How AI Is Changing Strategy: Rethinking What Your Business Is Designed to Do

Host Adi Ignatius opens by asking what competitive advantage can mean when every company can access AI for idea generation, analysis, software writing, and complex tasks. Ethan Mollick, Wharton associate professor and author of the forthcoming book Coexistence: The Next Phase of AI, says two developments have changed the strategic picture. First, AI’s capabilities are still on an exponential curve with no obvious sign of stopping. Second, agents have become viable: AI systems can work autonomously on real processes and important tasks rather than merely answering prompts. Mollick warns that treating AI as a productivity exercise creates more output without necessarily creating more value. If companies do not change their processes or key performance indicators, employees may simply produce more PowerPoint presentations or more code. Those productivity gains can be “eaten up by the organization.” Leaders therefore need to decide what the organization is actually trying to produce and whether an intermediate output, such as a presentation, was ever the real goal. He says AI can now regularly perform work that would have taken a human two to 18 weeks, at reasonable cost and with high success rates. That means jobs begin to blur—coders do design work, designers do coding, and product managers may do much more of everything. Leaders must reconsider job boundaries, organizational goals, values, and the processes used to reach them; this is already happening inside companies, not merely as a future academic question.

When Alison Beard asks how leaders can turn urgent conversations into decisions, Mollick recommends a three-part organizational model: leadership, lab, and crowd. Leadership must do more than tell employees to “use AI.” Senior executives need to decide how AI will be used, what the organization should do with it, and how incentives will change. The crowd means giving employees access to current, capable models. If people use inferior systems because the company has not upgraded, they develop a distorted understanding of what AI can and cannot do. The lab is a continuously operating group that is not primarily technical. It should include technical and organizational specialists who benchmark what systems can do, build future processes, and redesign the organization around AI. The lab combines ideas from employees using the tools and decisions from senior leadership. Mollick argues that this must be a whole-organization response rather than an isolated technology project. At the individual level, he sees willingness to “eat risk” as a major differentiator. Leaders who succeed accept that nobody knows exactly what will happen and experiment anyway, because refusing to act is also risky. They should use the newest systems themselves rather than waiting for reports, training programs, or consensus. Leaders who personally experiment learn where the technology is useful and where it remains unreliable.

Mollick explains that his earlier models of human-AI collaboration—“centaurs,” who divide work between human and machine, and “cyborgs,” who continually move back and forth between them—are being overtaken by what he calls “conjurer work.” In independent studies, AI has moved from reliably completing roughly 30 to 45 minutes of work to completing 16 or more hours of work from a single instruction, with a good chance of success. A conjurer effectively summons a temporary team. Instead of prompting an AI, inspecting the result, and prompting again, a person can write something resembling a project request or standard operating procedure and assign the work to an agent. The system then carries out the assignment. Mollick says this changes the meaning of management because everyone now has AI to manage, and AI skills increasingly resemble management skills. He cites an Anthropic study suggesting that people who are most successful at producing working code on the first pass are managers rather than coders, because they specify what they want effectively and let AI execute. He says the danger is that organizations with little imagination will use this capacity only to automate work or turn employees into verifiers of AI content. That would erase differentiation. Competitive companies should combine human and AI work, process, and culture to accomplish things neither could do alone, rather than becoming interchangeable cloud-based companies.

When Ignatius asks how companies can differentiate if they all have the same frontier models, Mollick contrasts a “dark factory” with what he calls a “twilight factory.” In a dark factory, robots do all the work without humans or even lights. He points to StrongDM as an example he has heard about: the company ships software with no human looking at or writing code, under rules that make the process 100% AI automated. Mollick considers that approach dangerous because it lets the code make all the decisions and removes human variation. A twilight factory would let AI do substantial work but proactively bring people into the process when their different views, expertise, approvals, or interest matter. Mollick argues that this integration—bringing humans in at the right points and at the right times—can become a genuine differentiator. He says leaders may have a swarm of 100 agents working on a problem and receive progress reports and demonstrations much as a manager receives updates from a team. Being “in the loop” no longer means approving every action. AI can also expand leaders’ span of control. Mollick has spoken with executives who diagnosed problems in plants that had never been raised to them because they could query all the information available to the company and ask the AI to investigate. Agents already make strategic decisions involving judgment, creativity, and trade-offs. Leaders must decide which decisions AI should execute, which options it should surface and pressure-test, and where the leader must choose.

Mollick says leaders cannot decide what to delegate by relying on permanent bright lines, because AI systems are “jagged”: they can be excellent at unexpected tasks and poor at others. The answer is disciplined process design. Companies need to define the boundaries between functions and determine where human judgment, culture, and organizational knowledge are essential. AI may eventually be better at pricing or systems engineering, but each company has to test that proposition in its own lab rather than assume it. For students and aspiring leaders, Mollick identifies four capabilities. Deep knowledge lets someone recognize whether a spreadsheet, argument, or proposed answer is wrong without reading every page of generated material. Wide knowledge helps people cross disciplinary boundaries and give AI better instructions and feedback. Taste helps them choose among many possible strategies, designs, or outputs and decide what fits the company. Agency is the willingness to act under uncertainty and experiment, which Mollick says is harder to teach. He rejects the idea that older leaders are automatically disadvantaged while younger “digital natives” understand AI. Unlike the internet, AI does not depend on an innate cultural background, and it rewards expertise. Studies involving BCG consultants found that less-experienced consultants produced worse results with AI, and other research, including from Anthropic, points in the same direction. Senior leaders may have an advantage because they possess domain judgment and experience with major changes.

Mollick says both large and small companies can benefit from AI. Large companies have resources, customers, and experienced employees, yet some executives respond to demonstrations by saying they will simply ride out the disruption. He considers that irrational because the people who can discover useful applications are already inside those companies. Startups, meanwhile, have rarely had a better opportunity: AI can make a small team competent at functions it previously could not afford, such as payroll, marketing, or software support. He identifies two major strategic mistakes. The first is anchoring on current capabilities and current costs. What seems to be the central issue today—such as token spending—may be less important in a few months as models become cheaper and organizations use multiple agents. The second is assuming that vendors or consultants have already solved the problem. Mollick says the technology is only about three years old, many purported success cases involve obsolete models, and companies need internal capacity to develop their own answers. His prescription is organizational ambidexterity: exploit what works today while continuing to explore what becomes possible next. Leaders should ask what AI has made trivial and therefore should be stopped, what the organization is building that seems impossible and could change the industry, whether it is building something that does not work yet in anticipation of better models, and whether leadership, lab, and crowd are in place. He also warns that unguarded future models could produce severe cybersecurity problems. Agents may attack internet-exposed systems not out of malice but while pursuing an assigned task, making security planning urgent.

Mollick cautions leaders against treating AI as a normal technology project with a long requirements process, a six-month request-for-proposal phase, and an eight-month discovery phase. Because capabilities change so quickly, a supposedly de-risked project may be obsolete when it launches. Companies need experimentation, permission to fail, and the organizational culture to surface what employees are actually doing with AI. Psychological safety, rewards, openness, and good management matter because employees will hide their uses if disclosure could get them fired or disrespected. On employment, Mollick says economists do not yet have a reliable answer. Junior workers may be squeezed because they lack the expertise that helps people use AI effectively. But AI could also increase hiring inside organizations by allowing companies to bring activities such as coding or marketing in-house, even as demand for centralized service firms declines. The AI labs’ explicit goal, he says, is to replace all human labor, but he is not convinced they are right and says society should not simply accept that outcome. He also worries about people being removed from the interesting parts of jobs and left only to verify AI output, about students cheating instead of learning, and about damage to the information environment. At the same time, AI can tutor students, help people who were previously stuck, and provide useful medical or legal information—though people should still consult doctors and lawyers. His conclusion is that policy and leadership should encourage beneficial uses, mitigate harmful ones, and ensure that AI augments rather than disempowers people.