When the agent screws up, who's liable — us, AWS, or the customer?

🎬 Cold Open

Alright, another conference hallway, another deal on the brink. Today’s scenario: the customer’s legal team wants to know, when that shiny new AI agent screws up, who’s on the hook? Us, AWS, or them? Get ready, it's a minefield.

🎙️ The Argument — When the agent screws up, who's liable — us, AWS, or the customer?

Marcus: So, just left that manufacturing deal. Legal team's in the room, totally threw a wrench in Dev's flow.

Dev: Hey! It was going great until Mr. 'General Counsel' wanted to talk about... 'agentic liability.' Honestly, it's fine. We just tell them our agents are super smart!

Priya: Uh-huh, and then what, Dev? You promise perfection? Because *McKinsey says* that inaccuracy is still a primary AI risk. It's the number one barrier to scaling agentic AI, actually.

Marcus: Exactly. I've watched deals evaporate when that question comes up. You can practically see the P&L flash before their eyes. The whole deal went south.

Dev: But... it *is* smart! I built it on Bedrock, it's got Guardrails! Those new Bedrock Guardrails cross-account safeguards are amazing for PII redaction and hallucination detection.

Priya: Guardrails reduce risk, Dev, they don't eliminate it. *Knostic analysis* highlights the need for layered controls – input filtering, output inspection, grounding. 'Amazing' isn't a legal defense.

Marcus: Alright, let's reenact. Priya, you're the customer's head lawyer, Ms. Albright. Dev, try to close this. Don't promise perfection. Just handle it.

Priya: (Deep, official voice) Mr. Dev, my client is excited about this supply chain optimization agent. But when it recommends a critical component supplier that later defaults, who is ultimately liable for the production shutdown?

Dev: (Nervous) Well, our models are, uh, designed to be very robust... and we have, uh, logs...

Priya: (Interrupting, as lawyer) 'Robust' isn't a legal defense, Mr. Dev. Nor are 'logs' sufficient evidence of due diligence. I need documented governance.

Marcus: See? Dead in the water. Dev, your 'it's fine' approach won't fly. You sound like a deer in headlights.

Dev: Okay, okay, but what *do* you say? It feels like a brand new problem!

Priya: It's not new, Dev. It's just 'governance' with a new hat. Remember the *EU AI Act*? August 2nd, 2026 – that's two months from now. High-risk obligations mean you *have* to document your governance.

Marcus: Documented governance is the new table stakes, kid. No, you don't document the agent's brain.

Priya: You document the controls, the auditability, and the accountability mechanisms. You show them what happened with Bedrock audit logging, with CloudTrail. Every action is traceable.

Dev: So, we're selling logs and fancy paperwork? Not just the cool agent?

Marcus: We're selling *accountability*, Dev. It's a feature. We show them the Well-Architected RAI Lens principles we followed. We show them we limited the blast radius.

Priya: Like IAM least-privilege. The agent literally CAN'T do the catastrophic thing because it doesn't have the permissions. If it suggests 'delete all data,' it can't execute it.

Dev: Oh! So it can recommend something bad, but it physically can't mess things up if we restrict its actual capabilities?

Marcus: Precisely. And for their legal team, you point to AWS Audit Manager – that's their compliance evidence. This isn't about promising the AI is perfect.

Priya: It's about demonstrating control, investigation capabilities, and the customer's own ability to oversee. Not guaranteeing infallibility. Accountability isn't a burden, it's a competitive advantage.

Dev: So, accountability isn't just about 'whose fault is it?' It's about 'how do we prove we're being responsible and giving *them* control?'

Marcus: Exactly, Dev. At the end of the day, it's their agent, in their account, running their business. Our job is to give them the framework to operate responsibly. Make that the sales pitch.

Priya: The answer to 'who's liable?' is always 'we gave you the tools to know, and to contain.'

Dev: And to limit the damage if I built it too fast and didn't test every edge case.

Marcus: Hey now. Next time, Dev, when legal's in the room, don't just smile and nod. Tell 'em about the logs. They love logs. And tell 'em about limiting blast radius. They love that even more.