Four Perspectives for Leaders Setting the Rules of AI Use

Most organizations now have an AI policy on file. But for boards, general counsels, CHROs and product leaders, many AI policies leave day-to-day use unsettled. Someone still has to decide, application by application, what success looks like, and how to answer for failure. 

Research and commentary published Oct. 26-28 by four Stellant Group speakers suggests that to realize the potential of AI in organizations, companies need more nuanced policies regarding AI governance, written by people with a grounded vision for AI use in the driver’s seat. 

Key Takeaways

  • Beth Simone Noveck and co-authors report that among public workers entering InnovateUS's foundational generative AI course, 51% had never tried the technology, and 19% of non-users didn't know whether they were allowed to use it at work. Public agencies need to do more to help workers understand and use AI, at all levels. 
  • Faisal Hoque proposes recency, proximity to AI and stake as tests for new board directors, rather than experience alone; in 2026, current and former CEOs accounted for 37% of appointments, and retired executives comprised 54% of new directors. 
  • Vasant Dhar argues that regulation should target how AI is used, starting with how much variance each use case can tolerate.
  • Kathy Pham offers product teams four principles for building responsibly now that AI makes building fast and cheap.

Beth Simone Noveck: Publishing a Policy Is the First Step to Realizing AI-Driven Gains

In "What we learned from 250,000 public workers about AI adoption," published in Fast Company on Oct. 26, Beth Simone Noveck, a professor at Northeastern University, and co-authors Gregory A. Porumbescu and Lee Rainie report survey results from InnovateUS, the nonprofit learning initiative Noveck founded, which has trained more than 250,000 government workers in AI and digital skills. The findings draw on 106,300 precourse responses and 67,002 postcourse responses collected between June 2024 and May 2026. Among workers entering the foundational generative AI course, 51% had never tried the technology. Another 14% had experimented personally but weren't using it at work.

The results indicate the distance between issuing a tool and building employees’ capacity to use it responsibly. Among respondents who weren't using generative AI, 28% said they didn't think it applied to their work; 23% didn't know how to use it; and 19% didn't know whether they were allowed to use it at work. Managers showed similar uncertainty: among managers who weren't using AI, 24% didn't know how and 22% didn't know whether it was permitted. Importantly, after workers received training, they reported both higher levels of confidence and vision for how AI could be used in their roles. 

For Noveck, Porumbescu and Rainie, however, simply equipping government employees to work faster won’t necessarily lead to better public services for people. "Providing access to a tool or publishing a policy leaves substantial work undone," Noveck and her co-authors write. "Employees need relevant examples, opportunities to practice, and managers prepared to support responsible use." They argue that success should be measured by what changes for residents: better services, more opportunity to be heard and participate in civic life, and solutions to the problems they care about. In that way, AI could be a substantial boon to rebuilding public trust. 

Faisal Hoque: In Hiring for Experience Alone, Boards Risk Leaving AI-Driven Gains on the Table

Erring on experience in hiring decisions could pose risks to institutional resilience as AI disrupts how companies deliver value and how work gets done, argues Faisal Hoque, founder of SHADOKA and NextChapter, in his recent piece for Fast Company, "Are boards preparing to fight the last war?" (Oct. 26). 

"Faced with intense turbulence and uncertainty, boards appointing new directors are favoring experience,” Hoque writes. “The logic is understandable: In tough times, you err on the side of picking people who have seen it all before." But experience, he argues, is by definition a record of past mistakes, so it protects a board against the previous era's errors and offers no such protection against those of the next.

Board leaders need the vision that accompanies familiarity with new tools to help shepherd companies through AI-driven change. However, fewer than 3% of directors disclose AI expertise, even as more than four in five S&P 500 companies now disclose AI as a risk. Meanwhile, current and former CEOs accounted for 37% of new director appointments in 2026, the highest share since 2012, and retired executives made up 54% of new directors.

Hoque proposes three tests for rebalancing boards without sacrificing experience: recency (are directors working in today's conditions?), proximity (are they close to where AI is actually being used?) and stake (will they live with the consequences?). He suggests allocating one or two seats to directors who pass those tests and creating an advisory panel of the people closest to how AI is used in the business, with a standing quarterly agenda item and a direct line to the full board. And he suggests requiring a “2040 Test” for any decision resulting in impacts lasting beyond a decade: a short memo written by the directors and advisors. The goal is to empower those board members who will actively live in the commercial landscape that those choices create over the next 15+ years. 

Vasant Dhar: Why AI Should Be Regulated by Use Case

As constructed today, AI is not aligned with human interests, argues Vasant Dhar and João Sedoc in this week’s "The Solution to AI's Alignment Problem is Precise Guardrails on How We Use It," (Tech Policy Press on Oct. 28). According to the authors, modern AI's complexity makes its internal reasoning very difficult to understand and control, and thus writing governance requirements into models themselves is almost impossible. The better approach is to instead “establish expectations and put guardrails around the use and consequences of AI-based systems,” writes Dhar and Sedoc, through “a combination of government legislation and company self-monitoring.”

Their experience building DBOT, a multi-agent system designed to mimic the valuation thinking of their NYU Stern colleague Aswath Damodaran, illustrates the problem: run DBOT on the same stock multiple times with no new information, and it can produce different valuations and narratives. "Random variances are generally undesirable in a machine, which is typically expected to yield the same output each time in response to unchanged inputs," the authors write. "Although variance is an inherent property of all machine learning systems, generative AI has made it much harder to identify its sources and to debug such systems."

Whether a given level of variance is acceptable depends on the use case: in high-stakes areas such as finance and mental health, stakeholders would not accept recommendations that vary widely for the same case, while a creative task might tolerate more. 

Dhar and Sedoc point to finance as a blueprint. In the 1990s, broker-dealers and banks were fined for failing their fiduciary responsibility to clients, and financial institutions went on to build compliance departments. The authors expect AI companies to follow as consumer protection rules emerge, and they call on lawmakers to understand the inherent variance in AI models, and design harnesses that control specific uses; a financial application, for example, might be required to disclose its expected variance. 

Kathy Pham: How Product Teams Can Build Trust When Building Is Cheap

Product teams have a responsibility to social outcomes too, says Kathy Pham, a Harvard Kennedy School faculty member and Shorenstein senior fellow (Oct. 26, "Why you should care about trust when building is this easy”, Mind the Product Chicago). Now that speed has changed the stakes of product development, Pham calls on product teams not just to build faster, but to do so with the good of people at the center of every decision: "Any tooling that makes building the right thing cheaper makes building the wrong thing cheaper too.”

The conference report summarized her four pillars for product teams:

  1. Solve the right problem: What are users actually asking for?
  2. Make evals the new acceptance criteria: use evaluations with proper tools and human domain expertise as guardrails.
  3. Treat trust as a feature: consider how the product affects everything around it.
  4. Orchestrate the shift: create new product and software lifecycles, redefine roles and rethink how teams connect.

Pham also encouraged teams to build a "phone a friend" network: know the product, users and domain well enough that anyone can call you for a fast, trusted answer when you need a reality check. The report closes on her final point: the hard question for product teams is whether they should build something at all, but whether it is right for the communities it serves.

Leaders Need to Know the Tech and the People They Serve

The four pieces converge on the same set of practical questions: How can AI work for us, in ways that help organizations thrive and keep communities safe? Who decides, case by case, if an AI application is working as intended? How do we hold AI accountable when it isn’t? And how do we do this work as the technology continues to evolve at an unprecedented speed? 

Policies alone won’t answer these questions, but nor can people create efficacious AI policies without a specific understanding of the technology itself. For government leaders, that means treating workforce learning as an investment and measuring what changes for the people agencies serve. For boards, it means rebalancing composition to include directors who will live with the consequences of today's infrastructure and talent decisions. For product teams, it means building trust into the development cycle and knowing whom to call for a reality check on a product design decision.

To keep pace with the speed of change, leaders need to stay focused on who their organizations serve and the affordances of today’s technology for their work. Holding both concerns simultaneously will be the key to driving real value from AI across teams and communities, while bolstering public trust. 

Beth Simone Noveck, Faisal Hoque, Vasant Dhar and Kathy Pham are each exclusively represented by Stellant Group for speaking engagements including keynotes, fireside chats and panel discussions with leadership teams thinking through AI governance and the accountability structures that responsible AI use requires. For our consulting or advisory offerings, please reach out through Stellant Group for a custom proposal.