How to Use AI in Your Business—Without Compromising Security or Compliance

Updated October 2026

To use AI in your business safely, decide what data each AI tool may touch before anyone uses it. Then buy business plans with no training on your data, block personal accounts for work, and keep a human in front of any output that reaches customers. Finally, check which state, sector, and EU rules apply to your use cases.

This guide is for business owners and operators rolling out tools like ChatGPT, Claude, Copilot, or Gemini. If you need a full governance program with risk tiers and audit evidence, read our data and AI governance framework after this one.

Why is AI adoption a security problem in the first place?

AI adoption becomes a security problem because employees adopt it faster than companies set rules. The 2026 Verizon Data Breach Investigations Report found frequent employee use of AI tools rose from 15 to 45 percent in one year. It also named shadow AI the third most common non malicious data leakage activity.

The cost shows up in breach data. The IBM 2025 Cost of a Data Breach Report found 63 percent of the 600 organizations studied had no policies to manage AI or curb shadow AI. High levels of shadow AI added 670,000 dollars to the average breach cost. Among organizations with an AI related incident, 97 percent lacked proper AI access controls.

None of this means AI is unsafe by default. It means the risk sits in ordinary decisions, such as which plan people use, what they paste, and who checks the answer. Those decisions are cheap to get right early and expensive to fix later.

How should you decide which data can go into AI tools?

Start with a simple data classification and tie every AI tool to the classes it may receive. Four classes are enough for most businesses. Everyone should be able to name the class of a document in a few seconds.

Regulated data needs a contract, not just a setting. For health data in the US, that usually means a business associate agreement. OpenAI says it can sign BAAs for its API Platform, and Microsoft says web search queries in Copilot fall outside its DPA and BAA. Confirm coverage for the exact product and feature you plan to use.

Classification also improves answers. A model that reads your official revenue definition gives better results than one guessing from raw tables. A study from April 2026 found that explicit business semantics raised accuracy by 17 to 23 points across the frontier models tested. Our comparison of RAG, knowledge graphs, and semantic layers covers how teams supply that context.

Which AI plans protect business data?

Business and enterprise plans from the major vendors say they do not train on your data by default. Consumer plans can differ, so the plan matters more than the brand. The table below reflects each vendor's own documentation as of October 2026.

OpenAI

OpenAI says business data from ChatGPT Business, Enterprise, Edu, and the API is not used for training unless you opt in. Data is encrypted with AES-256 at rest and TLS 1.2 or higher in transit. API inputs and outputs may be kept for up to 30 days for abuse monitoring, with zero data retention available for eligible endpoints.

Anthropic

Anthropic changed its consumer terms in August 2025. Free, Pro, and Max users can opt in to training, and opted in data is kept for five years. Claude for Work, the API, Claude Gov, and Claude for Education fall under commercial terms and are excluded from that change. Employees on personal Pro accounts are therefore on different terms than your Team plan.

Microsoft

Microsoft says Copilot Chat prompts, responses, and Graph data are not used to train foundation models. Copilot respects your identity model, sensitivity labels, and retention policies. The catch is that it also sees everything a user can already see, so oversharing in SharePoint becomes oversharing in Copilot.

Google

Google says Workspace does not use customer data for training without prior permission or instruction. Gemini follows existing Workspace permissions, so it cannot read a file the user cannot open. Feedback that users choose to share may be used to improve products, which is worth covering in training.

How do you stop shadow AI?

You stop shadow AI by making the approved path easier than the unapproved one. Blocking alone pushes people to phones and personal laptops. Give employees a capable approved tool, then close the gaps around it.

• Buy business seats for the tools people already use, and connect them to single sign on.

• Block or warn on personal AI accounts at the browser or network layer for managed devices.

• Review identity provider logs and expense reports for AI subscriptions bought on cards.

• Publish a short approved tool list and a fast way to request new tools.

• Register every approved tool with an owner, the data classes it may receive, and its retention setting.

Agents need extra care because they act, not just answer. Give each agent its own scoped identity and log every query it runs. Our guide to data governance for AI agents covers MCP authorization and audit trails.

Which AI laws and rules apply to a business in 2026?

The rules depend on where you operate, what data you use, and whether AI influences decisions about people. Most general productivity use is lightly regulated. Hiring, lending, housing, insurance, and health decisions draw the most rules.

State laws

The Texas Responsible AI Governance Act took effect January 1, 2026. It bans certain uses, such as AI built to discriminate intentionally, and requires health care providers to disclose AI use in some contexts. Penalties range from 10,000 to 200,000 dollars per violation, and only the attorney general can enforce it.

Illinois amended its Human Rights Act through HB 3773, also effective January 1, 2026. Employers may not use AI in hiring, promotion, or discipline in a way that discriminates, and they must give notice. Using zip code as a proxy for protected classes is barred.

Colorado repealed its original AI Act before it took effect. Governor Polis signed SB 26-189 on May 14, 2026, which centers on transparency rather than discrimination audits. Deployers must notify consumers, explain adverse decisions within 30 days, and offer human review, starting January 1, 2027 if rulemaking is done.

California's privacy regulator finalized CCPA rules on automated decisionmaking in July 2025. Businesses using ADMT for significant decisions must comply by January 1, 2027. Risk assessment obligations follow by December 31, 2027.

Federal policy

Executive Order 14365, signed December 11, 2025, created an AI Litigation Task Force to challenge state AI laws. It also asks agencies to draft a uniform federal framework. The order does not itself preempt any state law, so existing state duties still apply. Sector laws such as HIPAA, PCI DSS, and GLBA also apply to any regulated data you put into an AI tool.

EU AI Act

If you sell into the EU, the AI Act applies on a staged timeline. The AI Omnibus was published July 24, 2026, and it moved Annex III high risk duties to December 2, 2027. Article 50 transparency rules generally apply from August 2, 2026. The AI literacy duty now asks providers and deployers to support literacy rather than guarantee a level.

Who is responsible when AI gets an answer wrong?

Your business is. In Moffatt v. Air Canada, a British Columbia tribunal held the airline liable for negligent misrepresentation after its chatbot gave wrong bereavement fare advice. The tribunal rejected the claim that the chatbot was a separate entity.

Set review rules by the stakes of the output. Internal drafts can go straight to the author. Anything sent to customers, used in a contract, or used in a decision about a person needs a human check against an approved source.

• Ground customer facing assistants in approved documents, and log which source each answer used.

• Give users a visible way to reach a human.

• Keep records of AI assisted decisions about people, since Colorado and California will require explanations and human review.

• Track lineage so you can trace a bad answer back to its data. Our guide to provenance and lineage covers the tooling.

How do you secure AI features you build?

Treat any AI feature you build as a new attack surface with its own threats. The OWASP Top 10 for LLM Applications 2025 is the standard checklist, and prompt injection sits first. Classic controls like multi factor authentication and encryption still matter, but they do not cover these risks.

The most common mistake is putting access rules in the prompt. A model told to hide salary data can be talked out of it. Enforce permissions in the data layer, so the model never receives data the user cannot see.

How do you train employees to use AI safely?

Train with short rules and real examples, then repeat. People remember a two page policy and three stories better than a long handbook. Under the EU AI Act, deployers must also support AI literacy among staff who use AI.

• Show what each data class looks like, using your own documents.

• Demonstrate a confident but wrong AI answer, so people learn to verify.

• Explain why a personal account is risky even for harmless looking work.

• Give one place to report a mistake, with no blame for early reporting.

• Refresh training when you add a tool or a new law takes effect.

How do you roll this out in 90 days?

Roll out in three phases of about a month each. Visibility comes first, then controls, then habits. Smaller companies can compress this into a few weeks.

The NIST AI Risk Management Framework is a useful structure to borrow, with four functions named Govern, Map, Measure, and Manage. It is voluntary and scales down to small teams. NIST says the framework is being revised under the White House AI Action Plan.

Once the basics hold, measure value as well as risk. Our list of enterprise AI automation platforms helps teams compare tools for the next round of use cases.

Frequently asked questions

How can a business use AI without risking its data?

Classify your data first, then match each class to an approved tool. Buy business plans whose terms say prompts are not used for training, and turn on single sign on and retention settings. Block personal accounts for work data. Keep regulated data such as health or payment records out of any tool that lacks the right contract, such as a business associate agreement.

Do business AI plans train on company data?

The major business plans say they do not by default. OpenAI says ChatGPT Business, Enterprise, and its API are not used for training unless you opt in. Microsoft says Copilot Chat prompts are not used to train foundation models. Google says Workspace data is not used for training without permission. Consumer plans can differ, so check which plan employees actually use.

What is shadow AI and why does it matter?

Shadow AI is employees using AI tools the company has not approved. The 2026 Verizon DBIR calls it the third most common non malicious data leakage activity. IBM found high levels of shadow AI added 670,000 dollars to the average breach cost. The fix is to offer approved tools that are good enough that people stop reaching for personal accounts.

Which AI laws apply to a US business in 2026?

It depends on where you operate and what AI decides. Texas TRAIGA and the Illinois AI employment amendment took effect January 1, 2026. Colorado replaced its AI Act with a transparency law starting January 1, 2027. California ADMT rules under the CCPA start the same day. Sector laws such as HIPAA still apply to any data you put into AI tools.

Does the federal government preempt state AI laws?

Not yet. Executive Order 14365, signed December 11, 2025, directs agencies to challenge and review state AI laws and to draft a federal framework. The order does not itself displace any state law. Until courts, agencies, or Congress act, businesses should keep complying with the state laws that apply to them.

When does the EU AI Act apply to businesses?

Prohibited practices have applied since February 2, 2025. The AI Omnibus, published July 24, 2026, moved obligations for Annex III high risk systems to December 2, 2027 and Annex I systems to August 2, 2028. Article 50 transparency rules generally apply from August 2, 2026. Any business selling AI driven services into the EU should map its use cases now.

Is a company liable for what its AI chatbot says?

Courts have treated it that way. In Moffatt v. Air Canada, a British Columbia tribunal held the airline liable after its website chatbot gave wrong advice about bereavement fares. The tribunal rejected the argument that the chatbot was a separate entity. Any customer facing AI output should be checked against approved sources and should have a clear human escalation path.

What security risks are specific to AI tools?

The OWASP Top 10 for LLM Applications 2025 lists prompt injection first, followed by sensitive information disclosure and supply chain risk. Excessive agency matters once AI can take actions. Most of these risks need controls outside the model, such as scoped permissions, output checks, and human approval for actions that move money, data, or customer commitments.

What should an AI acceptable use policy include?

List the approved tools and the plan each person must use. Define which data classes can go into which tool, with concrete examples. Require human review before AI output reaches a customer, a contract, or a hiring decision. Name who approves new tools and how employees report a mistake. Keep it short enough to fit on two pages.

Do small businesses need an AI governance program?

They need a lighter one, not none. A small business can cover most risk with an approved tool list, business plans with single sign on, a short data policy, and a review step for customer facing output. The NIST AI Risk Management Framework is voluntary and scales down well, so it is a sensible structure to borrow from.