Getting started with AI agents can feel like a race against time. The pressure to adopt generative AI and autonomous systems is immense, yet so are the risks. How do organizations move fast without crashing? Two enterprise leaders offer a playbook that combines speed with caution, drawing on firsthand experience from large-scale deployments.

The challenge: speed vs. safety

AI agents—software programs that can perform tasks autonomously on behalf of users—are becoming central to digital transformation strategies. But their deployment raises fundamental questions about control, reliability, and accountability. Should companies move fast and furious, jumping on every new model? Or should they proceed with extreme caution, ensuring every step is measured? The answer, as with many complex technology shifts, is both.

1. The human is the loop

The first principle: never hand over the keys entirely. Scott Likens, global chief AI engineer at PwC, puts it bluntly: “Stop being the human in the loop. The human is the loop.” This means AI agents should augment human decision-making, not replace it. Every action initiated by an agent should be instigated and overseen by a person. Lasherelle Morgan, senior vice president of AI innovation and acceleration for NBCUniversal, agrees: start with the end user, not the technology. “Start with easily repeatable processes and data,” she advises. “Ask users: what are you struggling with? What are you spending five hours a day on?” By focusing on pain points, organizations ensure that AI solves real problems rather than creating new ones.

2. Experimentation is important

Moving fast doesn't mean skipping discipline. At PwC, AI experiments run in one-day or five-day cycles. This rapid prototyping allows teams to test hypotheses quickly and iterate based on feedback. Likens warns against a narrow cost focus: “All this talk of tokens just started a couple of months ago, and now there is a cost focus. That’s the wrong way to look at it. Experimentation is really important and so easy nowadays, and you get feedback fast.” However, this approach requires a cultural shift, especially among mid-level managers. “Many are not used to one- or two-week cycles,” Likens notes. “Top executives and new employees may be on board, but the frozen middle—experts and managers who don’t want to change—is a human challenge.” Overcoming that inertia is critical for scaling AI.

3. Blow up a bad process

AI can work miracles, but not on broken foundations. Morgan emphasizes that data must be clean and workflows well-defined. “You have to literally get a pen and paper and write out the process,” she says. “Show me who is the owner.” One thing AI is good at is exposing flaws. “It is really good at blowing up a bad process.” Before deploying AI, identify repetitive tasks that people hate doing—these are ideal starting points. PwC’s approach is to build a solid data architecture first, even before AI models came into the picture. Likens explains that the real challenge is extracting tacit knowledge: “How do you extract knowledge that usually sits in people’s heads?” Their solution focuses on telemetry and agent behavior, feeding into a knowledge base that makes AI both safe and scalable.

4. Governance and guardrails

Not all AI use cases carry the same risk. Morgan advises a risk-based governance model: “With a use case as simple as an agent that presets lunch on my calendar, that’s low risk—no human in the loop needed. But if it automatically sends messages to consumers, that’s a bigger deal.” NBCUniversal uses intake forms to track and measure potential impact. This allows them to scale guardrails appropriately. At PwC, AI responsibility is centralized among deep AI engineers who set standards and build the trusted chassis. Then, a distributed layer of builders across the business applies those standards to specific industries and functions. This structure ensures consistency without stifling innovation.

Both leaders agree that the path to successful AI adoption is neither purely fast nor purely cautious—it’s a dance between the two. Start with a clear understanding of the user’s pain point. Experiment aggressively but within a framework of clean data and sound architecture. Apply governance proportional to risk. And above all, keep humans at the center of every loop. By following these four principles, enterprises can deploy AI agents with confidence—moving fast, but not so furious that they lose control.

Source: ZDNET News