When Uber's COO admitted the company burned through its entire 2026 AI budget in four months without clear ROI, he wasn't just confessing to sloppy spending. He was exposing a chasm in modern enterprise operations: nobody knows who authorized the AI to spend that money in the first place.
This is the authorization liability gap. It's not about whether AI agents can make autonomous decisions. They can. It's about whether anyone in the organization actually approved those decisions, who's responsible when they go sideways, and whether your legal team knows this is happening.
The gap is becoming liability, and liability is becoming law.
The Authority Problem Nobody's Talking About
Most organizations deploy AI agents with vague permission structures. A team spins up a Claude instance to "improve customer service." Another group uses an agent to "optimize marketing spend." Engineering fires up autonomous systems to "streamline operations." Each one incrementally gains more autonomy, deeper access, and higher stakes but rarely with explicit board approval, CFO sign-off, or legal review.
The problem crystallizes when something breaks. When an agent makes a bad call, the blame ricochets. Was it the person who deployed the agent? The one who set its parameters? The developer who built it? The executive who greenlit the initiative? The board that didn't know it existed?
Real question: Can we know what the agent is actually doing?
Nobody has a clean answer because nobody asked the authorization question upfront.
The FSB just published draft guidelines calling for explicit "authorization points" and "human approval checkpoints" in agentic AI systems. They're not doing this because they're paranoid. They're doing it because liability is already fuzzy, and regulators know that ambiguity is a tort waiting to happen. When an agent causes harm, regulators want to know: who actually approved this thing?
Spending Without Guardrails Is Just the Start
Uber's four-month budget burn is the headline example, but it's the obvious case. The real liability exposure is architectural.
Consider what an AI agent does: it makes decisions without human approval in real time. That's the whole pitch. An agent can autonomously negotiate vendor contracts, revise pricing strategies, allocate marketing budget across channels, approve hiring candidates, adjust ad spend, negotiate payment terms, or modify content recommendations all without waiting for human review.
The reality: responsibility shifts to compliance teams with no framework.
Each of these decisions creates downstream liability that your authorization framework might not cover.
If the agent negotiates a bad contract, did the board understand the risk profile before it was deployed? If it makes a discriminatory hiring recommendation, is the HR team liable, the AI vendor liable, or the company that deployed it without oversight? If it tweaks product pricing in ways that trigger antitrust concern, who authorized that decision tree? If it blocks a customer from a service based on a credit score that wasn't their real credit score, who's responsible?
Most organizations have zero formal answer to these questions.
The CIO/CMO friction compounds this. Finance wants autonomous spend optimization because it cuts costs. Marketing wants agents that can act fast without bottleneck approval. Operations wants systems that need minimal human input because human input is expensive and slow. But nobody's asking the questions that matter: "Should an AI agent have this much authority? Does the board know? Does legal? What happens if it goes wrong?"
It just happens. Pilots become production. Production scales quietly. Liability piles up invisibly.
The Blame Loop Breaks Both Ways
Here's what regulators and plaintiff attorneys already understand: the more autonomous an AI agent is, the less you can claim human judgment protected you. And the more human oversight you add, the more liable those humans become for decisions they didn't fully understand or endorse.
This is the blame loop. You can't win either way.
Full autonomy path: You automate everything, minimize human touch, let the agent decide. You save money. But you accept zero accountability cover. If something goes wrong, you can't hide behind "a human made that decision" because no human did. The plaintiffs' bar will argue you were reckless, negligent, that you prioritized cost over safety. The regulators will ask: "Who authorized this level of autonomy? Where are your guardrails? Why wasn't this escalated to human review?"
Full oversight path: You oversee every decision the agent makes. You add review layers, approval checkpoints, human sign-off on major moves. Now you have accountability. But you've basically hired people to rubber-stamp machine decisions. The human reviewers didn't think deeply about the decision. They rubber-stamped it. If it goes wrong, they're liable too. The plaintiffs will argue the humans were negligent for not actually reviewing the decision, just signing off. The regulators will ask: "If humans are reviewing these, why is an agent making them at all? This isn't autonomous operation, this is automation with extra steps."
There's no liability-neutral middle ground right now.
Enterprises are trapped between cost advantage and risk exposure. The cost advantage of autonomous agents evaporates if you have to add so much human review that you've basically hired people to oversee machines. But full autonomy without explicit authorization is just regulatory risk on a schedule, waiting to happen.
What Authorization Actually Looks Like
The FSB's draft practices hint at what adult governance might resemble: boards establish explicit "delegation frameworks" that define what decisions agents can make autonomously, which ones require human sign-off, and what audit trails prove oversight actually happened.
But very few organizations have this yet. Uber didn't. The vast majority of shops deploying agents today don't either.
Real authorization would look like:
- Board approval for agent deployment, documented, with explicit risk acknowledgment (not just a tech team slack message)
- Written decision thresholds (agents can autonomously approve refunds under $100, marketing spends under $50k, hiring recommendations below a certain score but anything above requires human review)
- Audit trails that prove humans actually reviewed critical decisions, not just auto-generated logs that nobody reads
- Clear liability assignment (if an agent makes a bad call on a $500k contract, who's the accountable executive? The CIO? CFO? A specific human?)
- Regular authorization reviews as agent capabilities expand (the agent's role grows, the authorization needs to grow with it)
This sounds like bureaucracy, and it is. It's also the difference between defending yourself in court ("We had a governance framework, explicit authorization limits, and oversight logs") and getting obliterated in discovery ("We had no idea what the AI was doing").
The Timing Problem: Agents Are Outpacing Governance
AI agents are advancing faster than authorization frameworks. By the time your board approves an agent governance policy (if they do), your engineering team has already deployed three new agent types, each with different risk profiles, and nobody told the board because it happened through normal "innovation velocity."
This gap isn't accidental. Autonomous systems are sold as "agile" drop them in, let them learn, they adapt. The authorization question ("Does anyone actually approve this? Should the board know?") feels like an old-fashioned governance bottleneck to teams operating in startup velocity mode, even inside Fortune 500 companies.
But regulators aren't moving at startup velocity. The FSB guidelines, EU AI Act frameworks, and emerging FTC scrutiny of AI decision-making are all signaling the same thing: you can't hide behind "the model did it." Someone authorized the model to exist in production making real decisions. You just have to prove it.
The ones still operating in the shadows are setting themselves up. By the time they realize they need an authorization framework, regulators will already be asking where it was.
The Practical Liability Shift Happening Now
This authorization gap is already shifting liability in ways most enterprises haven't priced in. When an AI agent causes customer harm, regulatory fine, breach of contract, discrimination claim, or financial loss, the plaintiff's attorney will subpoena:
- Who approved the agent deployment?
- What authority was it explicitly given?
- What review happened before it was put into production?
- Did the company have a governance framework? (If yes, did it cover this agent? If no, why not?)
- Who was responsible for oversight?
- What audit trails exist showing humans reviewed critical decisions?
If the honest answer is "nobody knew it was happening" or "the tech team just deployed it," you're not just losing the lawsuit. You're exposing:
- Board-level negligence claims (the board didn't oversee risk management)
- Shareholder derivative suits (executives failed to govern enterprise risk)
- Regulatory investigation (did the C-suite have adequate risk management?)
- Personal liability for executives (depending on jurisdiction, the CEO or CFO might be personally liable for inadequate governance)
The authorization gap turns AI deployment into a governance liability that most organizations haven't quantified yet.
What's Next: The Regulatory Squeeze
The FSB guidelines are drafts now, but they'll become hard rules. The EU AI Act is already law. The FTC is actively investigating AI decision-making in hiring, lending, and marketing. State attorneys general are filing cases about AI discrimination.
All of them will ask the same question: who authorized this?
Organizations that move first on agent authorization frameworks boards that explicitly approve agent deployment, finance teams that define spending limits, legal teams that map liability, compliance teams that audit decisions will be the ones that survive the next wave of AI enforcement cleanly.
The ones that keep deploying agents in the shadows and hoping nobody notices? They're building tomorrow's horror story. And when discovery starts, the emails will be damning.
Bottom Line: Authorization Is the Missing Layer
You can't delegate authority without explicitly granting it. When an AI agent acts autonomously, someone authorized that autonomy, even if it was passive (nobody stopped it). The legal and regulatory world is starting to demand that authorization is conscious, documented, and defendable.
The authorization liability gap is closing. The only question is whether you close it deliberately, or whether regulators and plaintiff attorneys close it for you.