The slide toward surrender

Addy Osmani recently published a piece called Critical thinking when agents do the work. He wrote it for software engineers, but the idea at the center of it applies to anyone who uses AI at work. It's the clearest description I've found of a habit every team new to these tools should watch for.

He calls it a slide toward surrender.

Three stages, one slide

Osmani describes a spectrum. He borrows the term "cognitive surrender" from Wharton researchers Steven Shaw and Gideon Nave.

• Offloading. You hand AI a task and still own the result. You read it, fix it, and it's yours. This is the healthy way to use it.
• Deferral. The AI's output becomes your starting point, and you stop questioning it. It gets treated as settled.
• Surrender. You stop checking at all, and whatever the AI produced quietly becomes your own judgment.

Nobody decides to surrender. You slide there. AI output is usually polished and mostly right, so it rarely gives you a reason to look twice. Every time it turns out fine, you check a little less, until one day you aren't checking.

What the slide looks like in a small business

Quotes. You start having AI turn your notes into client quotes. The first few, you check every line. By the tenth, you glance at the total and hit send. Then one goes out with last year's rate, or a line item the client never asked for, and now you're negotiating against your own number.

Client emails. The reply reads well and the tone is warm. It also promises a turnaround you can't hit, or answers a question the client didn't ask while skipping the one they did.

Policies. A refund policy or handbook section sounds official and complete. Buried in it is a term you'd never agree to. Once it's posted, it's your policy.

In every case, the AI did the drafting and your name went on it.

Posters, graphics and visuals: where the slide moves fastest

In my view, visuals are where this happens quickest. Images are fast to generate and easy to approve at a glance, and a glance is exactly the problem. Four patterns show up again and again.

Taking the first decent one. Image tools are very good at "pretty good." The first result looks finished, so it ships. Looking finished doesn't tell you whether it says what you need, to the people you need to reach.

Wrong details. Generated graphics still get text wrong. A misspelled word, a date that's off by a day, a phone number with two digits swapped, a price that doesn't match your menu. A flyer with the wrong date does more damage than no flyer at all. Read every word on a generated image, and read names, numbers and dates character by character.

Off-brand drift. AI doesn't know your colors, your fonts or how your business looks. Each graphic seems fine on its own. Line up a month of your social posts, though, and it can look like five different companies made them.

Endless regenerating. This one looks like the opposite of surrender, but it's the same slide. You hit regenerate twenty or thirty times, hoping the next one will be the one. At that point you've stopped deciding and started waiting for the tool to decide for you.

The fix for all four is the same: decide what good looks like before you prompt. Write down who it's for, the one message it has to get across, the details that must be exact, your brand colors and fonts, and where it will run. Then judge every result against that list. If three tries miss, rewrite the brief instead of rolling again. Osmani gives the same advice for prompts in general: rewrite the request rather than endlessly tweaking it.

Where to draw the line

You don't need to check everything with the same care. Osmani's rule of thumb works well for small teams too. Let AI run freely when mistakes are cheap and easy to fix: brainstorming, a first draft only you will see, ten headline options to pick from. Keep the thinking yourself when someone else pays for a mistake. That means clients, customers, employees and your reputation. Quotes, contracts, policies, anything printed and anything public all fall on that side.

He makes one more point worth stealing. Decide what "checked" means before you start. If you don't know what you're looking for, a quick glance will feel like a review.

A review checklist you can actually use

Keep this next to wherever your team uses AI. It takes a couple of minutes per piece.

1. Did I decide what good looks like before I asked? Audience, goal, and the must-have details.
2. Did I read all of it? Not just the opening and the total.
3. Are the names, numbers, prices and dates right? Check them against the source, not against memory.
4. Does it promise anything we haven't agreed to? Turnarounds, refunds, features, discounts.
5. Is anything made up? A stat, a quote, a policy term, a service you don't offer.
6. For visuals: did I read every word, and is it on-brand? Colors, fonts and logo used correctly.
7. What would make this wrong, and did I check that one thing?
8. Would I put my name on this if I'd made it by hand?

One last check for the team as a whole: if AI-made work is piling up faster than anyone can review it, slow down. Volume is how the slide speeds up.

Stay the one who decides

AI is worth using for quotes, emails, policies and graphics. It saves real time. The skill is staying the person who decides what goes out the door, even when the draft looks good enough.

Credit to Addy Osmani for the framing in "Critical thinking when agents do the work." The examples and checklist above are my own.