Escalate or Approve?

Most people learning AI practise the wrong thing. They get better at asking the model for output. The skill that actually separates people is deciding where the human sits once the output exists.

That decision has a name. Every AI workflow you build either asks a person to approve each result, or asks the machine to escalate only the cases that break a rule. Approve or escalate. It sounds like a small design choice. It is the whole game.

Two Places a Human Can Sit

In an approval model, the human sits at the gate. The AI drafts something, then it waits. A person reads it, maybe tweaks a line, clicks yes, and only then does the work move on. Every single output stops at that gate.

In an escalation model, the human sits at the exception. The AI handles the routine, checks its own work against rules someone wrote down, and only raises its hand when something breaks one of those rules. The person shows up for the small slice of cases that actually need them.

Here is why that matters. With an approval step, your workflow runs at human speed, no matter how fast the model is. You put a faster engine in a car that still stops at every intersection.

The Number That Settles It

Stanford's Digital Economy Lab measured this. In The Enterprise AI Playbook (Pereira, Graylin and Brynjolfsson, March 2026), the authors compared how companies place people inside AI workflows. Escalation-based models, where AI handles 80% or more of the work autonomously and humans review exceptions, delivered 71% median productivity gains. Approval models delivered 30%.

Same models. Same vendors. Often the same use case. The gap came from one design decision.

The same playbook cites McKinsey on the behavior behind the result: 55% of high performers redesigned workflows around AI, against 20% of other companies. The report's own summary is blunt: high performers redesign workflows, not just deploy tools.

If you are learning AI right now, read those two numbers as a hint about what to practise. The people getting big results are not the best prompters. They are the people who can redesign a workflow and decide where the human line goes.

Why Students Default to Approval

When you are new, approval feels safe. You don't trust the output yet, so you read all of it. That instinct is fine for your first week with a tool. It becomes a trap when it turns into the design.

The real reason most people stay at the gate is not caution. It is that they cannot say, precisely, what an exception is. And you cannot say what an exception is until the workflow is written down end to end. If you can't name what wrong looks like at each step, "review the exceptions" quietly becomes "review everything and hope."

So the skill is not bravery. It is documentation. That is less exciting than a new model release, and it is exactly why so few people are good at it.

A Real Example: How Edge8 Publishes Its Blog

At Edge8, my company, I run the blog as an escalation model. It is the smallest complete example I have, so here is how it works.

• A person, usually me, writes the idea and starts the writer agent.
• The agent works through eight passes: drafting, editing, the SEO package, exhibits, the hero image, links, assembly and validation.
• When the run finishes, a person sets a publish date. The days before that date are the review window.
• If nobody pulls the post back, the daily routine publishes it at 11:00 Vietnam time.

Notice what is missing. Nobody presses Publish on the day. Nobody reads every line as a gatekeeper. A person still starts the work and still sets the date, but the routine no longer waits for a yes.

What makes that safe is the checklist. The writer agent names 14 failure conditions and checks itself against every one before a post ships. A few of them:

• An em dash anywhere in the copy. A brand rule a human used to enforce by reading.
• A slug that is already taken. A collision a tired reviewer could easily miss.
• An exhibit showing a number the body never states. A consistency check machines do better than people.

Each of those used to be something a person caught by reading. Now each one is a line the agent checks, and when a check fails, the failure names the rule it broke. A post that fails validation stays scheduled, is retried, and is named in the report. It never silently disappears.

That is the design in miniature. The human is not asked to catch problems. The human is told about them.

The Ratio Is a Dial, Not a Leap

You don't go from approving everything to approving nothing in one step. The Stanford playbook describes a financial services company that chose an 80/20 model for marketing content, AI generating and humans refining, and took time to market from seven weeks to six hours. The report treats that human 20% as transitional, expected to shrink as the AI improves.

That is the honest starting point. Edge8's blog sits roughly there too: there is an auto-publish switch, and for now it is off, because the brand rules are still being tuned. When the rules are tight enough, the switch flips and nothing gets rebuilt.

The lesson for a student: you are not designing for zero humans. You are designing so the human share is a documented, measurable slice that can shrink on purpose. An approval step with no rules stays at 100% forever, because nobody can say what would make it safe to remove.

The Exercise: Draw the Line on One Workflow

Do this once, on paper, this week. Pick one workflow you actually run, at work or in your own projects. Something with AI in it, or something you want to put AI into. Then:

1. Write it end to end on one page. Where it starts, every hand-off, what information each step needs and where that information lives, and where it ends.
2. Mark where the human sits today. Circle every step where a person reads, checks or approves.
3. Mark where things actually go wrong. From memory or from history, put an X wherever real problems have shown up.
4. Compare the circles and the Xs. In most workflows they are far apart. The human reads everything, and the real problems cluster in three or four places.
5. Turn each X into a rule a machine could check. Not "looks wrong." Something specific, like "a number in the summary that does not appear in the source." Aim for five rules.
6. Redraw the workflow. Move the human from the circles to the rules. Decide what happens when a rule fails: who gets told, and what the message says.

You have just done the job. That one page is a workflow redesign. If you want a companion skill for step 5, read how to write a definition of done for AI, because a good exception rule is just a sharp definition of done turned inside out. And if you are still deciding what to hand the machine in the first place, start with how to delegate to AI.

One Trap to Watch For

Your rules only work if the machine can see the data it needs to check them. Edge8 worked with a footwear retailer that had eleven systems and workflows already written down, and still could not run a single one, because the data those workflows needed lived in systems that did not talk to each other. Every process stayed manual and ended in a spreadsheet.

So when you write a rule in step 5, ask one more question: could a machine actually see the information to check this? If not, you have found the next thing to fix. A documented workflow without connected data can describe an exception, but it cannot detect one.

This Is What Gets Tested

I care about this skill because it is the one most people skip. Plenty of people can get good output from a model. Very few can take a messy real process, write it down, define its exceptions and move the human to the right place.

That is why AI Officer certification challenges use your own data and your own workflows, not a toy case. To certify you complete the core challenges, 8 elective micro-sessions and 4 live coaching sessions, with an AI Buddy coaching you along the way and weekly CAIO office hours when you get stuck. The exercise above is the kind of work you will be asked to do and defend. It is a certificate of proof, not attendance.

If you want to prove you can draw the human line on a real workflow, start with AI Officer certification. And if you want to compare your one-page workflow with other people doing the same exercise, bring it to the community.