Go with the flow

How do we overcome the temptation of AI autopilot?

How do we resist the pull of just clicking ‘accept’, the lure of letting it handle all the thinking for us?

My approach is: I don’t.

I accept my shortcomings, and plan around them.

Huge AI-generated planning document in need of review?

I may use AI to summarise it.

I may skim the doc, acknowledging I’m only going to have a cursory understanding, but trusting my pattern recognition to flag anything unusual.

Or I may simply trust the person who generated it, who’s going to be in charge of the project and running it day to day, to get on with it.

Do we even need the review in the first place?

---

With AI, we’re moving to higher levels of abstraction.

In my experience highest level of abstraction when it comes to software development is the human layer.

Yes, we need to construct guardrails, evals and rich context for agentic systems to operate effectively and autonomously.

But creating those conditions will require effective human communication.

• You need someone with an engineering mindset to consider the edge cases and failure conditions for guardrails.

• You need to talk to the domain experts to understand how to construct the right evals.¹

• You need a human vision, articulated clearly, for AI to enact it effectively.

In reality, human alignment has always been the most important determinant of successful software.

AI doesn’t change that.

It just forces us to abandon the idea that we ever owned the code, the architecture, the language or the syntax, the frameworks or the philosophies.

We only over owned our relationships.

To our customers, our colleagues, and to ourselves.

This endures.

1. Shout out to Hamel Husain for educating me on this one.