Nov 2024 STRATEGY

Are your workflows ready for AI?

A few years ago, if someone told you they were automating their business, it probably meant they had spent a fair bit understanding it first.

They had watched people do the work, argued over edge cases, and painstakingly written procedures. In doing so, they figured out which decisions deserved judgment and which didn't. Automation was almost an afterthought - a natural progression, after the hard part had already happened.

The go-to instinct for AI adoption today seems to be hiring AI PMs, AI marketers and AI engineers with an aspiration that they’ll be able to disproportionately affect outcomes. But most of these teams rarely have a codified sense of what’s actually under hood - the approval chains, lack of evaluation rubric and subjective conditions that cause the same decision to vary under different circumstances.

I've spent a fair amount of time automating workflows, both professionally and in my personal life. One thing I've started finding funny is how quickly teams arrive at "it depends." Sometimes it's only the second or third "why?" before the confidence starts fading.

That's usually a sign the hard part hasn't happened yet. The workflow was really just a collection of good people making good decisions.

Paul Graham famously wrote about doing things that don’t scale, mostly attributed to finding product-market fit. I think there’s another lesson hiding in there. You roll up your sleeves because that’s how you discover the work and the messy parts, where the process reveals itself.

My working theory is that the people making money, going viral, or completely rethinking their workflows with AI are, more often than not, the same people who would've done it anyway. AI has amplified capability more than it has created new ones.

That's why I ask myself whether someone could have solved the same problem without AI. If the answer is yes, and they also know where AI meaningfully changes the equation, I'm interested. When the conversation starts and ends with MCPs, RAGs, and agents, I usually get a sense of what's going on.

Every week there's another model capable of doing something remarkable, and every week teams rush to plug it into workflows that are still held together by tribal knowledge and good intentions. When outcomes are subpar, the instinct is to blame model quality.

I wonder if we're blaming the wrong thing.