they should have something along those lines, and most of them only think about it when somebody new starts, and then it gets half written and nobody looks at it again. Agents don’t give you that option. And once the documentation exists they keep it current themselves, which is the part that I love!
Start With One
One agent focused on one task, and a task that gets through all four of those questions. Give it a couple of weeks, and expect a lot of back and forth correcting it and redirecting it. Most of what goes wrong early turns out to be your instructions rather than the agent, and sorting that out is the actual work.
When starting out it’s slower and messier than anyone tells you. Worth it though. Once you’ve got one agent doing its job properly the next one is easier, and the one after that is easier again, and somewhere in there it stops feeling like a project and starts being genuinely good fun. Problems you’d written off years ago as just how the business runs start looking like something you could actually go and fix. Nobody warns you about that bit.