Spend time among management consultants advising on artificial intelligence (AI) and you will hear familiar prescriptions: use it to automate repetitive tasks, tidy up paperwork, process large datasets more quickly. All this will boost your firmʼs productivity, the argument runs, and if you adopt AI early, you will gain an edge over competitors.
The productivity gains are likely to be real. But an edge won by efficiency alone is unlikely to last. The world enjoys only a handful of large language models that matter (the type of AI that powers ChatGPT), and they are widely available at prices most firms can afford. When competitors use the same tools for the same efficiency-seeking strategy, they are more likely to end up with similar ways of working and less diverse offerings, not an advantage.
A more meaningful use of AI is needed to gain a true competitive edge, argues Ethan Mollick of the Wharton School of the University of Pennsylvania, whose work on AI in practice has made him one of the most influential voices on how firms can best use the technology. In conversation with Economist Enterprise—parts of which you can listen to across this article—Mr Mollick offers a framework for fostering AI-driven innovation built around three main elements: a “crowd” of employees empowered to experiment, leadership that sets the right incentives and a dedicated “lab” that captures and scales breakthroughs. Have all three and AI might indeed set your firm apart.
Look for innovation at the bottom
Popular accounts of how innovation happens often include a lone genius blessed with sudden insight. But inside most firms the process is more prosaic. Mr Mollick argues that real improvements begin with the “crowd”, the staff who face, each day, the frustrations, inefficiencies and occasional creative opportunities of their work. And because AI tools are now universally accessible, often cheaply, those closest to a task can easily experiment with how the technology might help. A marketing manager can test whether AI improves campaign ideation. A lawyer can explore how it can help to draft better memos. Each experiment costs little more than time and yields practical knowledge about what works.
The evidence supports this bottom-up approach of using AI for meaningful improvements. A field experiment by Mr Mollick and colleagues at Harvard University and the Massachusetts Institute of Technology (MIT), conducted in collaboration with BCG, a management-consulting firm, describes how the use of GPT-4, a model from 2023 made by OpenAI, augments the work of over 750 consultants in “realistic, complex and knowledge-intensive” assignments. The results are stark. On tasks where the modelʼs strengths—writing, ideation and synthesis—were relevant, consultants using it produced work that human evaluators judged to be 40% higher in quality than that of consultants working without AI.

Yet the same experiment also reveals that when AI is used in tasks that demand human judgement—including piecing together contradictory interview notes and weighing spreadsheet data—the model can become a liability. Consultants aided by GPT-4 were 19% less likely than those working on their own to reach the correct answer in such tasks, and the diversity of their ideas fell by 41% as the model steered them towards similar conclusions. It is a cautionary finding for an industry that sermonises about “differentiation” with AI, but subsequent research by Mr Mollick and colleagues at The Warton School offers hope. It finds that when a prompt asks GPT-4 to show its reasoning in steps, an approach known as chain-of-thought structure, the model widens its pool of distinct ideas by about a third and restores a level of diversity close to that of human brainstorming. Newer models, such as OpenAI’s GPT-5.2, promise even more diverse and "creative" outputs.
The pattern across successes and failures is that AI can amplify good judgement but cannot substitute it. The lesson for most firms is that expertise and careful thought are essential to making AI useful for innovation, but not using the technology to pursue improvements would almost always be a mistake.
Lead with assurances
Part of being a good boss in the age of AI is about striking the right tone of ambition and reassurance. Staff need to know that leadership expects the technology to deliver change, but also know that experimenting with it—and making it useful—will not lead to job losses. As Mr Mollick puts it, “workers understand perfectly well that 'productivity gains' with AI often means 'weʼre going to fire a bunch of people to keep our productivity exactly where it was before with fewer people'”.
What emerges when assurances are not provided is “shadow” use of unauthorised AI tools. A global survey of 5,200 desk-based knowledge workers by 1Password, an online security firm, finds that 27% of employees have used AI in the shadows in 2025. Another recent study of over 2,000 British workers by Microsoft, a tech giant, reports even higher levels: 71% have used AI in the shadows before, and 51% do so every week. Many firms will see this as a security concern—and it is—but it is also a sign of a failure in leadership that results in a missed opportunity to gain from how the “crowd” is making the technology deliver.

Mr Mollick sets out three clear tasks for leaders. The first is to use AI themselves—“C-level use often matters,” he notes, because it provides both permission and an example. The second is to articulate what the organisation could become. Mr Mollick criticises the “celebrated sets of memos” from firms like Amazon and Shopify that amount to instructions to “use AI or weʼre in trouble.” Such exhortations are counterproductive. “People are not inspired by vague statements about “we should use AI,”” he says. “It has to come down to what is your vision of what the company looks like in an AI-using world.” The vision must be specific enough to guide decisions yet broad enough to accommodate the unexpected applications that emerge from experimentation.
Bosses must also consider what their firms actually value. Mr Mollick argues that many have not thought clearly about the distinction between process and outputs. “If outputs alone matter, itʼs easy to get more productivity in outputs,” he says. AI can generate infinite PowerPoints, Excel spreadsheets or lines of code. But if the goal is merely volume, the organisation risks optimising for the wrong thing. “Is the point of producing this report the interactions weʼre having as we produce the report? Is it the outcome of the report itself?” These questions force leaders to articulate what they want from their employees—a task most firms have avoided by defaulting to traditional metrics like hours worked or tasks completed.
Scale what works, quickly
A culture of experimentation produces insights. But insights scattered across individuals are not yet organisational assets. This is where the third part of Mr Mollickʼs architecture for AI-driven innovation, the “lab”, comes in. He describes it as a firmʼs translation layer that turns the findings of the workers-turned-AI-experimenters into scalable tools and processes.

The lab has two main tasks. First, it needs to deliver speed. When an employee finds a prompt or workflow that markedly improves their work, the lab should validate it and release a usable version to as many as possible “in a day”, says Mr Mollick. This is the opposite of traditional IT deployments, which can take quarters or years. The urgency to scale improvements with AI matters for two reasons. AI models still improve at breakneck pace, so firms that move slowly lose out on compounding gains. Models are also still highly fungible. When a new model comes out, it can often slot into existing workflows without the need for much redesign.
The second task of the lab is to evaluate and tailor AI models. The headline scores that AI builders often talk about—say, a modelʼs performance on legal exams—tell a firm little about whether the same tool can draft the regulatory filings or customer updates that might matter to them. The lab should build an evidence base that shows where specific AI tools add value to real work. This is what brings genuine differentiation in a market where most firms have access to the same few AI models. Such evaluation also provides firms with the data needed to make intelligent decisions on whether and when to buy existing AI solutions or build their own.
Fostering AI-driven innovation
“Provide context” is generally good advice for prompting AI chatbots. It is also what firms should do when adopting AI tools. Some of that context is technical and includes training models on company data and embedding them in real workflows. But another sort, organisational context, matters as much but is too often ignored. What Mr Mollick offers is a general framework that can help firms to build that setting without prescribing what exactly they should do with AI. By working through repeated cycles of trying, learning and spreading what works, firms discover for themselves where AI adds value and where it does not. That capacity—to learn continuously in their own circumstances—is what will separate companies that settle for modest efficiency gains from those that use AI for genuine innovation. AI will soon be ubiquitous, but the organisational context that makes it valuable will not.
Sources
Harvard Business School. Working Paper: Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. September 22, 2023. Available at: https://www.hbs.edu/faculty/Pages/item.aspx?num=64700
Boston Consulting Group. How People Can Create—and Destroy—Value with Generative AI. September 21, 2023. Available at: https://web-assets.bcg.com/pdf-src/prod-live/how-people-create-and-destroy-value-with-gen-ai.pdf
Mack Institute for Innovation Management, The Wharton School. Working Paper: Prompting Diverse Ideas: Increasing AI Idea Variance. January 2024. Available at: https://mackinstitute.wharton.upenn.edu/wp-content/uploads/2024/02/for-web-AI-idea-variance.pdf
1Password. The Access-Trust Gap (2025 Annual Report). October 2025.
Microsoft UK Stories. Rise in ‘Shadow AI’ tools raising security concerns for UK organisations. October 13, 2025. Available at: https://ukstories.microsoft.com/features/rise-in-shadow-ai-tools-raising-security-concerns-for-uk/










