Sometimes it comes right after a video or an info session: "I hear AI can do things on its own now. Let's get one too." Sometimes it comes six months later, in a meeting room, in a much less excited tone: "We spent the money, nobody's any less busy. What went wrong?"
Between those two sentences sits a question nobody ever asked out loud. And you don't need to understand the technology to answer it.
01The one-question test: are the steps fixed?
Before you compare tools, read quotes, or sit through vendor pitches, put the job you want to hand off on the table and ask one thing: is it done the same way every time?
If the answer is "yes," what you want isn't an agent.
If the answer is "it depends," that's when an agent comes in.
It sounds too obvious to say, but most rollouts fail precisely because nobody asked it. A company spends six figures on something that can make judgment calls, then puts it to work on a job that needs no judgment at all. That's hiring a head chef and having them do nothing but switch on the rice cooker every day.
02The fixed-step half: what you actually need is a pipe
Let's get specific. Your company probably does these every day.
A customer fills in the form on your website. Someone copies the details into the client list, sends a confirmation email, pings sales in the group chat, and sends a reminder if there's no reply in three days.
An invoice comes in. Someone checks the amount, files it in this month's folder, and updates the master spreadsheet while they're at it.
At the start of each month, someone downloads a report from each of three back-ends, pastes them into one spreadsheet, works out a few numbers, and emails them to the manager.
These all have something in common: the steps are the same every time, the order is the same every time, situation A gets action A and situation B gets action B. You can write it down, and it doesn't change. This kind of thing has a very unromantic name: fixed automation. There are several tools on the market for it, and they all do the same thing: they connect the dozen-odd clicks you used to make by hand into a pipe. Data goes in one end and comes out the other.
It isn't smart, and it doesn't need to be. Its one talent is "the same every time," and that happens to be the most important virtue for this kind of work.
It has some very practical advantages too. It's cheap, and you can understand it; when it breaks, you know which section broke. A company with no engineering team can maintain it themselves once it's set up.
For a business owner, which tool you pick isn't that important. What matters is whether the job needs judgment. That's the dividing line that decides how much you spend.
03The half that needs judgment: this is where agents come in
So what does "it depends" look like?
A complaint email comes in. It could be a return request, a question about delivery, or someone who's angry but really just wants to hear an apology. It could be all three. You have to read it before you know which way to go. Along the way you check the order, check the shipping record, then decide what to reply and whether to escalate. It's the same task, "handle an email," but every email takes a different path.
A task like this can't be drawn as a flowchart, because there are too many branches to draw. You can't list every case in advance. All you can do is give it a goal and let it decide the next step as it goes.
That's the real definition of an AI agent: not "more powerful automation," but automation that plans its own steps on the spot.
The line extends in both directions. If all you want is "something that can answer my questions" — look up a regulation, rework a paragraph — you don't even need a workflow. Just open a chat tool and ask. One step further up, multiple agents collaborating and running across applications for long stretches, is the heaviest end, and the easiest to lose control of.
Laid out as four steps, it looks roughly like this:
- Fixed questions that need instant answers → a chat tool is enough
- Repetitive tasks with steps you can define in advance → fixed automation
- Steps that aren't fixed and need judgment → an AI agent
- Multiple agents collaborating, running across apps for long stretches → a heavier agent platform
Most small and mid-sized businesses spend their days on the first and second steps. The loudest voices in the market are always talking about the third and fourth.
04Even if it needs judgment, there's a second question: does it hurt enough?
Once you've established that a job needs judgment, don't rush to sign anything. Technically possible and worth doing now are two different things.
There's a more practical test here. Plenty of companies with resources hold off, not because they don't understand, but because the need doesn't hurt enough yet — the gap an agent could fill is too small. Instead of asking "should we do it," ask "which constraint am I under right now that's making the old solution stop working." There are roughly four such constraints. If any one of them is true, "no rush" flips to "we need to take this seriously."
A staffing gap. Key roles go unfilled, and senior people who can run things on their own are harder and harder to find. An agent here isn't a replacement for people. It lets one senior person work with a few agents and get through what used to take three. Like a head chef who used to need three apprentices and can now direct the whole kitchen alone.
Fuzzy boundaries on what you're asked to do. The jobs coming in are all over the place, and you can't keep every skill on payroll. The core team holds onto the few things it does best, and agents that can be deployed quickly fill in the edge skills. Like core chefs backed by an on-call caterer.
Busy months and quiet months. In peak season you never have enough people; in the off-season you're carrying heavy fixed costs. Core staff handle judgment and client relationships, while agents scale up execution at the peak and scale back down in the trough. Like expanding the fleet when the wave comes in and trimming the crew when it goes out.
A new line of business that needs fast validation. To find out whether a new direction is worth investing in, building a whole team to test it is too slow and too expensive. Let agents form a virtual squad to run the proof of concept first. Once you're sure it's worth it, build the real human structure.
Underneath all four constraints, the question that really belongs on the meeting-room whiteboard is the same one: which constraint is your organization under right now that's making your old solution stop working?
If you can answer it, it's worth talking about agents. If you can't, get the pipe from step two connected first. The return is much higher.
05Pick the wrong side and both sides hurt
Handing a job that needs no judgment to something that makes judgments costs more than just money.
It makes mistakes you didn't see coming. When a fixed workflow goes wrong, you know which section it went wrong in, and you fix that section. When something that plans its own steps goes wrong, you first have to figure out what it was thinking at the time. The untamable lobster from our last piece is the extreme version of this.
It also fools people into thinking they're doing AI transformation, when they've really just wrapped a simple job in expensive packaging. It's the same illness as in the piece before that: buying something that doesn't belong to the step you're on.
The reverse is just as bad. Force a job that needs judgment into a fixed workflow and you get an endless pile of exception rules — one more for every new situation — until six months later nobody can read the flowchart, including the person who wrote it.
There's only one test. The hard part is applying it honestly to every single workflow.
06The hard part isn't the test. It's looking honestly at your own workflows
It sounds simple. The difficulty is that you know your own workflows so well that you can't see which steps are actually fixed. Most companies, the first time they lay their workflows out, find far fewer steps that genuinely need judgment than they expected. More often, the step everyone keeps complaining "takes forever" doesn't need AI at all. Connect a pipe and it's solved.
That's exactly what our AI workflow diagnostic does: we lay your workflows out with you and mark them one by one. This part is fixed, so automation is enough; this part needs judgment, so it's worth an agent; this part can't leave human hands, so someone has to press the final button. Once it's marked up, the money worth spending separates itself from the money that isn't.
A question that will still work in ten years
Back to the line at the very beginning: "Let's get one too."
One of what? That depends on whether the job you want to hand off has fixed steps. Ask that one question and eighty percent of the anxiety about choosing tools disappears; hand the remaining twenty percent to "does the gap hurt enough" to filter.
Tools will keep changing. Automation platforms may rename themselves next year, agent platforms reshuffle every three months, and today's flagship is standard equipment by year-end. But "are the steps of this job fixed?" will be just as useful a question in ten years. It isn't asking about tools. It's asking about the shape of the work itself.
Let AI move fast. Let people hold the anchor. Deciding where to speed up and where to anchor has always been people's work, from start to finish.
Rooted in craft. Built for the new wild.