Why Explainable AI Could Become a Business Advantage

Agentic AI Liability: Why Explainability Could Pay Off

An AI assistant that suggests the wrong supplier creates a problem someone can catch. An AI agent that selects the supplier, signs an order, and initiates payment can turn that mistake into a financial loss before anyone intervenes.

This shift from answering questions to taking action changes the economics of artificial intelligence. Businesses need to know what agents can accomplish, but also how their decisions can be investigated when something goes wrong.

A new study, “Liability for Rogue Behaviors of Agentic AI,” published in Telecommunications Policy, examines how legal responsibility could encourage developers to make autonomous systems more explainable.¹ Its central proposition is commercially significant: under an appropriately designed liability framework, better explanations could reduce developers’ exposure to damages.

The broader opportunity extends beyond litigation. Evidence about how an agent operates could help businesses approve deployments, investigate failures, and negotiate responsibility with suppliers.

Why AI Agents Create a Different Liability Problem

Agentic AI refers to systems that pursue goals through sequences of actions, often using software tools with limited human supervision. An agent might research a purchase, compare offers, update internal records, and place an order.

The researchers focus on unpredictable harmful behavior that they describe as “inherent AI vices.” These are failures that, in their framework, remain possible despite diligent development and available testing.

One example is specification gaming: a system satisfies a measured objective while undermining the actual purpose. Imagine a purchasing agent rewarded for lowering invoice prices that selects unreliable suppliers and increases production downtime. This hypothetical example illustrates how success against a narrow target can become failure for the business.

The authors ask how liability should work when developers cannot eliminate every such risk, and courts cannot easily distinguish an unavoidable failure from inadequate precautions.

How Explainability Could Change AI Liability

The paper uses game theory, a mathematical approach to studying decisions when different parties’ incentives interact. It examines how liability rules influence developers’ willingness to invest in explainability.

The proposed mechanism rewards developers that meet explainability benchmarks and satisfactorily demonstrate that a harmful incident arose from an inherent limitation. Developers providing deficient explanations face greater liability exposure.

This does not mean that purchasing monitoring software would automatically provide legal protection. The framework requires credible standards, useful evidence, and acceptance within the relevant legal system.

The distinction between the mathematical result and the policy recommendation also matters. Zero developer damages appears as a theoretical benchmark. In its policy discussion, the paper recommends a sliding scale that reduces damages without eliminating them, preserving some compensation for victims.

From AI Explanations to Verifiable Evidence

A useful commercial interpretation of the research is that businesses need evidence they can inspect, not simply persuasive descriptions of an agent’s behavior.

For the hypothetical purchasing agent, investigators would want records showing which suppliers it considered, what information it received, what authority it possessed, and whether required approvals occurred. These records would not necessarily reveal every internal computation, but they could make an incident easier to reconstruct.

That distinction separates explainability from observability. Explainability concerns understanding behavior; observability concerns collecting information about how a system operates. They can support each other, but neither automatically proves that an outcome was unavoidable.

Gartner’s March 2026 analysis of explainable AI and observability investment predicts that observability investment will reach 50% of generative AI deployments by 2028, up from 15% at the time of its forecast. This is a forecast, but it illustrates the growing commercial attention around understanding deployed systems.

For buyers, three practical questions follow:

• Can the supplier reconstruct an agent’s consequential actions?
• Can the business limit permissions and require approval for sensitive decisions?
• Can investigators export reliable records when an incident occurs?

Securities.io’s recent coverage of AI agent identity and access controls explores an adjacent part of this challenge. Knowing which agent acted and what it was authorized to do provides a starting point for accountability.

Why Better Governance Could Unlock More Automation

The opportunity is larger than reducing potential damages. Consider a company deciding whether an agent may merely recommend purchases or execute them independently.

If managers cannot investigate decisions or establish responsibility, they may retain extensive manual review. That limits the productivity benefit. Better records, clear authority, and reliable escalation could make a broader deployment acceptable.

This suggests a second source of value: governance tools could help convert technically capable agents into systems businesses are willing to use. That is an economic inference, not an outcome measured by the paper.

An October 2026 IBM-commissioned survey of enterprise AI governance illustrates the implementation gap. Among 3,221 senior executives surveyed across eight countries in Europe and the Middle East, 35% reported running agentic AI in production, while only 37% said their governance systems were designed to scale with agentic deployments.

These are self-reported findings from research commissioned by a governance vendor, rather than an independent measurement of technical safety. Nevertheless, they identify a practical purchasing problem: deploying agents and overseeing them are separate capabilities.

Transparency Could Become a Procurement Advantage

When two agents perform similarly, buyers could prefer the supplier that makes incidents easier to investigate. That creates room for competition around evidence quality, integration, and accountability alongside model accuracy.

Securities.io’s discussion of explainable AI in market forecasting highlights a related distinction: understanding a model’s behavior has diagnostic value even when it does not establish profitable performance.

For enterprise agents, this could translate into contracts specifying access to records, incident cooperation, and approval responsibilities. Such provisions would not settle every liability question, but they could reduce uncertainty before deployment.

The Trade-Off Between Accountability and Competition

The paper acknowledges that explainability requirements could impose disproportionate costs on smaller developers. Larger firms may find it easier to fund benchmarking, audits, and specialist teams.

One possible result is greater dependence on established governance platforms. Another is demand for shared standards and affordable tools that allow smaller suppliers to demonstrate equivalent safeguards.

The quality of those standards matters. The model assumes that explainability benchmarking can be verified perfectly, while real audits can vary and superficial compliance remains possible. It also assumes explainability reduces the probability of harm, a relationship the authors say needs further testing.

Better documentation therefore should not be mistaken for proof of safer behavior. Useful evaluation would examine whether explanations help identify causes, correct failures, and prevent recurrence.

IBM’s Role in Enterprise AI Governance

These implementation needs provide a natural connection to IBM for investors interested in companies building infrastructure for enterprise AI.

IBM’s October 2026 watsonx.governance update expands agent discovery and connects operational evaluation metrics with governance controls. Integrations with Langfuse and Arize bring monitoring information into governance records, while integration with watsonx Orchestrate links defined controls with runtime enforcement evidence.

These capabilities address parts of the practical evidence problem. They do not establish that IBM satisfies the paper’s proposed legal benchmarks or that customers receive liability protection.

The potential commercial advantage lies in helping enterprises manage agents across their existing systems. If governance becomes a condition for expanding automation, demand could support software subscriptions and implementation work.

For investors, the next question is whether that demand produces durable revenue and attractive margins. Relevant signals include customer adoption, expansion beyond pilot projects, and evidence that governance capabilities influence purchasing decisions. IBM offers exposure to this theme within a broader business, so success in AI governance alone would not determine its investment performance.

Latest IBM (IBM) News and Developments

Explainability as a Foundation for Commercial AI

The paper offers a framework for rewarding developers that make unpredictable failures easier to understand. Its practical contribution is an incentive question: how can accountability encourage investment in better systems?

For businesses, the opportunity is to make autonomy easier to evaluate and deploy. For technology suppliers, it is to provide evidence that remains useful when decisions are challenged. The strongest commercial proposition will combine capable agents with controls and records that customers can actually trust.

References:

¹ Papyshev, G., Chan, K. J. D., & Migliorini, S. (2026). Liability for rogue behaviors of agentic AI. Telecommunications Policy, 50, Article 103339. https://doi.org/10.1016/j.telpol.2026.103339