Handing AI the Whole Project Is Not the Same as Leading It

November 15, 2026

In Deloitte's latest State of AI in the Enterprise survey of 3,235 leaders at the director level and above, 74% of companies plan to deploy agentic AI within two years, and only 21% report having a mature model for governance of autonomous agents. If your team now includes AI agents and you're tempted to hand them the whole project and walk away, this newsletter is your next move.

When I design curriculum, or when we are creating a learning or teaching module, I still believe in the 80/20 of agentic work. The agent can carry most of the labor. But the judgment comes first. Judgment-in-the-Loop starts at the beginning, not at the end.

I set the parameters. I share my expertise. I direct the AI: what to look for, what not to look for, and what to watch out for. I also ask it to challenge me, because my own perspective may be biased. We go back and forth.

Then the AI pulls everything together.

I come back in the last 10% to add my judgment and expertise. I do not give AI carte blanche and only show up at the end. If you arrive after the work already looks finished, what it produced can start to influence you. You end up reacting to its version of the work, instead of leading from the standard you set at the start.

Handing AI the whole project is not the same as leading it.

There's a saying in Spanish for this. In Mexico we say:

"Más vale prevenir que lamentar." (Better to prevent than to lament.)

It means it is better to take care at the beginning than to regret what you have to undo later. The English saying I often hear is "An ounce of prevention is worth a pound of cure." Same idea.

In my past newsletter Why Using AI Every Day Isn't Making You Better at It: Proximity Is Not Practice, I put the 80/20 split this way: the agent carries the repeatable 80%. You keep the 20% that takes judgment. What I want to add now is this: a lot of that 20% belongs at the start. You direct the work. Then you return at the end with your expertise still intact.

The goal isn't to do every task yourself. The goal is to set the direction first, stay in the loop for the judgment, and come back at the end to add what only you can add.

THIS WEEK'S INSIGHTS

None of these findings is set in stone, and none of them is all knowing. They are worth noticing, a way to stay informed, and a look at the writing on the wall.

1. Agents can now complete a much larger share of real remote work. On the Remote Labor Index, a benchmark from the Center for AI Safety and Scale AI that measures how often AI agents complete real remote projects, Claude Fable 5.2 leads with a score of 41.25, nearly double Claude Opus 5.5 at 21.25, as of November 14, 2026. A leaderboard score is not a claim about every workplace, but the direction is worth noticing. The work an agent can carry is getting larger, which is exactly why the leadership question is getting louder.

2. A more capable agent still may not notice when the work is not good enough. This week, Nathaniel Whittemore recapped new model tests on The AI Daily Brief. One point he highlighted comes from Wharton professor Ethan Mollick, who wrote about his early use of Fable 5.2 in Fable 5.2 and Jagged AGI. Mollick found that the model's failures were no longer mostly invented facts or bad math. In one test, he wrote, "Fable's first version was accurate and joyless." He wrote that Fable "put critics, fact-checkers or judges on nearly every project, but those checks found errors and added fixes. None of them asked whether the work was too long, too repetitive, or no fun." His conclusion is the line I keep coming back to: "Knowing when your own work is not good enough is a large part of what makes someone an expert, and it is the part Fable does least reliably."

3. Companies are moving to deploy agents faster than they are ready to lead them. In its 2026 State of AI in the Enterprise survey, Deloitte surveyed 3,235 leaders at the director level and above, across 24 countries, between August and September 2025. 74% of companies plan to deploy agentic AI within two years, and only 21% report having a mature model for governance of autonomous agents. Buying the agents is not the same as leading the work they produce.

Put those together and the hard part is no longer whether an agent can take a whole project. More often, it can. The hard part is directing the work from the start, using more than one tool when you need to, and making the call. When you don't stay in the loop, a project that looks finished can still miss the standard, and your name is still on it.

THE MYTH TO REFRAME

Myth: "If I hand the agent the whole project and walk away, I'm being an agentic leader."

Why we believe it: For most of our careers, letting go was the test of leadership. If you still touched every task, you weren't delegating. You were hovering. Now an agent can take a whole project, start to finish. Walking away starts to feel like the next level of that same skill. Staying close can feel like you don't trust the tool, or worse, like you don't know how to lead.

Reframe: Walking away from the project is not the same as leading it. An agentic leader doesn't do all the labor. An agentic leader also doesn't disappear. You set the intent. You define what "good enough" looks like. You place the checkpoints. And you make the final call. That is Judgment-in-the-Loop, and it starts at the beginning. Always keep a human in the loop. The human is not there to redo the work. The human is there to direct it, and to decide whether it meets the standard.

What to do: Before you hand a project to an agent, write down the outcome you want, the standard you will accept, what to look for, what not to look for, and what to watch out for. Ask the tool to challenge your thinking, because your perspective may be biased. Then stay for the checks, including the last 10%, when you add your judgment and expertise. If you lead a team, make that the expectation, not a personal habit.

TOOLS TO EXPLORE

This week's tool: a cross check, using more than one AI tool you already have

You don't need a new app. Use the AI tools your company already approved, and leave out client names, financials, and anything confidential.

In my experience, I don't use just one LLM. For me, each one has its own zone of genius, and current research on the latest models points in the same direction. I go to ChatGPT for logic and mathematical analysis. OpenAI describes GPT-6 Astra as built for complex reasoning, research, and document creation. I go to Claude for communication, learning design, and contextual knowledge. Anthropic positions Claude's latest models for reasoning, coding, and writing. I use Perplexity for research and for finding strong sources, because it is built for real-time, web-wide research and answers with built-in citations. Gemini helps me with search and recency, and Google's latest Gemini 4 models, announced in November 2026, are being used for complex knowledge work. And I use Grok when I want to be close to real time, because it searches the web and X live, so the answers reflect what is happening now. These are my working notes, not a ranking of the tools.

I also ask the tools themselves to help me define what each one is better at as new models come out. And I use a tool like ChatHub to put the same prompt into several models side by side, see which one answers best for me, and then decide which tool to use for that task.

I wrote about not betting on a single favorite in my past newsletter, Your Favorite AI Is Not Your Best AI. When you are directing a project, you can't hand the whole thing to one LLM and walk away. You direct it. You use bits and pieces. You wouldn't hand a whole project to one person on a human team and expect them to do every piece. A good leader knows who on the team is good at what, plays to their strengths, and divides up the tasks. Start treating LLMs as digital teammates the same way, with you as the human in the loop who keeps the judgment.

Run this 30 minute test:

  1. Pick one real project (3 minutes). Choose a learning module, a teaching plan, or another piece of work you would be tempted to hand off this week.
  2. Write the brief first (8 minutes). In a few short lines, write: the outcome you want, what to look for, what not to look for, what to watch out for, and where you want the tool to challenge your thinking.
  3. Run the first pass (6 minutes). Give that brief to one tool and let it pull the work together, up to a first checkpoint. Save what it gives you.
  4. Cross check with a second tool (8 minutes). Paste the first output into a different tool and ask it to challenge the work: what is missing, what is assumed, and where you or the first tool may have a blind spot.
  5. You decide (5 minutes). Compare both passes against the brief you wrote at the start. Keep what meets the standard, send back what doesn't, and add the judgment only you can add.

Prompt to steal:

Here is the brief I set for this project: [OUTCOME, WHAT TO LOOK FOR, WHAT NOT TO LOOK FOR, AND WHAT TO WATCH OUT FOR]. Here is the output I received from another AI tool: [PASTE THE FIRST OUTPUT]. Act as a thoughtful skeptic. Do not rewrite the whole thing yet. Tell me what is missing, what is assumed, where the first output may be biased or incomplete, and what I should watch out for before I accept this work. Challenge my thinking as well as the draft. Then wait for my direction before you change anything.

POWER TIP

Don't let the first finished draft be the first time you apply your judgment.

If you wait until the end, you are stuck reacting to a long deliverable that already looks complete. That is how work that is accurate, finished, and still not good enough slips through.

Set the parameters before any tool starts. Then, at the first checkpoint, run a cross check with a second tool and ask what is missing. As Mollick's tests showed, an agent can check for errors and still miss whether the work is actually good enough. You make that call.

TRY IT THIS WEEK

1. Direct the next module or project before any tool starts. Write the outcome, what to look for, what not to look for, and what to watch out for. Ask the tool to challenge you. Then, and only then, let it pull the work together.

2. Run one cross check. Take the first output to a second AI tool and ask what is missing or assumed. Notice what you would have accepted if you had only looked at the first draft.

3. If you lead a team, change the question you ask. When someone says an agent can take a project, ask "What direction will you give it, where will you check the work, and what does good enough look like?" Make judgment part of how the work gets led.

YOUR NEXT MOVE

Learning to direct AI with judgment in the loop starts with knowing how you work with it today. I have a special invitation just for you as a ¡Ay, Ay, Ay, AI! reader.

Take the FlipFactor™ diagnostic. It is the proprietary diagnostic from FlipWork™ that measures AI readiness. It is free, and it is built for you as an individual, whether you are an entrepreneur, a consultant, or anyone working outside a corporate role. You get a personalized report with specific next moves, and guidance on becoming a true agentic leader: an Orchestrator who directs AI with judgment in the loop, instead of handing it everything.

We also work with many leaders and organizations to roll out the FlipFactor diagnostic across entire teams with our FlipWork Pulse™ product, so they can get the lay of the land: how ready the team is, and how to meet people where they are. It identifies Guardians and Orchestrators, Explorers who are still building consistent habits with AI, and Pioneers who move quickly and may be outpacing the safeguards around them.

Take the FlipFactor diagnostic

👉 When you hand a project to AI, how do you decide that the work is good enough? Reply and tell me.

Closing Thought

AI can take more of the project now. That part is getting clearer every week.

The question is whether you are still leading it. Whether you set the direction at the start, stayed in the loop, and decided it was good enough, instead of arriving only after the work had already begun to influence you.

Direct the labor. Keep the judgment.

Más vale prevenir que lamentar. Better to prevent than to lament. The care you take at the beginning is what keeps the work yours.

¡Hasta la próxima, un abrazo fuerte!

(Until next week, a big hug!)

P.S. ♻️ Know someone who is building or leading a team that now includes AI agents? Send them this newsletter and invite them to subscribe to ¡Ay, Ay, Ay, AI!. If they want a clear read on how they work with AI, share the FlipFactor diagnostic with them too.

¡

The Newsletter

¡Ay Ay Ay, AI!

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.

Weekly · Five minutes · Free