As enterprises move beyond generative AI to autonomous, decision-making systems, the ethical stakes are rising sharply. In this interview, Reggie Townsend, vice-president of Ethics, Governance & Social Impact at SAS, explains to Sudhir Chowdhary why agentic AI calls for a fundamental rethink of accountability, trust and governance – arguing that the real challenge is no longer just building intelligent systems, but ensuring they act responsibly in the real world. Excerpts:
As the conversation shifts from generative AI to agentic AI, how do the ethical stakes change?
The distinction is simple but profound. Generative AI informs; agentic AI acts. When a system answers a question, humans retain the ability to evaluate and intervene. When it acts on your behalf, that moment of judgment has already passed.
What complicates matters is the spectrum of autonomy. At one end are fully autonomous agents that execute tasks independently; at the other are systems with humans involved at every step. Most enterprises will operate somewhere in between. Where they sit on that spectrum directly determines their ethical exposure.
This raises very real questions: who does the agent represent, what liabilities arise, and who bears the consequences if something goes wrong? These are not abstract concerns. Ethical thinking must precede deployment, not follow failure.
Who should be accountable when an AI agent causes harm?
Responsibility lies with both the builder and the deployer. This aligns with the direction of regulation, particularly the EU AI Act, which places primary responsibility on the deployer. The analogy is straightforward. If a defective product causes harm, both the manufacturer and the retailer may be liable. AI will follow a similar logic. Courts are still shaping this, but the trajectory is clear. Enterprises that fail to define accountability structures now risk exposure as legal frameworks mature.
How can organisations ensure auditability in systems that operate through chains of decisions?
Auditability cannot be retrofitted; it must be foundational. At SAS, the approach rests on four pillars. First is culture – creating an environment where people feel accountable for AI-driven decisions. Second is operations – rethinking ownership, workflows and metrics in an AI-driven context. Third is regulatory awareness – navigating a fragmented global landscape. Fourth is controls – monitoring, audit trails and escalation mechanisms. The real test is simple: can you explain what your AI did, why it did it, and how you managed risk? If not, governance has failed.
How does bias evolve when AI systems act independently?
Bias in agentic AI extends beyond the model into the entire sociotechnical system. A critical but under-discussed factor is human behaviour during training. If employees fear displacement, they may resist or degrade training inputs – consciously or not. That “noise” becomes embedded in the system. When such systems act autonomously, the impact compounds.
This is why bias and trust are inseparable. If people distrust the intent behind AI, data quality suffers. Addressing bias therefore requires leadership, transparency, and clear communication about AI’s role in augmenting – not erasing – human work.
What does ‘trustworthy AI’ actually require in an agentic environment?
“Trustworthy AI” risks becoming a slogan unless grounded in practice. Nearly half of employees still do not fully trust AI in their organisations, and that distrust is often directed at leadership, not the technology.
Building trust requires cross-functional governance. At SAS, oversight includes legal, HR, finance, marketing and technology. It also requires investment in AI literacy – something regulators are beginning to mandate.
Crucially, governance must be treated as a business discipline, not just compliance. When SAS mapped its internal AI use cases, it found multiple approaches to the same task. Governance created visibility, enabling standardisation and better outcomes. That is what trust looks like operationally.
What ethical challenges should fast-growing AI markets like India prepare for?
The focus should not be on adoption, but purpose. “Adoption for what?” is the more important question, because it reframes ethical decision-making. AI literacy is key. It enables organisations to recognise when AI is not the right solution. Not every problem needs a generative model layered on top. There is also reason for optimism. Markets like India have a second-mover advantage. The foundational infrastructure already exists; the real opportunity lies in building applications that prioritise inclusion and human capability, rather than just efficiency.
What is the biggest ethical blind spot in AI today?
The lack of focus on vulnerable populations. Too often, the people most affected by AI decisions are excluded from the conversation. There is a real risk that AI amplifies existing inequalities. But that outcome is not inevitable. In fact, there is strong potential – particularly in emerging markets – to use AI for inclusive growth.
Real-world examples, such as rural women using drone technology for agriculture, show what responsible deployment can achieve. But these outcomes require deliberate intent. The system will not naturally optimise for inclusion – you have to design for it. That is the blind spot: not a lack of capability, but a lack of sustained commitment to making inclusion a priority.