Stop Running So Many AI Pilots

In late 2023, when the management team at the consumer packaged goods company Reckitt considered adopting gen AI, potential use cases spanned the business—from drafting presentations to delivering customer support to optimizing procurement contracts. Many of the use cases guaranteed time savings and an immediate return on investment, but they applied to disparate tasks. Executives at Reckitt were pleased with the time gen AI could save, but they knew the effort wouldn’t transform the company’s strategy or create a meaningful advantage. They were hoping for something more dramatic, not just marginal efficiency improvements.

So instead of green-lighting every AI project that would yield a positive ROI, they decided to focus on a single domain: marketing. Here gen AI could be used across many interlinked tasks, including insight generation, content generation, and new product development. The tasks were related to one another, using information pulled from the same data, customers, and market research. The results from one task connected to others. For instance, real-time consumer insights lead to superior product innovation and better segmentation. Better and faster product innovation helps companies bring relevant products to market, thereby strengthening loyalty, retention, and advocacy. Focusing their efforts just in marketing, executives believed, would help the company improve at scale.

Reckitt had existing strengths in marketing, such as rich customer data and a team that was proficient with advanced technologies, including predictive AI. Adopting gen AI with a focus on improving marketing operations helped Reckitt discover new ways that the department could operate. It forced the marketing team to dramatically rethink its entire approach, recasting many processes as the team learned how gen AI could improve the whole marketing operation, not just the odd task or two.

Less than two years later, Reckitt credits its deliberate adoption of gen AI with helping it generate product concepts as much as 60% faster than it used to. It also credits gen AI with making brand and marketing communication processes more efficient by 30% or more (depending on the process). Those gains wouldn’t have been possible if Reckitt hadn’t focused its AI experimentation and investment efforts on a set of interconnected functions or tasks within a single domain.

Reckitt’s experience illustrates a vital lesson: To obtain a lasting competitive advantage, companies must avoid two common and related gen-AI implementation mistakes. The first is spreading your efforts and resources across one-off use cases throughout the company. The second is focusing solely on immediate ROI. This approach—having many unrelated use cases, each justified by the strong promise of immediate ROI—is often described as a “shallow and broad” deployment. Many executives we spoke to said they believe they should deploy gen AI as widely as possible to see what works. While this technique will generate some immediate cost efficiencies, it’s unlikely to deliver a competitive advantage, and companies that use it aren’t as successful as they could be.

Here’s why. Let’s say a firm automates a few disparate processes in accounting, some in the supply chain, some in marketing, some in finance, and some in HR. Two things happen: First, executives have little motivation to completely revamp functions when only 5% to 10% of the tasks are using gen AI. Second, running use cases with no focus on a larger objective often leads to initiative fatigue. And when the use cases provide little competitive advantage or are easily copied by competitors, employees can grow skeptical or disillusioned. (At Reckitt, by contrast, the larger objective was to drive top-line growth by introducing new products to market faster.)

When it comes to deployment of AI, firms should follow Reckitt’s lead—what we call the “deep and narrow” approach. It’s an unusual strategy: Only 4% of companies take a focused approach on a few priorities where they transform deeply, according to a 2024 BCG survey of 1,000 CXOs and senior executives. But those that do achieve twice the ROI over time.

We have decades of experience as practitioners, educators, and consultants for global brands on marketing, strategy, and technology. And since the introduction of gen AI to the market in 2022, we’ve observed that the deep and narrow approach has many advantages. It requires you to fundamentally rethink how work is done within a single part of your business. It leads to better productivity by leveraging your competitive strengths. And it encourages the discovery of new ways of working that are unique to your company’s culture, making it difficult for rivals to copy. Once you master gen AI in one business area, then you can apply what you’ve learned to others.

When we first proposed the idea of deep and narrow deployments, many executives worried that it seemed counterintuitive—and more risky than the shallow and broad approach. After all, shallow and broad deployments aren’t wrong: They do save time and money. And executives should be comfortable conducting a few shallow and broad experiments, especially as they begin to test the organization’s willingness and readiness to adopt AI. But they need to understand that one-off, disconnected use cases are not transformative deployments. The executives were also uneasy about the significant investment required to rebuild processes and feared deviating from conventional wisdom, which says you need to run hundreds of AI projects to succeed. Additionally, many had failed to conduct an honest and intense appraisal of their firm’s core competencies and data maturity—a necessary step to decide where to go deep—because they considered that process too difficult and time-consuming. Over time, however, many of the skeptics came to see why deep and narrow makes sense.

In this article we’ll explore why so many companies mismanage their adoption of gen AI. Then we’ll provide a guide for successfully going deep and narrow for your own AI deployments, using examples from IKEA, L’Oréal, and others.

Why Go Deep and Narrow?

Leaders tend to demand concrete ROI projections for every use case. Only after immediate ROI is demonstrated, they told us, do they think they should deploy gen AI elsewhere. But that approach ignores the reality of how AI adoption really works: ROI accumulates at a different speed and magnitude for each use case. At firms that we believe have deployed gen AI effectively, the impact may take some time to materialize, and progress shouldn’t be measured by short-term ROI alone.

Compared with shallow and broad, a deep and narrow implementation takes a lot of time and effort. It’s exceptionally difficult to do it simultaneously across multiple domains because it requires extensive change management. You’ll be introducing technologies that are new to your employees. You’ll be changing work processes. Your team structures and ways of working will have to change, too. Of course, introducing any new technology will require learning and some adaptation of work processes, but going deep requires wholesale changes and a reimagining of how work is done.

Deep and narrow implementations can focus on a single function or on end-to-end processes. L’Oréal, the world’s largest cosmetics firm, has deployed AI extensively within its marketing function, as Reckitt did. Acme Bank (not its real name), by contrast, focused on transforming its end-to-end mortgage process. Its original process was remarkably cumbersome: First, a mortgage adviser collected borrower documents and manually entered the data into the loan origination system. Next a loan officer verified the documents and used additional documents (such as title searches and flood certifications) to aid in the evaluation. Finally, an underwriter assessed the borrower’s creditworthiness and set the terms of the loan. Acme now uses gen AI to automate the processing of all mortgage paperwork. The bank also built a conversational gen AI interface for underwriters. They can converse with the system and ask it to perform tasks, such as running predictive AI models and writing an explanation for an applicant’s approval or rejection. This easy interface enabled loan officers to probe the model’s decisions, and they found that providing better information on a customer’s context improved the outcomes for both customers and the bank. As a result, Acme Bank improved the quality of its decisions, its speed, and its responsiveness to customer concerns.

How to Begin

Before you implement gen AI, you need to identify where it would be most strategic because ultimately you want gen AI to strengthen your existing competitive advantages. To do that, we recommend a four-step process.

Step 1: Identify the most promising opportunity. Gen AI is a general-purpose technology with a spectrum of uses. Office productivity tasks, such as summarizing meetings and creating presentations, fall on the routine end of the spectrum. At the other end are value generation activities, such as using gen AI to enable new business models or, when it is infused into products and services, to overhaul customer experiences. In the middle of the spectrum is domain reinvention—the end-to-end reimagination of work processes or functions.

Office productivity is about improving the efficiency of handling mundane tasks. It is not likely to yield a competitive advantage. Enabling new business models using AI is an exciting opportunity but may be difficult for companies selling products in which AI cannot be infused easily (say, toothpaste). Domain reinvention—which covers functions and processes—is an area of opportunity for almost any firm.

Let’s return to the example of Acme Bank. It is one of the largest lenders in the region, and mortgage loans are a big part of its portfolio. Mortgage financing has big margins even in low-interest-rate environments, which makes mortgages a strategically important product for the bank. Acme used mortgage loans to improve customer relationships and cross-sell other products. The team mapped the essential steps in the process, evaluated where it made sense to do things differently, and eliminated redundancies. The team developed a more-effective process, and gen AI made it even faster and easier.

Step 2: Identify areas of lasting advantage. Whenever you deploy AI deeply, your goal should be either to protect or enhance an existing competitive strength or to create a new advantage that is hard to copy.

L’Oréal used gen AI to create a new competitive advantage. The €43 billion French cosmetics retailer recently combined its cosmetics expertise with powerful gen AI tools to reinvent customer interactions. Nicolas Hieronimus, the company’s CEO, said L’Oréal shifted its vision from beauty for all to beauty for each by putting personalization at the center of its mission. How? It focused its gen AI efforts on the consumer journey—deep and narrow. Its Beauty Genius chatbot conducts an in-depth diagnosis of each user’s skin and then gives customers expert analysis, product recommendations, and personalized routines. It can do this because L’Oréal has been building its augmented reality (AR) capabilities steadily since 2015 and because it can access the company’s advanced research on skin biology. L’Oréal focused its gen AI program on its domain expertise and trove of customer data. It reinvented the consumer journey when it introduced Beauty Genius, which makes it easy for consumers to understand their skin (diagnosis), get advice, select the right products, and share with friends. Moreover, every time a customer interacts with the chatbot, L’Oréal gets more valuable data.

The project has been a success. Beauty Genius has conducted more than 400,000 conversations in the United States in its first six months. L’Oréal is integrating Beauty Genius into WhatsApp, which we expect to lead to a surge in usage. L’Oréal’s strategy is to provide a slew of Beauty Tech services, of which Beauty Genius is a key component. These services—such as virtual try-ons and Hapta, a computerized makeup applicator—leverage AI extensively and have been used more than 110 million times. In Southeast Asia, the Middle East, and North Africa, Beauty Tech has doubled the conversion rate of its 50 million users. With success like that, many copycats are bound to emerge, but they can’t offer the vast cosmetics research that L’Oréal does.

IKEA is taking a slightly different tack: It’s using gen AI to amplify its strengths. The company, renowned for its affordable furniture, also offers interior design services. But for customers, price is often a barrier because highly paid designers are required to provide the services. Now IKEA is looking to use gen AI to make design services more affordable. Based on early results, Francesco Marzoni, IKEA’s chief data and analytics officer, projects that IKEA’s professional designers will be able to use gen AI to complete design projects in 10 minutes and then use their expertise to enhance the designs. That’s a big change from using consultants whose fees are typically $99 per room and take a few days to deliver designs.

This new way of designing projects enhances IKEA’s existing competitive strength in design. It already has the world’s largest community of interior designers, and it carries out far more interior design projects than do its competitors. Now it can feed its gen AI models in real time with high-quality data from thousands of projects. These two elements, expertise and scale, will enable its gen AI system to learn faster and better than others. It will be hard for other furniture retailers to provide a comparable service at the same quality and price.

Step 3: Choose the right sequence. Gen AI can boost profits by enabling cost efficiencies and growth. Should you prioritize efficiency, growth, or a combination of the two? When you’re going deep and narrow, starting with cost-efficiency gains within a single domain is often the smartest option, because you can typically reduce your costs more quickly than you can achieve top-line growth.

That’s what Reckitt did. The company identified five gen AI pilots to demonstrate immediate marketing efficiencies. One experiment, for example, showed that gen AI could develop media campaign analyses in hours rather than days. Early wins like that helped gain support from executives. Once leaders were convinced, Reckitt went deep and narrow within marketing to drive both efficiency and growth.

Next Reckitt took a systematic approach to deployment. It categorized all the tasks performed by its more than 2,000 global marketers. This was a painstaking activity involving both surveys and one-on-one interviews. That research identified 300 separate tasks. Marketers provided estimates of how much time each task took. The company then plotted each task along two axes: the degree of possible automation (based on type of gen AI applicable) and the opportunity size for the company (based on time spent on the task). This analysis helped the team target approximately 100 tasks that would be most suitable for gen AI use. Because so much of the team’s work would now involve AI, Reckitt’s leaders recognized that marketers would have to do a wholesale reconsideration of the way the department worked. As gen AI picked up routine tasks, human marketers took on more-demanding work, including tasks that had previously been performed by their managers. Reckitt focused its gen AI deployment on efficiency because those gains would make an immediate contribution to the bottom line. Once executives saw that, they green-lighted the use of gen AI for growth objectives, which normally take longer to achieve. Marketers used gen AI to achieve corporate growth by translating consumer insights into better and faster product innovation and speedier time to market.

Because the deep and narrow approach requires more focus on transforming processes and has a bigger impact on employees’ day-to-day work, it is more managerially intensive than shallow and broad. Change management is 70% of the challenge. Perfecting data and using gen AI effectively are only 20% and 10%, respectively. As companies try to build on their early successes and move beyond the initial department where they went deep, they must be ready for a full-scale change-management initiative.

Step 4: Monitor the competitive landscape. Just as you are using gen AI to strengthen your competitive positioning, so too are your competitors. Ask yourself: Can our top competitor replicate a valuable strength of ours by using gen AI? The other company does not have to be as good as you at the strength. To cause trouble, it must provide only a “good enough” solution that is simpler and cheaper. To avoid losing your advantage, you must deploy gen AI in ways that will increase the gap between your firm and your competitors, as L’Oréal’s Beauty Genius does. When you prioritize gen AI within your operations, it will reveal opportunities to change how you work. Take them.

. . .

We’re still in the early days of gen AI and agentic AI, but already a clear pattern has emerged. Most companies deploy AI widely across the organization; they use the shallow and broad approach and hope that a few pilots will deliver meaningful value. Many of these efforts do provide ROI, but collectively, their impact is often limited and temporary because the shallow and broad deployments are easily replicated.

Going deep and narrow reinforces your firm’s strengths in a deliberate manner. IKEA, L’Oréal, Reckitt, and others used the deep and narrow approach to improve functions or complex end-to-end processes. They asked themselves tough questions about their AI goals. In some cases, it led them to create new lines of business. And the results have been worth the effort. One BCG study found that firms using the deep and narrow approach have double the ROI of firms using shallow and broad deployments. We predict the gap will continue to widen.

Copyright 2025 Harvard Business School Publishing Corporation. Distributed by The New York Times Syndicate.

Explore AAPL Membership benefits.