AI tools now give your managers a forecast, a score or a recommendation. The decision that follows is still the manager's, and the better the tools get, the less often managers practice it.
On the night of June 1, 2009, Air France flight 447 was cruising at 35,000 feet over the Atlantic, from Rio de Janeiro to Paris, on autopilot. The captain had left the cockpit to rest. Two copilots were at the controls.
At 2:10 a.m. (Universal Time), ice crystals most likely blocked the probes that measure airspeed. The speed readings became incorrect, and the autopilot disconnected. The copilot flying the aircraft said, "I have the controls," and pulled the nose up.
Investigators later found that the copilots had never trained in flight, at high altitude, for unreliable airspeed or for flying the aircraft manually. The aircraft stalled and stayed stalled until it hit the ocean, about four and a half minutes after the autopilot disconnected. All 228 people on board died.
The automation worked for almost the whole flight. When it stopped, two people who had never practiced that situation in flight had to make the hardest decision of the flight.
What Does This Mean As We Move Into the Age of AI?
As AI makes more of the predictions, the decisions it leaves to your managers become less frequent and more difficult. Managers have to practice those decisions deliberately, before they must make one at work.
In MIT Sloan Management Review this October, Jeffrey Camm, Michael Fry and Scott Shafer describe a retailer whose data science team builds accurate demand forecasts, and a CEO who then asks how much inventory to hold. They conclude: "The models predict demand but do not tell management what to do with those predictions."
Elizabeth Renieris and her coauthors, reporting on a panel of AI experts convened by the journal and BCG, add a second point. "Judgment is a skill that will atrophy if it isn't exercised," they write, and organizations "should deliberately build in opportunities to exercise judgment outside of high-stakes moments."
Your managers already act on AI output: a renewal forecast, a credit score, a recommended price. Most of the time, the model is right, and the manager can act on it. The decisions left to the manager are the cases the model cannot handle well: a fact the model was not given, a case it was not designed for, or data too limited for a confident answer. In each of these cases, the manager makes three decisions.
Decision 1: Ask What the Model Wasn't Told
The common move: Accept the recommendation because it looks complete. A model cannot warn you about a fact it was never given. The renewal model says the account is healthy, so the manager approves the plan and goes to the next account.
The cost: The model was right about its data and wrong about the account. The client's executive sponsor left the company last month, and no system recorded it. The renewal does not happen, and nobody expected that because nobody asked.
The move that works: Treat the model's confidence as a statement about its data, not about the account. Then ask the person who works with the account most one question: what has changed that no system would record?
Decision 2: Override, or Don't, and Say Why
The common move: Override the model based on instinct, or never override it, because nobody wants to be the person who questioned the system and was wrong.
The cost: In both cases, the organization learns nothing. Nobody can check a decision based on instinct. A team that never overrides the model has let the tool make the decision.
The move that works: Measure your own decisions the way the model is measured. Before the result is known, write down why you overrode the model, or why you didn't. When the result is known, check that reason with your team, so the team learns where your judgment is better than the model's and where it is worse.
Decision 3: Act Before the Data Is Complete
The common move: Wait for the model to be certain. "Let me pull another report."
The cost: The model is often least certain when the decision matters most. Waiting is also a decision, and you can lose the chance to act while you wait.
The move that works: Before you ask for another report, say what result in that report would change your decision. If one would, ask for that report and nothing more. If no result would, the report is only a delay: decide now, and set a date to check the decision.
Why a Course on the Tool Doesn't Solve the Problem
Most AI rollouts include training on the tool: the features, the prompts, where the buttons are. That training is necessary, but it does not cover the decision the model leaves to the manager. The better the model, the less often managers make that decision, so daily work no longer gives them practice in it.
The French investigators found the same problem in pilot training. "Current training practices do not fill the gap left by the non-existence of manual flying at high altitude," they wrote. They also noted that in simulator exercises, trainees "are more or less familiar with the failures they will have to deal with." Replace "manual flying" with the decisions your managers no longer make without a model, and the sentence describes most AI rollouts.
Knowing the right answer is not the same as making the right decision when the time to decide is short. A course on the tool fails for the same reason most leadership training doesn't change behavior.
The Cost Nobody Tracks
The cost appears in the business number the decision affects: inventory held, credit losses, renewals kept. Every manager who accepted the same recommendation in the same way adds to that cost, and nobody sees it until the quarterly results. Because the model's prediction was correct for its data, nobody connects the loss to the manager's judgment.
How Blueline Solves This Problem
Camm, Fry and Shafer advise business leaders to "start modeling your decisions." Blueline's high-stakes performance simulations start the same way, then give the person who makes the decision practice in making it.
Consider the renewal from Decision 1. Decision-Consequence Mapping™ (DCM™) maps that decision, starting from the renewal number you name: the common move, the move that works, and the consequence of each. In a simulation built from that map, each manager talks with the account lead. The account lead mentions that the sponsor has not answered email for weeks. Nobody labels that as the moment to decide, and unlike the pilots in the French investigators' simulator exercises, the manager is not told which signal matters. Asking for another report is a choice too, with a consequence. A manager who ignores the account lead's comment finds the account lead less willing to share in the next conversation.
What lets every manager on your team practice that conversation is BluEQ™, Blueline's proprietary behavioral intelligence infrastructure. It carries the DCM™ design into every conversation, so each character remembers what the learner did and responds the way the real person would.
Afterward, you see each manager's decision path, for a whole team, before a real renewal depends on it: who asked about the sponsor in time, who approved the forecast, and at which point. You name the metric and provide access to the data that measures it, and that metric becomes the baseline for ROLI (Return on Learning Investment).
Start Here: Pick One Decision
Choose one decision your managers make with help from an AI tool. Pick the one that costs the most when they choose wrong.
This week, look at the last few times your team made that decision. For each one, check:
- Did anyone ask what the model wasn't told?
- Did anyone write down why they followed the model or overrode it?
- Before asking for another report, did anyone say what result would change the decision?
You'll see which managers already do this, and which don't.
At DevLearn? Come to booth 333 on Wednesday, November 4, or Thursday, November 5. Play a Blueline simulation yourself and see each decision you made and where it led. Then bring your decision, and we'll talk about what a simulation of your decision would look like. Not at DevLearn? Bring it to a private demo.
If someone else is responsible for the results of that decision, send them this post and the ROLI Blueprint, which explains the business case.
The French investigators wrote that for situations that disconnect the autopilot, "only specific and regular training can provide the skills necessary." Your managers need the same practice for the decisions your AI tools leave to them.
Sources: BEA (Bureau d'Enquêtes et d'Analyses), Final Report on the accident on 1st June 2009 to the Airbus A330-203 registered F-GZCP operated by Air France, flight AF 447 Rio de Janeiro – Paris (July 2012), pp. 17, 22, 24, 185, 198, 204, 205. Jeffrey D. Camm, Michael J. Fry, and Scott M. Shafer, "Why Data Modeling Is Not Enough," MIT Sloan Management Review, October 19, 2026. Elizabeth M. Renieris, David Kiron, Steven Mills, and Anne Kleppe, "The Defining Challenge of Responsible AI Is Designing Systems That Preserve Human Judgment When and Where It Counts," MIT Sloan Management Review, October 15, 2026. DevLearn 2026, Schedule at a Glance, accessed October 23, 2026; booth 333 as published in Blueline's DevLearn 2026 Expo post.
