How to Train Your AI ‘Coworker’

“Ben” has been a very productive employee benefits consultant at OneDigital, completing 115,000 assignments regarding client issues since starting the job in February. The prolific output makes sense: Ben is one of the company’s 11 artificial intelligence-powered “coworkers.”

So many human OneDigital consultants rely on Ben as an expert regarding benefits (hence the name) that Shelley McLean, a principal in OneDigital and Ben’s “manager,” views it as much more than a set of tools. Conversing about Ben, she automatically uses male pronouns  and only sometimes follows up with “it.”

“I call him ‘him’—and I question myself on that sometimes,” McLean says. “But it does help me to think about ‘going to Ben for help.’”

Ben’s Origin Story

Ben is the brainchild of Vinay Gidwaney, a former software company owner and now the chief product officer at OneDigital. Co-Founder and Chief Growth Officer Mike Sullivan convinced Gidwaney to join OneDigital and assigned him to the “10X” project to train AI using the company’s databases and presentations on its insurance, financial services and human resources verticals. Sullivan says top-performing employees joined the project so AI would not just perform transactional tasks like writing summaries and drafts, but also collaborate on solutions and brainstorming.  The standout AI project was Ben.

“Ben is becoming the repository of all of that intellectual capital, all that formulation of best practices,” Sullivan says.

The employee chosen to manage Ben was McLean, a three-decade benefits veteran who Sullivan calls a “natural pick.” McLean spent months making sure Ben was correctly analyzing data and coming up with correct, actionable suggestions. However, she spent more time helping her human coworkers overcome fear that AI was “taking jobs.” She wanted her coworkers to realize that AI could handle menial tasks and increase their bandwidth for interpersonal work.

“We’re able to do better work for our clients instead of doing some of the minutiae that takes up a lot of time,” McLean says.

When Ben was first tested in October 2024, McLean says it was a “jaw-drop moment,” showing gr eat potential—as well as great risk, if responses were not grounded in proper training. McLean says several months of “super cautious” training and testing were needed before Ben was officially launched to OneDigital consultants in February 2025.

Best Prompts for Best Results

McLean taught her team to be extremely specific when communicating with Ben. Not only should Ben be told what to find, but it should also be instructed on how to present its findings and reminded to show the work that led to the conclusions.

“Our rule of thumb is you have to ask four questions to get really what you want,” McLean says.

As an example, McLean says Ben could be given a dataset and asked:

• “I specifically want you to outline 10 insights”;
• “I would like you to bullet-point them”;
• “I would like for you to give me quantifiable data to support the recommendations”; and
• “I would like you to sound like a top-tier employee benefit consultant.”

The last point is extremely important, since it makes sure the response’s tone matches company expectations. Ben can also be instructed to tailor the output for specific audiences, such as: “You’re addressing the chief financial officer of a company.”

Sometimes, McLean finds that the best answer from Ben is a prompt. She gives an example of preparing Ben for a client meeting by asking for the best prompt to surface the most quantifiable information from the client’s data to support that meeting. That way, Ben is creating instructions for itself and ensuring the best results.

Most importantly, if Ben is not finding the right answers, employees must share what should be done differently.

“You have to have specific expectations of what you want to get from Ben,” McLean says. “If you don’t get it, you tell ‘him,’ ‘No, that is not what I was looking for. Try again.’”

Picturing Ben

“Ben” (Courtesy of OneDigital)

After working with Ben for a year, not only does McLean have her own view of what “he” looks like, but the company has official portraits of Ben—a bespectacled coworker with kind eyes and an inviting smile—and the 10 other AI coworkers. That way, OneDigital workers are encouraged to think they are interacting with particular people, rather than anonymous chatbots.

“We have men, we have women, we have all different types of coworkers,” McLean says. “People are going to do their best work if they feel that there is a connection, this balance of human and machine.”

AI coworkers like Ben and human collaborators like McLean are likely to become more common. A survey of financial compliance officials—conducted in September and October and set to be released this week by the National Society of Compliance Professionals and the ACA Group, a compliance and governance adviser in financial services—found that 71% of respondents’ firms used AI—a 26-percentage-point increase from last year. More than half of compliance leaders—53%—said AI improved efficiency.

McLean agrees AI is not going away and says companies that incorporate it into their services will have much to gain.

“I would embrace [AI],” she says. “Winning organizations of tomorrow are going to invest in this today.”