Hi Reader,

Confession time: I’m one of those w*nkers who’s built an AI agent chief of staff. A whole digital workforce, actually. Seven of them. Named after fictional detectives. And eight weeks in, I’m in two minds about whether to turn the whole thing off.

Is it giving me a productivity gain? Dunno. But there have been benefits, just not in the form I expected.

Why I built it

I’m a small business owner. I’d rather spend my time doing interesting work with fun people i.e. clients. But running a business comes with a whole load of stuff you have to be on top of, and no one starts a business because they love admin.

Also, yes: I like playing around with tech. I learn by doing, not just reading or listening. And, let’s be honest, maybe this AI is modern-day procrastination.

What did I build?

My workforce are fictional detectives. My chief of staff is Columbo, because just before we leave the room he always suggests one more thing. My memory keeper is Nancy Drew. She’s got a nose into everything, and not just the content of what I’m working on. She watches the workforce for how I do things. That memory layer is what makes the whole thing work: the workforce gets smarter, because it knows not just what we’re doing but how we’re doing it, and how I prefer things done.

What did I get in reality?

It’s been as frustrating as delegating to humans. Misunderstandings. Taking things literally. And sub-standard output. But it’s made me do a lot of thinking about thinking. Which is weird.

I never sat down and wrote down how I think so that agents could run it. Nobody does. What happened is slower and more useful: every time I trained an agent on what I really wanted, I understood a bit more about how I actually think.

Example. Fletcher, my marketing director (Jessica, of Murder, She Wrote), drafted me some content. She had a pool of my greatest hits to work from: the insights, the things I always say, the topics in my own IP. And she came back with something that sounded like everybody else. Generic AI output is the thing I find most useless about these tools, and my own workforce was producing it.

My first reaction was annoyance at the slop. My second was more uncomfortable: I hadn’t been precise enough about what I wanted. I don’t want AI drafting my content. I only want it suggesting topics out of my own canon but the writing stays mine. That assumption was in my head and I’d never said it out loud.

Delegating to an agent is like delegating to a human, with one exception. As Matthew Treagus and I put it in our ten principles for AI agents: you’re not supposed to micromanage people, but you absolutely want to micromanage your AI agents.

The bit you shouldn’t outsource

The bigger lesson is one I’m relearning rather than learning, because my master’s degree taught it to me first: inefficiency is part of the process of developing knowledge and insight.

When you assimilate information, you gather a wide corpus and then you decide what’s important. Deciding is a cognitive step. It’s where the thinking happens. You shouldn’t outsource it.

So Bosch, my researcher, follows his namesake’s rule: get off your *ss and go knock on doors. He brings back interesting articles. I read them myself. I decide what matters and what the takeaways are. We’ve outsourced the scanning. We have not outsourced the interpretation.

The hard part of having a digital workforce

Discipline. That’s the hard part, and it shows up three ways.

First, ruthlessness about what you actually need. You can play with this stuff forever. Which skills really move the needle? Which tasks are quicker, cheaper and better in a spreadsheet or an off-the-shelf CRM? Claude is as Claude does: leave it to its own devices and you end up with a web built entirely out of AI, when half of it should be boring software that does the same thing every time, with AI as the layer that orchestrates it. AI is an ingredient, not a meal.

Second, feeding it. A workforce is only as useful as what it knows, and mine needs to know about the things I do when I’m not talking to it. That means tracking what I’m doing and planning my time. The surprise benefit: a digital workforce has professionalised my back office, mostly by making me behave.

Third, and least glamorous of all: the engineering. Nobody tells you how much traditional software you build on the way to an AI workforce. Mine runs on documents that are usage logs, architecture decision records and product requirement specs in all but name. The chartered engineer in me should have seen it coming. The shiny AI sits on top of systems thinking and the same discipline that holds up any system that works.

Why should you care?

You could reasonably say: big wow, who cares?

But read across to a business and the lesson holds. How do you codify what’s tacit? The “manual of me” has been around for ages. But do you really know how you prefer to work? I didn’t. I found out by teaching Columbo what good looked like.

So no, I’m not turning it off just yet. The jury's still out on productivity. But training it has been the most useful thinking about my own thinking I’ve done for ages. Columbo earns his keep the same way the real one did by always asking just one more thing.

Every exec team is having the AI conversation right now. Try my AI Confidence Snapshot to tell you where to focus: scorecard.helen-dawson.com. Or reply to this and we’ll find 15 minutes.

What’s coming up

The Innovation Crowd podcast is back: guest interviews, plus short solo episodes. Listen on Spotify, Apple or YouTube.

Until next time,

Helen

P.S. Reply KEEP if you think the workforce stays, or KILL if you think it’s procrastination with better branding. I’ll report the score next time.