The General Counsel's Guide to AI and Agentic AI

When software takes actions inside the business, the legal questions arrive immediately: who is accountable for what the agent did, what the vendor agreed to, how privacy and intellectual property rules apply, and which regulations attach to each use. This guide gives general counsel a practical structure for governing agentic AI and for using it in legal operations. It is general guidance, not legal advice; the rules vary by jurisdiction and are changing.

Why does agentic AI change the legal picture?

Earlier AI produced outputs that people reviewed before acting. Agents act: they send communications, post transactions, change records, and make commitments within their permissions. Three consequences follow for counsel.

1. Accountability stays with the company. Customers, employees, and regulators will treat an agent's action as the company's action. Recourse against a vendor is a separate question from responsibility to third parties.
2. Controls become legal evidence. Permissions, approval gates, logs, and evaluation records are what demonstrate reasonable oversight after an incident.
3. Obligations attach per use. An agent that handles customer data in one jurisdiction and employment decisions in another carries different obligations for each.

FISTA's guide on who is accountable when an AI agent fails covers the accountability model in more depth.

What should counsel require in AI vendor contracts?

Term	Why it matters	What to look for
Data use and training	Prevents your data from improving a shared model	Explicit prohibition on training; clear retention and deletion
Security and breach notice	Your obligations flow down	Defined controls, certifications, notice periods
Change and deprecation notice	Models are retired and behavior changes	Minimum notice; migration support; version pinning
Service levels	Agents depend on availability and latency	Measurable SLAs with remedies
Liability and indemnity	Allocation of loss and IP claims	Caps proportionate to risk; IP indemnity for output
Subprocessors and audit	Data may flow to third parties	Transparency, notice of changes, audit rights
Portability and exit	Avoids lock-in	Export of prompts, data, logs, and evaluation assets

The LLM vendor lock-in guide explains the technical protections that complement these terms, and the how to negotiate an AI development contract guide covers services agreements.

How do privacy obligations map to agents?

Map data flows per agent: what personal data enters, where it goes (models, tools, logs), how long it is retained, and what decisions it informs. Then apply the rules that attach to each flow, which may include lawful basis, purpose limitation, minimization, transparency to individuals, restrictions on automated decisions with legal or similar effects, and cross-border transfer requirements.

Two areas need particular attention. Logs often retain more personal data than the process itself, because agents record inputs and outputs for observability. Automated decisions in employment, credit, insurance, and similar contexts are regulated in several jurisdictions and may require human review, explanation, or notice. The AI data privacy compliance guide provides a mapping method.

What are the intellectual property questions?

Three, and they are distinct:

• Inputs: does the company have the rights to use the content it feeds to models, including licensed data, customer content, and third-party material?
• Outputs: who owns generated output, and is it protectable? Vendor terms typically assign rights to the customer, but copyright protection for purely machine-generated work is limited in several legal systems.
• Confidentiality: does sending material to a model risk trade-secret status or breach confidentiality obligations to clients and partners?

The practical response is a policy per data class, vendor terms that assign output rights and prohibit training, and documentation of human contribution to valuable output.

How should counsel maintain regulatory readiness?

Keep an inventory that maps each AI system to the obligations that attach in each jurisdiction where it operates: sectoral rules (financial services, healthcare, employment), horizontal AI laws where they exist, privacy law, consumer protection, and advertising rules. Assign a risk tier to each system, and require the controls and records each tier demands. Review the inventory when systems change and when laws change. The AI governance framework guide describes an inventory structure that many legal teams adopt, and the EU AI Act explained guide covers one of the horizontal regimes.

What records make AI decisions defensible?

The records that show the company acted reasonably: the inventory and risk tiers; permissions and approval gates for each agent; evaluation evidence before release and on a schedule; logs of agent actions; human oversight assignments and their records; incident reports and remediation; and the policies employees were trained on. Counsel should confirm these exist before an incident, not discover their absence after one.

How should legal use agents in its own work?

Legal operations has strong candidates: contract review against playbooks, clause extraction, intake triage, matter summaries, and research support from approved sources. The controls are the same ones counsel asks of the business: a lawyer accountable for every output, restricted sources, logging, and no external communication or advice without review. FISTA's AI in legal operations whitepaper covers the use cases and controls in detail.

How can FISTA Solutions help general counsel?

FISTA Solutions builds AI agents with the permissions, approval gates, logging, and evaluation records that make oversight demonstrable, and works with legal and executive teams through its AI enablement practice to establish inventories, risk tiers, and vendor requirements. Since 2017, FISTA has delivered 150+ projects for 50+ companies across 12+ countries.

To review the controls and records behind agents your business is deploying, talk to FISTA on WhatsApp, or start with the AI risk management guide.