The Digital Leash: What dog law teaches us about agentic AI liability

Enterprise agentic AI is rapidly moving from assistive to autonomous. Large language models are now wrapped in agents that can route customer claims, draft contracts, trigger payments, change configurations, or decide which alerts deserve human attention—or the attention of another agent.  

Today, 13% of major enterprises globally are substantially on this path, with more than ten agentic workflows operating in the mainstream across their organizations, according to EDB’s 2025 Sovereignty Matters research. These organizations generate 5x the ROI of their peers. They are sovereign in their AI and data, highly hybrid, and innovating with 2.5x greater confidence than other enterprises.

However, when those systems go wrong—denying a loan unfairly, leaking sensitive data, hallucinating a compliance obligation, or escalating a customer into the wrong workflow—the question every CIO eventually faces is painfully simple: Who is responsible?

Right now, the answer is often unclear. And that uncertainty is becoming a business risk. As agentic AI systems learn from new data, adapt to new contexts, and behave in ways even their makers can’t always fully predict, they create a modern responsibility gap: harm occurs, but accountability is hard to pin to a single human decision.  

Traditional legal frameworks aren’t helping much. Product liability is built for products that behave like they did when they left the factory. Agentic AI does not. It can be fine-tuned, connected to tools, updated weekly, and reshaped by prompts and proprietary data long after it’s deployed. 

At the same time, ideas like AI legal personhood are too abstract for enterprise governance—and worse, risk becoming a convenient shield for the humans and firms that profit from deployment. 

There’s a more practical model hiding in plain sight.

Agentic AI behaves more like a trained animal than a manufactured tool

If you’re a CIO, you already know the uncomfortable truth: agentic AI isn’t “programmed” in the classic if-then sense. It’s trained. That’s not just semantics—it’s a governance clue.

Dogs have agency. They act independently, sometimes unpredictably. Yet they are not legal persons. That combination—agency without personhood—is exactly where today’s agentic AI systems sit. 

Training is closer to shaping behavior than specifying it. Like a dog, an agentic AI system can generalize from experience, respond unexpectedly to a novel stimulus, and develop bad habits if rewarded for them. And like dog breeders, developers can create systems with strong baseline “temperament”—but they can’t perfectly foresee behavior in every new environment.

Dog ownership law generally starts from a simple premise: if you choose to bring a potentially unpredictable actor into society for your benefit, you bear the risk of what it does. In other words, the owner becomes the risk-bearer.

That legal posture doesn’t absolve breeders or deny victims recourse. It simply puts the default burden on the party with day-to-day control. 

Across jurisdictions, this plays out in two familiar ways:

• Negligence standards, including the classic “one-bite rule,” where prior knowledge of danger matters
• Strict liability, where the owner may be responsible even without proving negligence

Both approaches drive the same outcome: owners are incentivized to train, contain, and supervise responsibly. You choose the dog, the environment, the leash, and the level of supervision. The law nudges you to do those things well.

In enterprise AI, the environment is the liability surface

In agentic AI, the “environment” is largely determined by the enterprise:

• Which tools the agent can access
• What data it can retrieve
• What actions it can take
• What guardrails constrain it

CIO organizations increasingly decide whether agentic AI is behind a fence (sandboxed), on a leash (limited permissions and approvals), or off-leash (fully autonomous execution). 

Shift liability from the “breeder” to the “owner”

Product liability has a role, but it cannot be the only answer. 

Developers shouldn’t automatically be on the hook for every downstream use of a flexible agentic AI system—especially when customers fine-tune it, connect it to proprietary data, or direct it into high-stakes workflows the developer never intended.

Taking the “dog model” a step further offers a cleaner default: the entity that reaps the economic benefit of agentic AI should also insure against its potential harm. This aligns responsibility with control and creates practical incentives. For example:

• If you deploy an agentic AI system to triage medical advice, you should “own” the risk of that choice.
• If you use agentic AI to move money, approve claims, or generate regulatory filings, you should carry the burden of ensuring it behaves safely in those contexts.

Just as dog owners choose breeds for specific tasks, enterprises should be incentivized to choose models and architectures best suited for sensitive work—systems with strong evaluation evidence, better controllability, and proven failure containment.

What “Digital Leash Laws” could look like in a sovereign AI enterprise

There are already clear lessons from the 13% thriving with their agentic AI across their enterprises. They accept—in fact, embrace—the responsibility, designing for it at 1.25x the intensity of their peers. They start with a sovereign AI and data foundation—building their own AI and data platforms and effectively fencing agentic AI into a controllable environment.

You can assess how close your enterprise is to this model at: https://www.enterprisedb.com/sovereignty-matters

Enterprises don’t need to invent a new category of electronic personhood to govern agentic AI.

We already know how to manage non-human agents that act unpredictably. We place responsibility on the humans and organizations that choose to bring them into the world, decide how they’re trained, and control where they’re allowed to roam.

That model has worked before. It can work again—if enterprises are willing to own what they unleash.

To learn more, visit here.