If you can't delegate to people, you can't delegate to AI either

Originally published on LinkedIn.

Most people using AI at work are delegating projects they couldn't delegate to a person, then getting frustrated when the computer can't read their mind.

It does exactly what you specify

Last year I started a firm building AI tools for campaigns. The hardest problem is getting people to articulate what they want a project to achieve. Try putting into words how you actually decide on a client's messaging strategy and the research projects you need to support it. It's probably incoherent. It contradicts itself. It relies on unspoken norms we expect staff to absorb osmotically over years. We get away with it because humans fill gaps with intuition.

AI doesn't do that. It has perfect integrity: it does exactly what you specified. The gap between what you wanted and what you got is the gap between what you said and what you actually value.

AI skeptics overestimate how special our corner of "knowledge work" is. AI is doing useful work in radiology, legal research, and software engineering. It's absurd to think what happens in a strip-mall campaign office is innately more complex than that. The underlying problem is that smaller industries haven't invested in product design. We're trained on old processes, arcane bibles for how to Google the right way and format a memo about our findings. The real skill is finding answers to difficult questions, and that's jagged by nature. No process manual covers it well.

Nobody wrote it down

Jaya Gupta and Ashu Garg from Foundation Capital call this the context graph problem: the gap between what professionals know and what they've written down. Your AI isn't in the conference room when you have the idea. It doesn't have access to the Signal thread, the strategy doc you read last week, or the vibes in your heart and head. That's the gap it falls into.

Ethan Mollick gets there differently: "In figuring out how to give these instructions to AI, it turns out you are basically reinventing management." What does "done" look like? Teams defer these hard definitions because urgency doesn't arrive until the polling memo is due. Humans muddle through because they can course-correct late. A language model can't. It just goes and does the thing, and without a committed definition of done, it falls straight into the gap. The user expects magic and receives nonsense.

Be smart or persistent

Most people disappointed with AI treat their first prompt like a delivery instruction: "Build me a messaging strategy that will win the campaign." They're skipping the hard part where you figure out what you're building, for whom, under what constraints, and what you're willing to get wrong. This is a learnable skill. You either need to be smart about it or persistent; best results if you're both.

Smart means answering the product questions before you touch the technology. What makes a research finding useful? When is a project done, defined as something you'd be proud to show your client — versus the deadline hit and whatever you had is what you sent over? Before I got any staffer or AI system to produce work I'd trust, I had to answer these questions. None had anything to do with AI.

Persistent means staying with it when the first output is bad. AI is good at helping you figure out what you actually want. When one produces something wrong and you have to explain what's off, you're extricating your own decision trace. Your first prompt is a hypothesis. The output shows you what you left undefined. You iterate the prompt, not the output. Most people try once, get slop, and walk away. The ones who stay find something useful because bad outputs surface requirements they didn't know they had.

We never had to write any of this down because people were there to fill our gaps. AI is forcing the definition meeting we've been skipping for decades.