I often hear AI consultants give this advice: “Look for a process that’s painful, and figure out how to make it less painful with AI.”

I hear the same thing on prospect calls. When a company comes to us with a process in mind to automate, it’s usually the biggest, hairiest process in the business. Their plan is to point an agent at it and watch the problem go away.

Both start in the wrong place. The ugliest process is rarely the right place to start - for two reasons.

Visible doesn’t mean valuable

Pain and size are measures of effort. They tell you a process is time-consuming, inefficient, and annoying for the team. But they don’t tell you whether augmenting it with AI will move revenue, retention, margin, or whatever metrics your company needs to drive.

Let me give you an example. The marketing team at Section runs ~80 virtual events per year - each of which involves a slew of tasks like contacting guests, managing schedules, writing copy, managing Zoom, etc. 

No one would disagree that the process is involved, visible, and time-consuming. But most of that work is logistics. Automating it wouldn’t change how many attendees book a meeting or what those deals are worth. We could spend a lot of time on it and ultimately only save 20% of one person’s time.

It's never one agent

The second problem is scale. Companies picture a single agent that takes over procurement, or onboarding, or the month-end close.

A big process is a set of workflows. Each workflow usually needs its own cluster of agents, each one reading another's output and handing off to the next. On a recent client engagement, one workflow required six different agents working together - and that’s on the low end of what we often build.

When you start where the payoff feels biggest, you tend to get into the build and find out the project is far larger than you expected. No one can figure out how to eat the elephant, so momentum stalls and you give up. 

How to choose the right workflows

You're not trying to clean up something messy. You're trying to move a metric. So start there. 

I’ll use the finance team as an example.

1. What is this department responsible for, which handful of numbers tell you whether it's working, and which work actually drives those numbers?
   
   For example: Finance is responsible for making sure the company has the cash to run, and the numbers to watch are free cash flow, days payable outstanding, and operating margin. The work that drives them is mostly how cash moves in and out of the business: collecting from customers, paying suppliers through procure-to-pay, and controlling spend before it happens.

2. Which work, if you improved it, would move those numbers the most?
   
   For example: Of those, procure-to-pay moves the numbers most, because it’s the only one that touches all three: it decides how much you pay, when the cash goes out, and whether you catch overbilling before it hits the margin.
   
   
3. Within that process, which steps or sub-workflows have the biggest potential for AI augmentation? We judge this on three Vs. Value: how close the step sits to the metric. Velocity: how fast you could ship it. Viability: whether AI can do it reliably today.
   
   ‍For example: As the finance team, you’d break down all the steps that drive the procure-to-pay process and grade them across the three Vs. 

In pulling apart your “big, messy workflows,” you’ll find that not every step is equal. Some sit much closer to key metrics than others. Some are a much better fit for what AI can do today. Some you could ship this week, and some would take a quarter. The steps that score well on all three are where you start. The rest can wait.

Quick wins - with intention

If you're early in your agentic work, weight heavily toward quick wins. They show your team what building with AI involves before you commit to something hard, and they build the confidence a stalled mega-project destroys.

But a quick win doesn’t have to be disconnected from impact. Say finance wants to cut down on invoices getting stuck. You don’t need a suite of agents for that. A simple automation that checks every incoming invoice for a PO number, and sends it straight back to the supplier if one is missing, is a tiny build. And it stops a big share of problems without human intervention. 

Your biggest problem deserves AI. It just doesn't deserve to be your first agent.

See you next week, 

Michael 
Your fellow Head of AI