Verify every action against every policy. Start with the law.

One check before every action: law, industry rules, contracts, and internal policy. The strictest layer wins. Every answer cites its source.
lending-agent-v3
POLICY OS
Dashboard
Actions
Policies
Settings
Active agent
Overview
Monitoring
Evals
Audit log

Policy stack check

Run again
Send credit denial email
Running check Live
LAYER 1
Law & regulation
ECOA, 15 U.S.C. §1691(d): adverse action notice lacks specific reasons
Queued
LAYER 2
Regulator guidance
CFPB Circular 2022-03: algorithmic denial needs specific reasons
Queued
LAYER 3
Contract
Client agreement §4.2: data use and notices
Queued
LAYER 4
Internal policy
External communication policy v2.1
Queued
LAYER 5
Brief
Client brief: tone and required disclosures
Queued
Blocked by Law layer. Stricter policy wins.
1,000+

Global regulators monitored

1,000+

AI-related acts mapped

50+

Countries covered

Per-agent

Policy checks, guardrails, and audits

running against every policy layer
rule change
Error-resolution window shortened
12 CFR §1005.11
policy update
Sanctions list refreshed
Federal · OFAC
legal check
Filing threshold enforcement
31 CFR §1010.311
Agent action: transfer request
Checked against all three
See it work

Watch an agent get checked.

Rule changes, policy updates, and legal checks — running against live regulation.
Four pillars

Three pillars ask how the agent behaves.
The fourth asks if it's allowed to.

Safe, secure, and reliable each have a framework. Policy verification doesn’t.
Safe

Does it avoid harmful output?

Secure

Is the system protected?

Reliable

Does it behave consistently?

Policy verified

Is this action allowed under every policy that applies?

Coverage

The questions that create liability
aren't on any checklist.

Seven layers of policy, run across the frameworks your buyers already trust.
Policy layer Example question SOC 2 / ISO NIST / AIUC-1 CARVER
Law & regulation Is this action allowed where this user is? SOC 2 / ISO NIST / AIUC-1 CARVER
Regulator guidance & enforcement Would a regulator read this denial as needing human review? SOC 2 / ISO NIST / AIUC-1 CARVER
Industry & licensing rules Is this advice only a licensed professional can give? SOC 2 / ISO NIST / AIUC-1 CARVER
Contracts & client terms Does this use of data breach a customer agreement? SOC 2 / ISO NIST / AIUC-1 CARVER
Internal policy Does this message follow our external communication policy? SOC 2 / ISO NIST / AIUC-1 CARVER
Team briefs Does this creative stay inside the client brief? SOC 2 / ISO NIST / AIUC-1 CARVER
Change over time Has any layer changed since this agent went live? SOC 2 / ISO NIST / AIUC-1 CARVER
Checked per action
Governance-adjacent
Not covered
Compliance asks if your controls are sound. Policy verification asks if this action is allowed under every policy that applies. Your SOC 2 report answers the first one.
Cases

Nobody was breached.
The systems worked as designed.

That was the problem. The exposure came from the law, not the system. These are all law-layer failures: the strictest policy layer.
RealPage

Software priced rents using competitor data. The model ran as built.

DOJ + 8 states
2024
SYSTEM
Reliable and consistent
LAW
Sherman Act §1
GAP
Automation doesn’t exempt price coordination
Correct prices. Coordinated ones.
UnitedHealth

An algorithm predicted care length, allegedly used to cut coverage.

nH Predict
class action
SYSTEM
Consistent, high throughput
LAW
CMS individualized-review rules
GAP
Overriding clinicians is a statutory issue
Reliable at the restricted thing.
Cigna · PXDX

Automated review allegedly denied batched claims in seconds.

CA bad faith
class action
SYSTEM
Fast, deterministic, auditable
LAW
CA unfair-claims statutes
GAP
Speed became the allegation
Fast, consistent, allegedly unlawful.
Unlicensed Practice

An AI agent provides advice tailored to a person's specific legal situation.

AI + legal services
UPL risk
SYSTEM
Accurate, responsive, personalized
LAW
Unauthorized practice of law
GAP
Capability can cross the licensed-profession line
Helpful advice. Potentially unauthorized practice.
Product

Policy verification infrastructure for agents.

One layer between your agent and its next action.
Regulator feeds aren't enough. Carver maps acts, case law, and enforcement actions, and your contracts and internal policies: how a rule is actually applied, not just that it exists.

Your agent asks one thing: is this allowed, here, now? Carver answers with the governing rule and its source. Anything ambiguous goes to your legal team, not through the gate.

MCP server

Drop into any stack
Checks available where the agent acts.

SDK

Check at runtime
In-request call, answer plus citation.

Evals

Policy evals
Adversarial testing against the laws and policies that apply to you. Catch it before you ship.

Frameworks

Where agents get built
LangChain, CrewAI, and the rest.

Impact models

Run your own liability model
Weight exposure by jurisdiction, product, contract, or policy. Score against your risk appetite.
Support

FAQs

Quick answers on how Carver verifies agents against law and policy. More in our knowledge base.
What is policy verification?
Policy verification checks an agent’s actions, and the work it produces, against every policy that applies: law and regulation, industry guidance, contracts, organizational policy and team briefs. Law is the strictest layer, and Carver maintains it across 50+ countries. When layers conflict, the stricter one wins, so a client brief can’t override the law.
What is legal risk in AI agents, and how is it different from safety risk?
Safety risk is about harmful output. Legal risk is about whether an action is allowed under the law that governs it: jurisdiction, statute, case law. An agent can pass every safety and security check and still take an action that’s illegal. These are separate risks, and most frameworks only cover the first.
Can an AI agent be safe, secure, and reliable but still act illegally?
Yes. Safe, secure, and reliable each have an established framework: red-teaming, SOC 2/ISO 27001, NIST AI RMF. None of them check whether a specific action is legal in the jurisdiction where it’s taken. An agent can meet all three standards and still violate the law that applies to that action.
What is the difference between AI compliance and policy verification?
Compliance asks if your controls are sound: do you have the right processes, certifications, and audits in place. Policy verification asks if a specific action or asset is allowed under every policy that applies, starting with the law. A company can be fully compliant on paper and still take individual actions that break the law or its own policy.
Does SOC 2 or ISO 27001 certification mean an AI agent’s actions are legal?
No. SOC 2 and ISO 27001 demonstrate that your systems and controls meet defined requirements: that the agent behaves as designed and data is protected. Neither certification checks whether a given action is permitted under the statute, regulation, or case law that governs it in a specific jurisdiction.
What happens if an AI agent takes an action that violates a law the developer didn’t know about?
Not knowing about an applicable law doesn’t necessarily protect a company from liability. An agent can work exactly as designed and still create legal exposure if its actions violate the laws that apply to it. This is why legal risk needs to be evaluated at the action level, not just through system-level controls and audits.
Does Carver block an AI agent from taking an illegal action?
Carver determines the applicable legal requirements and recommends whether an action is allowed, restricted, or requires review. The agent governance system then decides whether to allow, block, or steer the action based on those recommendations and the organization’s own policies.
Which jurisdictions does Carver cover, and how is it integrated?
Carver’s law layer covers 50+ countries, with the deepest coverage in the US and EU, and it keeps expanding across jurisdictions and regulatory domains. It can be integrated throughout the AI application lifecycle: from development and testing to runtime decisions, monitoring, and audit, so the same legal intelligence can be used before and after deployment.
How does Carver check if an AI agent’s action is legal in real time?
Carver checks an agent’s intended action against the laws and regulations that apply to its location, industry, user, and activity, then determines what is allowed, restricted, or required, with the underlying legal sources for auditability. Our regulatory coverage is continuously expanding across jurisdictions and domains to reduce blind spots as the legal landscape evolves.

SOC 2 proves behavior.
Carver verifies policy.

See a policy check run against your own agent.