November 2026 | Nano Tools | 

Three Ways to Work with AI, and How to Choose

Three Ways to Work with AI, and How to Choose

Nano Tools for Leaders® are fast, effective leadership tools that you can learn and start using in less than 15 minutes — with the potential to significantly impact your success as a leader and the engagement and productivity of the people you lead.

Goal

Decide how much of a task to keep, share, or hand off to AI, so you get the speed without surrendering the judgment.

Nano Tool

AI is very good at some tasks and surprisingly bad at others, and it is hard to tell which is which in advance. Work that looks trivial can defeat it; work that looks hard can be well within its reach. Wharton management professor Ethan Mollick calls this uneven boundary the jagged frontier, and it keeps shifting as models improve. That is why working with AI well is harder than it sounds, and why most of the advice about it, which is really advice about prompting, starts too late. The more important decision comes earlier. Once a task is worth using AI for at all, there are three distinct ways to work with it, and choosing the wrong one costs time, quality, or both.

Two of the modes keep a person in the loop from start to finish. In centaur work, the human and the AI divide the labor, with deliberate handoffs: one does the analysis and the other drafts the write-up, or one generates ideas and the other selects among them. In cyborg work, the two are intertwined. An experienced user might call on AI a dozen times in a single page, to finish a sentence, verify a fact, or check whether the argument still holds, while making every decision about when to ask. Both modes give control. Neither scales, because one person can only be in so many loops at once.

The third mode, conjurer work, is what agentic AI makes possible. The person defines the task, assigns it, and evaluates the result, all while being absent as the work gets done. It is the most powerful of the three and the riskiest, because it treats the AI as an independent worker rather than a tool, and it calls for a different skill: not steering in real time, but delegating clearly. A Harvard study by Ben Weidmann, Yixian Xu, and David Deming found that the ability to direct AI agents correlates closely with the ability to lead human teams. In an experiment by economists Alex Imas, Kevin Lee, and Sanjog Misra, people wrote instructions for AI agents negotiating a used-car deal on their behalf. Everyone had the same goal and the same model; only the instructions differed. Those differences explained 73 percent of the variation in outcomes, a bigger gap than the one between experienced and novice human negotiators. 

Action Steps

  1. Separate the work. Use a centaur approach when the task has parts that split cleanly, and one of them calls for judgment the AI doesn't have or is work you want to keep. Decide what the AI does and what you do, and hand off between the two deliberately.
  2. Work together. Use a cyborg approach when you need to stay engaged throughout but want help at specific moments: to finish a sentence, verify a fact, or check whether the argument still holds. You stay in the flow and make every call about when to ask.
  3. Hand it over, but do the math first. Use a conjurer approach for tasks you don't need to be present for: the AI can do the work, you can evaluate the result, and the time saved justifies the setup. Three variables decide whether that's true: how long the task would take you, how likely the AI is to succeed, and how long it takes to check what comes back. A one-hour task that takes half an hour to verify is a trap; a ten-hour task with an hour of verification pays off even after three tries. The longer and more routine the task, the stronger the case.
  4. Brief the AI like a new hire. Every profession has a delegation document, from product requirements to a creative brief, because intent gets lost in translation. Give the AI the same things: the objective and its context, what it may change and what it must leave alone, what finished looks like, the outputs required, and the checks it must run before declaring itself done. The more precisely you define done, the faster the review goes.
  5. Read everything, then diagnose before you retry. You weren't there during execution, so review is where your expertise does its work. For high-stakes checking, use more than one model; they have different blind spots. When a result misses, don't just say "try again." Work out whether the task was unclear, the context missing, or a constraint misread, and fix the instruction. Two or three focused rounds usually reveal whether the AI can get there or whether you've hit the jagged frontier.
  6. Reassess the division of labor. Because the frontier is jagged and moving, no task belongs permanently in one mode. If the AI keeps failing, do the work yourself, note what went wrong, and retest with newer models later. When it finally succeeds at something it couldn't do before, have it write down how, so future sessions can draw on it. And watch your own capacity. A Berkeley field study of a technology company, published in Harvard Business Review, found that AI made employees more productive and more burned out, because the old limits on a reasonable workload no longer applied.

How Leaders Use It

Klarna’s experience shows what it can cost to skip the choice. In early 2024, the payments company reported that its AI assistant was doing the equivalent work of roughly 700 customer-service agents and estimated that it would improve 2024 profit by $40 million. Fourteen months later, CEO Sebastian Siemiatkowski acknowledged that the company’s aggressive push to reduce costs had gone too far and that service quality had suffered. Klarna began bringing human customer-service workers back into the mix, with AI handling routine, high-volume questions, while humans remain available for more complex cases and when customers want human help. The frontier didn’t move; Klarna found out where it was.

When Shopify CEO Tobi Lütke told employees in 2025 that reflexive AI use was now a baseline expectation, he didn't leave the mode to chance. Before asking for more people or budget, teams had to show why AI couldn’t do the work, and every team lead was asked to picture their area with autonomous agents already on the team. The same memo described AI as a thought partner, researcher, critic, tutor, and pair programmer. It also added AI use to performance reviews. The expectation also applied to Lütke and his executive team.

McKinsey has taken conjurer work to scale. In early 2026, CEO Bob Sternfels said the firm’s workforce was 40,000 people and roughly 25,000 AI agents, up from a few thousand agents 18 months earlier, with a goal of at least one agent for every employee. The skill the firm now hires for, according to a senior partner, is the ability to move between a consultant's judgment and an engineer's mindset while working with AI as part of delivery.

Knowledge in Action: Related Executive Education Programs

Leading AI and Digital Transformation

Generative AI and Business Transformation

Analytics for Strategic Growth: AI, Smart Data, and Customer Insights

Contributor to This Nano Tool

Ethan Mollick, Associate Professor of Management and Ralph J. Roberts Distinguished Faculty Scholar at the Wharton School, and co-director of the Generative AI Labs at Wharton; author of Co-Existence: The Next Phase of AI (Portfolio, 2026) and the New York Times bestseller Co-Intelligence: Living and Working with AI (Portfolio, 2024).

About Nano Tools

Nano Tools for Leaders® was conceived and developed by Deb Giffen, MCC, Director of Innovative Learning Solutions at Wharton Executive Education. It is jointly sponsored by Wharton Executive Education and Wharton's Center for Leadership and Change Management, Michael Useem, Director. Nano Tools Academic Director is Professor John Paul MacDuffie, Professor of Management at the Wharton School and Director of the Program on Vehicle and Mobility Innovation (PVMI) at Wharton's Mack Institute for Innovation Management.

Download this Nano Tool as a PDF