AI Does the First 70%. The Last 30% Is Still Yours.

Here's a pattern I see in almost every business that starts using AI seriously. The first week, someone drafts a proposal in four minutes that used to take two hours, and the whole team lights up. Three weeks later, that same team is quietly frustrated, because the AI keeps getting them "almost there" and never quite all the way.

Both reactions are right. That's the confusing part.

AI really is that fast at the first chunk of almost any job. And it really does stall out before the finish line. Once you see the shape of that, a lot of the hype and a lot of the disappointment both start to make sense.

The shape of the work changed

Think about any real piece of work you do that involves words, numbers, or decisions. Writing a proposal. Posting a job opening. Comparing three vendor quotes. Drafting a customer email sequence. Putting together a simple policy.

For most of these, the work splits into two uneven parts.

The first part is getting something real on the page. Research, structure, a solid first draft, the boring scaffolding. Call it the first 70%. This used to eat most of the clock, and it's exactly the part AI is shockingly good at. Pointed at the right task, it can produce a competent first version faster than you can find the old template. If your drafts come back generic, that's usually a prompting habit, not a ceiling. There's a whole post on getting better answers out of AI.

The second part is smaller but stubborn. It's the judgment calls. The context only you have. Deciding what's actually right for this customer, this deal, this moment. And somebody owning the result when it goes out the door. Call it the last 30%.

AI didn't shrink your work. It moved it, from producing to finishing.

Here's what trips everyone up. The first 70% got ten times faster. The last 30% barely moved at all. So the finish now feels slow by comparison, even though it takes the same time it always did. You just never noticed it before, because the slog of the first 70% hid it.

Two ways this burns people

Once the work has this shape, there are two predictable ways to get hurt. I've watched both happen plenty of times.

Failure one: you mistake the fast 70% for the whole job. The draft looks good. It's clean, confident, well organized. So it goes out. And then a client notices the proposal quotes the wrong scope, or the "personalized" email clearly wasn't, or the policy has a line that doesn't match how you actually operate.

This is where the bad reputation comes from. When people complain about generic, slightly-off AI output flooding their inbox, they're not describing what AI can do. They're describing what happens when someone ships at 70% and calls it done. Trusting a confident draft without reading it the way the reader will is its own well-documented trap.

Failure two: you expect the AI to deliver 100%. The other team waits for the machine to hand them something finished, ready to send, no thinking required. It never comes. The AI gets them to a strong draft and then goes vague or generic exactly where their specific situation matters most. So they decide "AI doesn't really work for our business" and go back to doing the whole thing by hand.

Both teams misread the same thing. One trusted the 70% too much. The other expected the 30% to disappear. The fix for both is identical: plan the work around the shape instead of fighting it.

Working with the 70/30, not against it

A few practical shifts, none of them complicated.

• Stop doing the first 70% by hand. This is the easy money, and a surprising number of people still skip it. If you're writing proposals, job posts, or routine emails from a blank page, you're spending two hours to save thirty minutes of review. Let AI do the draft. Every time.
• Protect your attention for the 30%. The time you save on the draft doesn't vanish. It should move to the finish. Read it like the customer will. Check the one number that has to be right. Add the context the AI couldn't know. The finish is now the highest-value thing you do on that task, so treat it that way instead of rushing it because the draft came so fast. This is also a good spot to point AI back at your own work as a second set of eyes before it goes out.
• Decide which 30% is non-negotiable, by task. Not every job needs the same finish. A low-stakes internal note can ship at 85%. A pricing proposal, a contract clause, or anything a customer will hold you to needs a careful human pass every single time. Write down, for your common tasks, what a person has to check before it goes out. That short list is worth more than any clever prompt.
• Know when a task is all 30%. Some work has almost no first 70% to hand off. A delicate conversation with an unhappy client. A judgment call about letting someone go. A strategic bet with real money behind it. Starting those with AI is starting at the wrong end. Use it to pressure-test your thinking if you like, but the work itself is yours from the first minute.

Why the last 30% stays yours

It's worth being clear about why that last slice is so resistant, because this isn't a temporary limitation the next model quietly fixes.

AI is good at the first 70% precisely because it has seen a million versions of it. Most proposals, most job posts, most policies look a lot like a million others, so AI produces a solid generic version in seconds. But your last 30% is the part that is not like a million others. It's your pricing, your client's quirks, your risk tolerance, the thing you know about this deal that was never written down anywhere. That gap, between what AI knows in general and what only you know about your business, is exactly where the finish lives.

That's the same reason a sharp new hire can draft you a decent memo on day one but can't make the final call on a key account for months. The draft is general knowledge. The decision needs your knowledge, and nobody but you and your team actually has it.

And there's the part no model takes off your plate: when it goes out with your name on it, you own the outcome. A tool can produce the words. It can't carry the responsibility for them. The finish is where that responsibility lives, and it doesn't transfer.

The honest version

The "70/30" is a way to think, not a measurement. For some tasks AI does 95% and you barely touch it. For others it gets you 20% and the rest is all you. The ratio isn't the point. The point is that almost every task now has a fast first part and a slow, human finish, and the two don't move at the same speed.

So the next time AI gets you "almost there" and you feel that flash of frustration, try reading it differently. It didn't fail to finish. It did the part it's built for, fast, and handed you the part that was always going to be yours. That handoff isn't the tool falling short. It's the tool working exactly as it should.

The businesses getting real value out of AI aren't the ones with the fanciest tools. They're the ones who figured out where their finish line actually is, and stopped confusing "fast" with "done."

If your team is either shipping half-baked AI output or quietly giving up because the tools never seem to finish the job, that gap is usually a workflow problem, not an AI problem. Designing the handoff so AI nails the 70% and your people own the 30% is exactly the kind of thing we help businesses sort out at Humanity AI.