The 10-80-10 Model, Updated for AI
Back in 2016 I taught a model called 10-80-10 to hundreds of agencies, and a lot of them used it to scale. The idea is simple. Put your energy into the first 10% and the last 10% of any job, and hand the middle 80% to someone else.
It worked when "someone else" meant a person. It works even better now that "someone else" can be an AI agent. There's just a sharper edge to it now, and it's easy to grab the wrong end.
The model, quickly
The first 10% is definition. What does great look like here? What's the real outcome I'm after, not the busywork version of it? How will I know when we've hit it? And the part that's easy to skip: what exceptions do I need to be told about?
The middle 80% is the work. Someone (or something) else does it. This is the part you're trying to get off your plate.
The last 10% is improvement. Feedback and training, so the same job gets done better next time. This is the bit that's easiest to drop, and it's the bit that compounds.
When I first explained this to agency owners, the resistance was always the same. They'd hire a VA, get the big red pen out, correct everything, and decide it would've been faster to do it themselves. Today, yes, it would. But if you train three people who each do a 60-80% version of what you'd do, and you keep improving them until they're at 90-95% of you, you've freed yourself up for the business, the family, or riding a mountain bike through a forest.
The AI trap is the red pen all over again
The trap with AI is the same red pen, just faster.
You get an agent doing real work. It drafts the email, builds the report, writes the first version of the page. And then you check it. All of it. Every time. You read every line, tweak the tone, fix the one number that's off, and ship it.
It feels responsible. It is also where the gains quietly die. I keep seeing the figure that roughly 40% of the time AI saves you gets eaten right back up in checking and correcting its output. That matches what I see. If you approve everything the AI produces, every piece of work still has to queue behind you. The agent got faster, but you're still the bottleneck it all has to pass through.
The fix lives in the first 10%, same as it always did
The original first 10% asked one quietly powerful question: what exceptions do I need to be told about? That question is the whole game with AI.
Swap the approval gate for an escalation gate.
An approval gate means nothing goes out until you've checked it. You're in the loop for everything.
An escalation gate means the agent runs on its own inside a clear boundary, and only calls you when it hits the edge of what it's allowed to do. You're in the loop for the edge cases, and nothing else.
How do you draw that boundary? Three questions:
- •How costly is the mistake if it gets this wrong?
- •How confident is the system in what it's doing?
- •Can the action be undone?
If it's cheap, confident and reversible, let it run. If it's expensive, uncertain, or a door that only swings one way, that one comes to you. That's the boundary. And it's the same "what do I need to be told about?" judgement you'd make handing work to a person, made explicit and handed to a machine.
Then the last 10% does the compounding
This is where it gets good.
Every time the agent escalates and you make the call, you feed that decision back in. The ambiguous case you just resolved becomes a rule. Next time that situation shows up, the agent handles it on its own. The boundary widens. Less reaches your desk. The loop gets tighter.
Do that a few times and something strange happens. The moment you've drawn the line clearly, this the AI handles, this it brings to me, the whole thing accelerates in a way that's hard to believe until you watch it happen. Work that used to crawl through your inbox starts moving at the speed of the agent, not the speed of you.
Where to start
Don't try to automate the whole business at once. That's the mistake. You'll build ten half-loops that all still funnel back through you, and you'll have rebuilt the bottleneck ten times over.
Start with one tightly-scoped job. Something that happens often enough to matter and is contained enough to reason about. Define what great looks like. Draw the boundary with the three questions. Then close the loop, end to end, so the agent can run it without you stuck in the middle.
Get one loop closed. Then move to the next one.
That's where 10-80-10 always pointed, even back in 2016. AI just makes it real. The faster you close your loops, the faster everything moves.
So what's one loop in your business you could close this week? And when will you actually do it?