The Ugly Truth About AI Accountability Decision Making

AI accountability decision making starts with one fact: when an AI system gives an answer, the company that deployed it owns the result. One panelist at HumanX 2026 put it plainly, and the legal and business consequences come with it.

That same panelist spent years building AI products at Microsoft and Qualcomm before starting Credo AI. After watching high-profile failures in products those companies launched, it became clear that a last-mile problem needed solving.

Who owns an AI decision once it ships?

The company does. Whatever answer a chatbot gives today belongs to the business that put it in front of customers, and agentic workflows add new risk on top of that.

One panelist warned that enterprises can no longer brush off lawsuits, pointing to the $1.5 billion Anthropic copyright case. Another argued that the number of suits involving Warner Brothers, the New York Times, and Universal Music means ignoring copyright is no longer worth the risk.

What does the gap between builders and risk teams cost?

The first panelist described the divide between data scientists and governance staff this way: they hate being in the same room and have no common language.

Jemiah Sius, VP, Market Strategy and Developer Relations at New Relic, said his team has seen production databases get deleted. At scale, he noted, downtime can cost upwards of $2 million in lost revenue. Builders and risk teams who do not talk to each other pay for it.

What happens when a vendor will not indemnify you?

The fix is to bring legal and compliance in before anything breaks. One panelist recalled working with a chief legal officer on a new technology where the vendor refused to indemnify. After six months of back and forth, the chief legal and privacy officers decided they had to adapt. In the panelist’s words, you have to get comfortable with the discomfort and then put guardrails in place that you can live with.

Training takes longer than planned. One panelist rolled out a new application to a non-compliant team, assumed three training sessions would fix it, and ended up running about 12. They also built tools for the compliance team so the team could do its job better. Those tools teach compliance staff how the technology works, and that takes some of the fear out of it. Bringing compliance in early and giving them tools lets a company move faster, and regulatory obligations did not stop the work.

How does shadow AI break an audit?

Sius argued that one engineer deploying an unapproved tool creates shadow AI, which brings compliance and regulatory headaches right away. Businesses now have to track every internal AI deployment, and Sius said failing an audit is close to inevitable without a discovery tool such as an AI catalog.

He expects new governance laws taking effect by the end of 2026 to make this worse. Untracked deployments make risk hard to understand and accountability hard to maintain. Agentic AI in particular is a black box right now, Sius said.

How should you structure AI accountability decision making?

One panelist said some Fortune 500 companies are building centralized AI governance teams that mix data scientists, governance and risk staff, privacy specialists, and psychologists. That mix is needed to understand risks beyond traditional fairness and security, such as adversarial attacks and toxicity.

The same panelist said enterprises are finally moving past policy talk and building continuous runtime guardrails. Drawing on work with Fortune 500 clients across 10 sectors, the panelist called governance a critical feature for enterprise adoption. Another panelist recommended embedding governance in the workflow itself, which gives a real-time, end-to-end view of agentic workflows.

Can people enforce accountability at machine speed?

One panelist recommended multiple kill switches, at the agent, connector, tool, and employee levels, to keep control centralized. Blake Brannon at OneTrust said relying on humans to review every automated scenario is a losing battle. His answer is to fight fire with fire and use AI agents to govern other AI agents.

Another panelist suggested executives spend the next 90 days reviewing consensus best practices such as ISO 42001, the international standard for responsibly developing an AI management system.

You cannot outsource accountability to a vendor. When an autonomous system breaks, the enterprise deals with the legal and business fallout, so AI accountability decision making needs a named owner before launch.

To hear more from the people putting AI into production, see the HumanX 2027 program preview.