Five Common Sense Elements for AI Safety in CX
A confluence of forces, from pending trillion-dollar IPOs, recent jailbreak agentic artificial intelligence announcements, ongoing data center fears, and rising tension in geopolitics, have created quite a conundrum for AI companies in the West. Calls by Anthropic's CEO, Dario Amodei, for an industrywide slowing of AI research are in direct conflict with the AI race with China, stoking national security fears. David Sacks, the former White House AI czar, called out this doomerism and pointed out that tech leaders should step aside if they cannot control the AI systems they are building.
Customer-Facing CX Flows Should Have Five Elements for AI Guardrails by Design
Most agentic AI use cases focus on automating tedious work, improving customer engagement, and delivering on mass personalization at scale. The proliferation of agents in a cross-platform, multi-agentic AI world will require improved guardrails.
For example, flows such as automated data capture, CRM enrichment, agentic sales development, sales outreach, pre-meeting deal briefs, intelligent lead scoring, automated customer support and resolution, pipeline management, IT and HR support requests, product discovery, commerce purchases, or customer feedback and sentiment analysis bear responsibility for customer-facing outcomes. With a greater volume of interactions and frequency of interactions, five minimum guardrails should be incorporated by design:
- Use the kill switch requirement. Kill switches stop AI from making decisions, taking actions, and automating decisions. The kill switch should isolate the system and cut the AI off from other systems, the outside world, and data. Security access should be revoked, and security tokens and API keys should be blocked. These safeguards should operate independent of the AI’s own reasoning software, operate in physical and digital forms, and not be overridden by AI. The recently introduced AI Kill Switch Act, if enacted, would provide a significant foundation for AI safety.
- Apply Asimov's Three Laws of Robotics to AI. The famous American writer and professor of biochemistry, Isaac Asimov (1920-1992), wrote and edited more than 500 books. In the Foundation series, he introduced the Three Laws of Robotics, which apply to AI. The first law—Robots cannot harm humans or allow them to come to harm through inaction; the second law—Robots must obey human orders, unless they conflict with the first law; the third law—Robots must protect themselves, provided it does not conflict with the first or second law; and the zeroeth law—a robot may not harm humanity, or, by inaction, allow humanity to come to harm.
- Ensure security by design. This approach to technology development builds security controls from the ground up rather than adding them post production. Use least privilege, which gives users and agents the minimum access needed; defense in depth, which applies multiple layers of security controls; fail securely, allowing systems to revert to a safe state when errors or failures occur instead of opening up access; and secure defaults, where products are shipped safe out of the box with no additional costs.
- Commit to energy ratepayer protection and (water). The Ratepayer Protection Act (H.R. 9340) and related Ratepayer Protection Pledge stops large-load data centers from driving up rates to consumers by consuming energy supplies and passing on infrastructure costs. In fact, the act also allows AI companies to build their own electrical generation capacity. Similar requirements on water consumption will probably follow.
- Submit to national security review. A combination of mandatory and voluntary review frameworks by different agencies should be in place to address national security, cybersecurity, consumer protection, and critical infrastructure risks.
The Bottom Line: Proactively Design and Build CX Agents for AI Governance
CX leaders have focused on compliant execution by automating the deterministic processes. Agents that enforce privacy laws or regulatory compliance will include prescriptive and very clear parameters. When well-defined deterministic processes are automated, organizations can subsequently focus on improving orchestration excellence through greater precision in probabilistic processes. As precision increases, confidence levels must anticipate new CX scenarios.
This type of feedback loops and signal intelligence will improve the confidence of data to decisions in CX. In each scenario, CX leaders who seek to proactively deliver AI governance should incorporate the five elements for AI guardrails by design.
R "Ray" Wang is author of Everybody Wants to Rule the World: Surviving and Thriving in a World of Digital Giants (HarperCollins Leadership) and founder of Constellation Research.