In short
The signal that matters
- If your agents keep sending work back for you to check, correct and re-explain, faster and in greater volume than before, the AI is probably not your problem.
- Your workflow grew layer by layer for people who fill the gaps with judgment, and your agents run every step exactly as written, so they hit every one of those gaps.
- Take one workflow, review it end to end as if you were designing it today, and cut or merge every step that isn't strictly needed, aiming for a 50% reduction and no less than 20%.
- Every step you remove takes away up to two places where things can break, which is how one client process went from 14 steps to 5, and from 20 minutes of failing back and forth to under 5 minutes with one round of interaction.
- Before you put agents on any workflow, start with the chain instead of the tool, and give every handoff a clear input, a clear value added, and a clear output.
Most AI implementations don't fail because of the AI. They fail because of the process we hand it.
I see it all the time. People take the workflows their teams have been running for years, and they give them to agents exactly as they are. And honestly, I get why, because it feels like the safe move. The process works, so why would you touch it? But it's the most common mistake I see when founders start putting agents to work.
How workflows get heavy
Nobody I know ever sat down and designed a messy process on purpose. It grows on its own. Someone makes a mistake one week, so you add a step. A few months later something else slips, so you add a check after it, and then there's a sign-off in there because someone once forgot something. Give it a few years and you've got layer sitting on layer, and half the team can't tell you why some of those steps are even there.
We humans can live with that pretty well. We skip what doesn't matter, we fill in the gaps with our own judgment, and when the instructions run out, we just use common sense.
Agents don't share that common sense.
What happens when agents run it
An agent runs every single step the way it's written. So wherever a person would have quietly filled a gap without even thinking about it, the agent hits that gap, and wherever the information is incomplete, what comes out the other end is off.
And then it all lands back on us, only faster and in greater volume than before. The owner ends up checking, correcting and re-explaining the same things, until they're completely brain-fried. You've got the same noise you always had, just on steroids now.
To be fair to that workflow, there was nothing wrong with it when only people ran it, or when it ran on simple automation with fixed rules. It just wasn't built for agents.
Every step is a place to break
Here's something most founders never stop to count. Every handoff in a process has an "out" and an "in," and only the first step (it has an output only) and the last step (it has an input only) get away with one point. So a 10-step chain doesn't have 10 places where it can break. It has 18.
Wait, what? 18? Yes, because every step in the middle takes something in and hands something on, and each of those two moments is a chance for something to get lost or misread. You've got eight steps in the middle with two points each, then the first one and the last one with one each, and that's 18.
That's why I don't treat simplicity as a nice-to-have once agents are doing the work. Every step you take out removes up to two places where things can go wrong.
The fix
- Review your whole value chain. Look at your workflows end to end, like you were designing them from scratch today.
- Cut every step that isn't strictly needed, or merge steps together.
- Design a clear handoff between every link, so each step knows what it receives, what value it adds, and exactly what it hands to the next one.
My tip here is to aim for cutting 50% of the steps, and to not settle for less than 20%.
The fewer steps you have, the better your chance of success, with AI or without it.
A real example
I was working with a client recently, and we sat down and reviewed one process in their organization, so just one part of their value chain.
That one process had 14 steps. Once we really looked at it, it turned out most of those steps got added over time because of errors and issues they'd had in the past. All our human fears and worries had been poured into the automation, so those mistakes would never happen again. And with AI doing the work, most of those steps just weren't needed.
We took it from 14 steps down to 5.
It still gets them the same outcome, from the same inputs, with the same outputs. All we changed was how it's organized, so it fits the way agents actually work.
So the process that used to take 20 minutes of back and forth, kept failing and left the owner brain-fried, now runs in under 5 minutes and finishes successfully, with one round of interaction.
Before you automate, simplify
If you're about to put agents on a workflow, my advice is don't start with the tool. Start with the chain. Cut what you can, merge what's left, make every handoff clear, and only then let your agents run it.
It's the same thinking I shared in You Do Not Need an AI Chief of Staff, where I suggest you start with one process and one agent. And a simpler chain is easier to keep safe too, since there are fewer places for an agent to reach into, which goes right back to what I wrote in Easy AI Setup Is a Real Business Risk.
If you'd like to look at your own value chain with me and get it simple enough for your agents to run, click the button below.