The race to integrate artificial intelligence into business operations is accelerating, but the path forward is fraught with both opportunity and risk. Industry leaders from PwC and NBCUniversal recently shared their approaches at a conference hosted by Section, a consultancy led by NYU professor Scott Galloway. Their insights reveal that the key to successful AI agent deployment lies not in choosing between speed and caution, but in skillfully combining both.
As organizations grapple with the promise of AI agents—autonomous systems that can perform tasks, make decisions, and interact with users—the central question becomes: How can companies move fast without falling into the traps of costly errors or ethical missteps? The answer, according to these experts, involves a disciplined yet agile framework rooted in human oversight, experimentation, process improvement, and governance.
1. The human is the loop
One of the most critical insights comes from Scott Likens, global chief AI engineer at PwC. He reframes the common phrase "human in the loop" to something more profound: "the human is the loop." This shift in thinking emphasizes that AI should not operate independently of human judgment and context. Instead, every AI action should be instigated and guided by human intent, especially in high-stakes environments.
Lasherelle Morgan, senior vice president of AI innovation and acceleration at NBCUniversal, echoes this sentiment. She advises starting with the end user and their pain points. "Don't just bring in an AI tool. Ask, 'What are you struggling with?' 'What are you spending five hours of your day on?'" This user-centric approach ensures that AI solutions address real problems rather than creating new complexities.
The practical implication is that organizations must resist the temptation to hand over the keys to AI agents without a clear human-led framework. Human oversight is not a bottleneck but a necessary enabler of safe and effective automation.
2. Experimentation is important
Likens advocates for rapid experimentation cycles to discover where AI can deliver the most value. At PwC, teams run AI-driven experiments in one-day or five-day cycles, allowing for fast feedback and iteration. This approach contrasts with traditional long-term planning, which can delay learning and adaptation.
However, experimentation requires a cultural shift, particularly among mid-level managers who are accustomed to longer planning horizons. "Top executives and board members may be on board, as are new employees that have already been there. It's that frozen middle, those experts and managers who don't want to change their ways," Likens notes. Overcoming this resistance is a human challenge, not a technical one.
Moreover, Likens warns against an overly narrow focus on cost savings. "All this talk of tokens just started a couple of months ago, and now all of a sudden there is a cost focus with AI. That's the wrong way to look at it." Instead, companies should prioritize value creation through rapid experimentation and a willingness to explore new possibilities beyond small efficiency gains.
3. Blow up a bad process
AI can amplify both good and bad processes. Morgan emphasizes the importance of clean data and well-defined workflows before introducing AI. "You have to have clean data, and a workflow that is clean from start to finish. You need to literally get a pen and paper and write out the process." She warns that "one thing AI is really good at is blowing up a bad process," meaning that flawed workflows will become even more problematic when automated.
The recommendation is to start with repetitive, hated tasks that are ideal for automation. Identify data-rich, repeatable processes that can be easily improved by AI. This not only yields quick wins but also builds organizational confidence.
At PwC, the foundation for AI success was laid well before the recent surge in generative AI. The company addressed data issues in regulated areas such as accounting and auditing, establishing a reliable data infrastructure. Likens explains that the challenge is extracting tacit knowledge—the informal, experience-based insights that often reside in people's heads. AI agents can help capture and systematize this knowledge, but only if the underlying data is trustworthy.
PwC's goal is to create a "tacit knowledge collection" system that tracks what AI agents do and feeds into a continuous learning loop. This requires a focus on architecture first, ensuring that the system is safe, scalable, and provides appropriate data access.
4. Governance and guardrails
Governance is not a barrier to innovation but a framework that enables responsible scaling. At NBCUniversal, the approach is risk-based. Morgan describes the concept of "blast radius"—the potential impact of an AI agent's actions. Low-risk use cases, such as an agent that schedules lunch appointments, need minimal oversight. But high-risk applications, such as automatically sending messages to consumers, require human approval and strict guardrails.
NBCUniversal uses intake forms to track and measure the potential impact of AI initiatives. This governance process helps teams identify risks early and apply appropriate controls. At PwC, AI responsibility is centralized among a small group of deep AI engineers—about 1% of the organization—who set standards and create trusted "chassis" (platforms). Another 10% are hands-on builders distributed across business units, combining technical skills with domain knowledge.
This two-tier structure balances centralized control with distributed innovation, ensuring that AI development remains both safe and agile.
In summary, deploying AI agents requires a deliberate strategy that blends speed with caution. By putting humans at the center, running rapid experiments, cleaning up processes, and implementing risk-based governance, organizations can capture the value of AI while mitigating its risks. The experiences of PwC and NBCUniversal offer a blueprint for moving fast—without breaking things.
Source:ZDNET News