If it's a shitty process, it's still going to be a shitty process, and AI just gives you a faster version of it with fewer people around to notice.
Most AI presentations skip that, and it decides whether you ever see the returns they promise. The gains mostly start with the boring stuff (organizing, documenting, and deciding what good actually looks like), and you don't get any of that by buying a license.
And I'm not pointing fingers here, because it's happened to us many times.
Start with the problem, not the agent
This is one of my favorite topics. Everybody is focused on agentic stuff right now, and is it super cool? Yes. But it's not magical and it doesn't fit every problem, and people skip over the most important part, which is what problem you're trying to solve. Most people start with the solution (agents, cool) and work backwards.
So what's the problem, and is this actually going to make it better? More importantly, what business process is it going to use, and is the data any good? If the data isn't good, you have to fix the data first. If the process isn't good, you have to fix the process first.
The four-rung test
Graeme Crouch's four-rung ladder (Graeme runs delivery at TruNova) is the test I keep coming back to here. You take one process and see where it falls off:
- Knowable: does anyone understand it end to end, or does it live in one person's head?
- Documented: is it written down anywhere outside that head?
- Repeatable: if three different people ran it, would you get the same answer?
- Consistently followed: how close is what the SOP says to what people actually do?
If it lives in one person's head, you have a business continuity problem before you ever have an AI problem.
The rung that catches everybody
The last rung catches nearly everyone, and it's sneaky because the SOP looks perfectly fine on paper. Someone documents a process but forgets to put it in a central place everyone knows where to find it. Or it was rolled out a long time ago and we didn't put in checks and balances to make sure we kept doing it the same way over time. Or the process got updated, but no one trained the team on the new way.
Our sales process at TruNova is a good example. We have processes for how our pricing and billing work, but what was documented (what we originally decided, how it was going to work) was different from how the team was doing it. That's not the team's fault, they just did the best they could based on what the others on the team were doing, and the original documented process was never fully rolled out.
Then a new hire learns it the "right way", but the rest of the team isn't doing it that way, which causes issues. An AI agent is basically that new hire, except it never goes to lunch with the team and learns how things really work 🙃
Build an agent against the documented process and the tool will behave exactly as designed and still produce wrong answers, because the documented process isn't what actually happens. In a lot of companies there's a documented process, and then people do other things, or different people do different things. Is that good or bad? It depends, but it's definitely unpredictable.
A newer model won't fix that, but an afternoon comparing your documents to what people actually do will (bring snacks, it's usually a longer afternoon than you planned), and it pays off before any AI is involved at all.
It's not a technical task
Doing this stuff is a lot harder than everybody thinks. It takes a lot more background work, discovery, understanding and design than people expect. The technical part is in there, but it's towards the end.
And then somebody has to own it. Who's going to maintain it and who's going to monitor it? How do you know it keeps working, and what happens when the process changes? Who makes those changes? And it's not just general maintenance. When new models come out, they work differently, so how do you test that the process still works, and how do you validate and govern it?
Then there's cost. How do you even know the ballpark cost of these things, and is this process worth that cost? How do you predict it and monitor it, and what happens if it gets out of control and costs a bunch of money? Does it just stop? Do you put a budget on it, and what happens in the business if that process stops?
These are all things you have to think about and plan ahead of time so you don't get surprised, or more importantly, so your client or your business doesn't get surprised. It's really tricky, and it's the part that almost never makes it into the demo.
Readiness happens one process at a time
The wrong reaction to all of this is "we're not ready, come back next year." Your CRM can be a complete disaster while your contract review process is perfectly ready, and those two facts have nothing to do with each other.
So the useful question is which of your processes is ready first, and then whether AI actually helps there. Would an agent work, or is it more about using AI in certain pieces of the process, or maybe a human just using AI to do their part better? Sometimes that's the better answer for the business. Technical people (me included) tend to think "ah, but it's better to use X," and maybe it is, but what's better for the business? What's the ROI, and what's actually going to make people's lives better? And what will the humans like, because the humans at the business actually matter here. A ten-person company might have exactly one good candidate and a big one might have hundreds, so size really just decides how many options you get to pick from.
And when you find the one that's ready, it's usually not the flashiest or most mission-critical process you have. It's the one that's known, written down, repeatable and actually followed, which tends to be the kind of process nobody brags about at a conference (and if they did, I'd probably sneak out for coffee, or a good bourbon, depending on how bad it is).
Pick one process this month, walk it up the ladder, and see where it falls off.



