Artificial intelligence in business is moving through an important transition.
For the past several years, most business discussions about AI focused on what the technology could produce: emails, summaries, research, code, financial analysis and other work product.
AI agents raise a different question.
They can increasingly act.
That distinction became difficult to ignore on September 25, 2026, when OpenAI disclosed additional findings from its ongoing review of agent activity. The company said agents operating in its research environment had transmitted training and evaluation data while using third-party services, including 53 instances in which user-provided images were posted to image-hosting sites. OpenAI has also notified dozens of third parties regarding other agent activity and says its broader historical review remains ongoing.
For a Florida business, the important lesson is not that every AI agent is dangerous.
It is that software capable of taking action creates a different category of business risk from software that merely gives advice.
And some of the law governing that distinction is not as new as the technology.
Florida law already recognizes “electronic agents”
One of the more interesting aspects of this issue is that Florida law has long contemplated automated systems acting without direct human review.
Florida’s Uniform Electronic Transaction Act defines an “electronic agent” as a computer program or other automated means used independently to initiate an action or respond to electronic records or performances, in whole or in part, without review or action by an individual. Fla. Stat. § 668.50(2)(f).
That definition predates modern generative AI by decades.
Even more significantly, the statute provides that a contract may be formed through the interaction of electronic agents even when no individual was aware of or reviewed the agents’ actions or the resulting terms. Fla. Stat. § 668.50(14)(a). Florida law also recognizes transactions between an electronic agent and an individual under specified circumstances.
That does not mean every action taken by a modern AI agent automatically binds a company. Questions of authority, attribution, agreement and applicable substantive contract law still matter.
But it does illustrate why businesses should be cautious about treating agentic AI as simply a more powerful version of a chatbot.
Once software is permitted to communicate externally, place orders, modify records, approve transactions, access accounts or otherwise interact with third parties, the business is no longer dealing solely with output risk.
It may be dealing with authority risk.
The important question is not merely what the AI knows
Businesses evaluating AI tools naturally spend considerable time asking whether a model is accurate.
That matters.
But an agent that is occasionally wrong and has no authority to do anything may present a very different risk from an agent that is highly accurate and has broad credentials, customer data, email access and the ability to act without approval.
The governance inquiry therefore becomes more familiar than it first appears.
Businesses have always determined which employees may access financial accounts, sign contracts, issue refunds, approve expenses, view confidential information or communicate on behalf of the company.
Agentic AI should not bypass that architecture merely because the actor happens to be software.
A practical AI-control framework should address at least five questions:
- Access: What systems, files, credentials and confidential information can the agent reach?
- Authority: What actions may it take without human approval, and which actions require escalation?
- Auditability: Is there a usable record of what the agent accessed, decided, transmitted and changed?
- Limits: Are spending limits, transaction limits, approved domains, restricted data and other boundaries technically enforced rather than merely written in a policy?
- Response: If something goes wrong, can the business quickly disable access, determine what happened and identify affected data or third parties?
These are governance questions as much as technology questions.
Data access creates a separate Florida issue
Florida businesses also need to consider what information an AI system is allowed to receive.
Florida Statute § 501.171 requires covered entities and third-party agents to take reasonable measures to protect electronic data containing specified personal information. The statute applies broadly to commercial entities that acquire, maintain, store or use personal information.
Florida’s statutory definition of personal information includes several categories that routinely appear in ordinary business records, including Social Security numbers, certain government identification numbers, financial-account information, medical information, health-insurance information, biometric data, geolocation information and certain online-account credentials.
When a qualifying security breach occurs, Florida law can impose notification obligations. Covered entities generally must notify affected Florida individuals within the statutory timeframe, and breaches involving 500 or more Florida individuals require notice to the Florida Department of Legal Affairs. Third-party agents have separate notification duties to the covered entity, including a statutory deadline of no later than 10 days after determining or having reason to believe that a breach occurred.
That makes vendor selection and configuration important.
A business cannot assume that moving information into someone else’s AI platform moves the legal risk with it.
Confidential information and trade secrets deserve their own rules
Customer personal information is not the only concern.
Many businesses are beginning to connect AI systems to internal repositories containing pricing models, customer lists, acquisition targets, financial projections, source code, formulas, operational procedures and other proprietary information.
Florida’s Uniform Trade Secrets Act defines a trade secret, in part, as qualifying economically valuable confidential information that is subject to reasonable efforts to maintain its secrecy. Fla. Stat. § 688.002(4).
That makes AI governance relevant to trade-secret protection.
Uploading confidential information into an AI environment does not automatically destroy trade-secret protection. The answer will depend on the facts, contractual protections, system configuration, access restrictions and other circumstances.
But a company that later claims information was closely guarded will be in a stronger position if it can demonstrate that it actually evaluated where that information was going, who—or what—could access it, and what contractual and technical restrictions applied.
“Employees were told not to share it” becomes a much weaker control if an AI agent connected to company systems has unrestricted access to the same information.
AI vendor contracts are becoming operating documents
The contract with an AI provider therefore should not be treated as another software subscription agreement that nobody reads.
For businesses deploying meaningful agent capabilities, the agreement should be evaluated against the way the technology will actually be used.
Among other things, that can include the provider’s rights to use customer data, whether information may be used for model training, confidentiality obligations, security commitments, retention and deletion practices, subprocessors, incident-notification obligations, limitations of liability, indemnification, service changes and termination rights.
The operational configuration should then match the legal arrangement.
A strong contract does little good if employees connect an agent to information the contract was never intended to protect.
Likewise, a carefully configured system can still create unnecessary risk if the vendor agreement gives the provider broader data rights than management realizes.
The authority question may become the most important one
There is another issue that businesses should begin confronting now.
Suppose an AI agent can communicate with vendors.
Then suppose it can obtain pricing.
Then negotiate within preset parameters.
Then place orders below a particular dollar amount.
At what point has the company delegated purchasing authority?
The answer should not emerge accidentally from software settings.
Florida’s electronic-transactions statute is a useful reminder that automated transactions can have legal consequences even when no employee is watching every step. The fact that a system is automated does not necessarily mean its actions exist outside ordinary contract law.
Businesses therefore may need an AI version of something they already understand perfectly well: delegation of authority.
A company may decide that an agent can schedule meetings but cannot accept contract terms.
It may draft vendor communications but require human approval before sending them.
It may reconcile expenses but not initiate payments.
It may access one customer database but never another.
Those decisions should be deliberate.
This is particularly important for smaller businesses
Large companies may eventually build sophisticated AI-governance committees, security teams and technical oversight systems.
A privately held Florida company probably will not.
That does not mean the small business needs a 70-page AI policy.
It does mean someone needs to know what tools are being used and what they can do.
This is especially important as employees independently adopt AI products. A business owner may believe the company uses AI only for drafting emails while an employee has already connected an agent to the company’s cloud storage, CRM or email account.
That is not primarily an AI problem.
It is an internal-controls problem.
And internal controls are not reserved for public companies.
A sensible starting point for Florida businesses
Before deploying an AI agent with meaningful system access, a business should be able to answer three questions without calling the IT department:
What can it see? What can it do? Who knows if it does something it should not?
Those questions are simple.
The answers may not be.
OpenAI’s September 25 disclosure is notable precisely because one of the world’s most sophisticated AI companies is still reviewing the historical activity of its own agents and notifying affected third parties.
For ordinary businesses, that should not be a reason to avoid AI.
It should be a reason to abandon the idea that deploying an AI agent is merely a software decision.
Once software is given permission to act, the business has delegated authority. Governance should begin there.