Andrew McAfee: 4 steps to fast-tracking generative AI

The greatest mistake organizations can make with generative AI: sitting on the sidelines. 

That’s according to Andrew McAfee, principal research scientist at the MIT Sloan School of Management and co-founder of generative AI company Workhelix. “When a technology this powerful comes along where you have to learn by doing, finding reasons not to do it is a pretty big error,” he says.

Despite the mass embrace of the technology in its first year of release—3 in 4 surveyed tech companies are already using it in some way, a recent VentureBeat report revealed—most organizations remain cautious about mass adoption. Two-thirds of enterprise risk executives surveyed by Gartner consider gen AI a top emerging risk. Among their biggest concerns: exposing intellectual property through publicly available generative AI models, revealing the personal data of users to third-party vendors or service providers, and securing the AI itself from criminal hackers.

McAfee counters that such risks are manageable.

“These risks are things you have to worry about with any other large-scale database technology project—but they’re not terrifying, and you have a great deal to gain,” says McAfee, who recently published “The Geek Way: The Radical Mindset that Drives Extraordinary Results.” The potential benefits of generative AI, of course, are huge. They include worker productivity gains, improved quality and consistency across dozens of enterprise functions, improved personalization for customers and employees, and cost savings.

“There’s a level of planning that is necessary for generative AI,” says McAfee, “but there’s also this idea of a minimum viable plan: Figure out what your strategy is, what the sequence of efforts is that you want to undertake, and then dive in.”

To identify opportunities and determine the potential ROI for generative AI applications, McAfee advises that CIOs and other senior leaders consider these four basic steps.

1. Inventory existing knowledge-work jobs

Generative AI is useful for almost all knowledge worker roles in the enterprise, including, for instance, customer service agents, programmers, marketers, and lawyers, McAfee says. The technology is best-suited for language-based tasks within those jobs. 

“The way to get started,” says McAfee, “is to think about the different jobs that are done in your organization and then get a rough idea about what percentage of the tasks for those jobs are amenable to generative AI. A good place to start is with the jobs where a lot of the tasks can have their productivity improved substantially.”

For instance, if what you’re creating follows a well-established template, such as an HR employee review, why start from scratch every time? says McAfee. “Why not let generative AI take the first crack at it, edit it, fill in the blanks, and then let the human worker review it?” he says.

Other writing and language-based tasks include financial report summarization, email and newsletter writing, and software coding. “A good ground rule is the more tasks a worker does with language, the more generative AI can assist that person,” freeing up experienced workers for high-level strategic work and lowering the bar for newcomers, he says. “That’s pretty wild because up until recently, technology was pretty lousy at human language. Now, it’s quite good.”

2. Consider off-the-shelf AI

After identifying roles that lend themselves to gen AI applications, consider whether the individual would benefit from having a “competent but naive gen AI assistant”—akin to a worker who excels at programming, writing, or preparing and summarizing data but doesn’t know anything about the organization, McAfee says. This type of AI assistant can be delivered through a pre-built, off-the-shelf AI solution. 

“Someone who is a new coder to a company can start to be productive pretty easily,” says McAfee. When it comes to writing a piece of code to test software or debug errors, the coder could hand that off to a digital assistant, which could do it well and quickly, he adds.

Publicly available large language models (LLMs) can act as this sort of naive assistant, writing code or creating a project management plan. It won’t know the organization’s systems integration needs or what projects other workers are working on—but that information largely isn’t necessary for performing the task at hand.

3. Consider bespoke AI

Some knowledge-work jobs that lend themselves to gen AI require more experienced digital assistants. A customer service agent needs the kind of institutional knowledge and case resolution expertise that only a veteran can provide.

“If you’re trying to troubleshoot a problem for a customer, you might want an experienced assistant—someone like a mentor or coach to whisper in your ear how to solve this customer’s problem,” McAfee says.

Read also: AI is posed to solve the tech overload problem

In these instances, an off-the-shelf generative AI system isn’t enough; organizations will need to combine it with another system trained on internal data to achieve the output of the more experienced assistant, says McAfee. These custom AIs require a more detailed development process in which organizations train the models on specific data sets in order to tailor the AI to the company’s needs.

Some of this data may include customer information, such as demographics and buying behavior, in order to personalize recommendations and customer support; sentiment analysis from customer feedback to proactively address concerns or capitalize on positive feedback; industry-specific knowledge, such as trends and jargon, to improve the accuracy of responses; and product or service data to provide customers with recommendations. 

4. Prioritize potential projects

After identifying the roles best-suited for naive or experienced digital assistants, leaders must identify and prioritize the most promising gen AI projects, McAfee says.

“Think about where the most productivity benefit is to be found and the percentage of those tasks that are amenable to generative AI,” he says. Some 75% of the value that generative AI use cases could deliver falls across four areas, according to McKinsey research: customer operations, marketing and sales, engineering, and R&D. Organizations that apply the new tools to customer operations, for example, stand to realize the most significant productivity gains: 30%-45% of current functional costs.

Many organizations, from consumer packaging and manufacturing to healthcare, have been successful in identifying gen AI applications through these four steps, McAfee says. 

“Success means having a clearer idea of where the big potential benefits are to be found,” he adds. “I see over and over again that our customers say that now they have an idea of where to go next. Maybe it’s not going after opportunity #1 because of other priorities as a company, but with their management team, they can pick and choose among those—and that clarity is helpful.”

We want to hear from you! Please send us your feedback, and get informed about exciting updates from The Works. Drop us a line: theworks@freshworks.com.