How to Use AI With Client Data in 2026: A Step-by-Step Compliance Guide for Professionals

In the first months of 2026, several things changed how professionals should think about AI. On January 1, a generative AI exclusion from the Insurance Services Office (ISO) became available for general liability policies, joining broader AI exclusions some carriers already use on management and professional liability lines.^{1, 2} FINRA's 2026 Annual Regulatory Oversight Report discussed the risks of AI agents for the first time.³ In February, OpenAI patched a ChatGPT vulnerability that researchers showed could quietly exfiltrate conversation data through a hidden network channel.⁴ In May, AI evaluation startup Braintrust disclosed unauthorized access to a cloud account holding customer API keys and told every customer to rotate them.⁵

If you use AI tools with client data, this is your problem. The good news is that the path forward is clear and the steps are well-defined. Here is the complete 2026 guide to using AI with confidential client information without violating compliance obligations, losing professional privilege, or discovering — too late — that your insurance policy excludes the very claim that was just filed against you.

The Three Risks You Are Actually Running

When professionals use mainstream AI tools with client data, three things go wrong at once. Most firms are aware of one or two. Almost none have addressed all three.

1. Compliance risk. Regulators and professional bodies — FINRA, the SEC, the AICPA, state bar associations — have made clear that existing confidentiality, supervision, and data-protection obligations apply to AI tools, and a growing number of states are adding AI-specific laws on top.

2. Privilege risk. In February 2026, in United States v. Heppner (S.D.N.Y.), Judge Jed Rakoff held that documents a criminal defendant generated with a consumer AI platform, on his own and without his lawyers' direction, were not protected by attorney-client privilege or the work product doctrine. The court pointed to the platform's privacy policy, which allowed it to collect user inputs and disclose them to third parties, as showing the communications were not confidential.⁶ The ruling turned on its facts — a federal court in Michigan reached a different result on work product the same month in Warner v. Gilbarco⁷ — but it shows how a court may treat confidential material typed into a public AI tool.

3. Insurance risk. ISO endorsement CG 40 47 01 26, available to carriers since January 1, 2026, excludes bodily injury, property damage, and personal and advertising injury arising out of generative AI from commercial general liability coverage wherever an insurer attaches it.¹ W.R. Berkley's Form PC 51380 goes further: an “absolute” AI exclusion for D&O, employment practices, and fiduciary liability policies that reaches claims arising from any actual or alleged use of AI.² If you have not checked your policies for AI language, your coverage may not respond to the very category of claim AI use can create.

Step 1: Inventory Your Firm's Actual AI Usage

Before you can govern AI use, you have to know what is happening. Most firms radically underestimate their exposure because employees use AI tools without formal authorization — a phenomenon now widely called “shadow AI.”

The scale of shadow AI is significant. Clio's 2024 Legal Trends Report found that 79% of legal professionals were using AI in some capacity, up from 19% in 2023.⁸ In a 2025 study by KPMG and the University of Melbourne, 44% of US workers said they had used AI tools at work without proper authorization or in inappropriate ways, and 46% admitted uploading sensitive company information and intellectual property to public AI platforms.⁹

How to inventory:

1. Send a confidential, anonymous survey to all staff asking: which AI tools do you use weekly, what types of work do you use them for, and have you ever input client information?
2. Review browser history reports if your firm uses a managed device program (focus on chatgpt.com, gemini.google.com, claude.ai, perplexity.ai).
3. Check expense reports for AI tool subscriptions — many employees are paying for personal accounts to use at work.
4. Inspect any browser extensions installed firm-wide — in January 2026, OX Security researchers reported two Chrome extensions with more than 900,000 combined users that were sending users' ChatGPT and DeepSeek conversations to remote servers.¹⁰

Document what you find. You cannot create defensible governance without an honest baseline.

Step 2: Map Your Regulatory Obligations

Your obligations depend on your profession, your jurisdiction, and the type of client data you handle. The major sources for 2026 are:

For Financial Advisors and Broker-Dealers

FINRA 2026 Annual Regulatory Oversight Report (December 2025) — Includes a standalone generative AI section that expands on the previous year's coverage and, for the first time, discusses the risks of AI agents: systems that can autonomously perform tasks on a user's behalf. FINRA emphasized that its existing rules, including Rule 3110 on supervision, continue to apply when firms use generative AI, just as they apply to any other technology, and suggested firms consider keeping prompt and output logs.³

SEC's amended Regulation S-P (compliance required from December 3, 2025 for larger entities and June 3, 2026 for smaller entities) — Requires broker-dealers, registered investment advisers, investment companies, and transfer agents to maintain an incident response program that includes oversight of service providers through due diligence and monitoring, and to notify affected individuals of breaches involving sensitive customer information. An AI vendor that receives customer information is a service provider, and a consumer AI tool with no contractual commitments is hard to fit into that oversight.¹¹

For Attorneys

ABA Formal Opinion 512 (July 2024) — Says lawyers must obtain informed client consent before inputting information relating to a representation into a generative AI tool with self-learning capabilities, and that boilerplate language in an engagement letter is not enough.¹² State bars have issued their own guidance, including Florida (Ethics Opinion 24-1, January 2024) and California (practical guidance first published in November 2023 and updated in May 2026).^{13, 14} Using AI without understanding how a tool handles client information puts a lawyer's duties of competence and confidentiality at risk.

In January 2026, the ABA's Law Technology Today published a practical checklist for using AI responsibly in law firms, covering tool selection, data security, human review of AI output, and client disclosure.¹⁵

For CPAs and Tax Professionals

AICPA Statements on Standards for Tax Services (SSTS) — Section 1.4 defines a “tool” to include artificial intelligence, and members remain responsible for exercising due professional care when relying on one. Section 1.3 requires reasonable efforts to safeguard taxpayer data, including data transmitted or stored electronically.¹⁶ Tax professionals retain all professional obligations whether or not they use AI.

IRC §7216 — Makes it a misdemeanor for a tax return preparer to knowingly or recklessly disclose or use information furnished in connection with preparing a return, except as the law permits. Penalties include a fine of up to $1,000, up to one year in prison, or both.¹⁷ Whether entering client tax information into a consumer AI tool is a prohibited disclosure has not been settled: the regulations permit certain disclosures to providers of auxiliary services, but commentators have flagged external AI tools as an open compliance question.¹⁸

For Healthcare Professionals

HIPAA Security Rule (proposed revision, January 2025) — The Department of Health and Human Services Office for Civil Rights proposed the first update to the Security Rule since 2013, citing the rise in cyberattacks on health care. The proposal would remove the distinction between “required” and “addressable” implementation specifications and introduce stricter expectations for risk management, encryption, and resilience for any system that handles ePHI, AI tools included. As of September 2026 it is still a proposed rule; the federal regulatory agenda now lists final action for July 2027.^{19, 20}

State Laws

The Colorado Artificial Intelligence Act, one of the first comprehensive state AI laws, was delayed and then rewritten in May 2026: the revised law drops the original duty of care and impact assessments in favor of notice and consumer-rights requirements for automated decision-making technology, effective January 1, 2027.²¹ California's CCPA regulations on automated decision-making technology, risk assessments, and cybersecurity audits took effect January 1, 2026, with the automated decision-making requirements applying from January 1, 2027.²² Illinois HB 3773 (effective January 1, 2026) prohibits discriminatory use of AI in employment decisions and requires employers to notify workers when AI is used in those decisions.²³ All three introduce new transparency and accountability requirements.

Step 3: Read Your Insurance Policy — Today

This is the step most firms skip. It is also the most urgent.

AI exclusions are now appearing on both general liability and management liability policies. Two forms are worth knowing by name:

• ISO Form CG 40 47 01 26 — A generative artificial intelligence exclusion endorsement, available to carriers from January 1, 2026, that removes coverage for bodily injury, property damage, and personal and advertising injury arising out of generative AI from general liability policies.¹
• W.R. Berkley Form PC 51380 — An absolute AI exclusion spanning D&O, employment practices, and fiduciary liability policies.²

As exclusions spread, insurers are also starting to sell coverage back. On April 2, 2026, Gartner advised general counsel to evaluate new “affirmative AI insurance” offerings as insurers add AI exclusions to traditional policies, and predicted that by 2030 insurers will require strong AI risk controls before offering that coverage.²⁴ Munich Re's HSB introduced AI liability insurance for small and mid-sized businesses in March 2026, designed to cover some AI-related losses that general liability policies may exclude.²⁵

What to do right now:

1. Pull your current E&O and general liability policies.
2. Search the policy documents for “artificial intelligence,” “machine learning,” or “generative” — be thorough; some exclusions are buried in defined terms.
3. If you find an exclusion, contact your broker the same day and ask: (a) when did this take effect, (b) what AI endorsement options are available, (c) what governance evidence is required for the endorsement.
4. If your policy renews in 2026, expect this to be a contested negotiation point. Plan for it.

Step 4: Build Your Tool Tier System

Once you understand your obligations, the question becomes which AI tools you can use for which tasks. The framework that works for most professional services firms is a three-tier system based on what happens to your data.

Tier 1: Public Consumer AI

ChatGPT (free), Gemini (consumer), Claude.ai, Perplexity, and similar.

What happens to your data: Transmitted to the provider, processed on their servers, potentially used for training, retained according to the provider's policy, and subject to subpoena. In May 2025, in the New York Times copyright case, a federal magistrate judge ordered OpenAI to preserve ChatGPT output logs that would otherwise have been deleted. That broad order ended on September 26, 2025, but OpenAI must still retain some logs, including those tied to accounts the Times flagged — a reminder that data held by a vendor can be swept into litigation.^{28, 29}

Approved use: Research and drafting that contains zero client- specific information. General learning. Template-style content with no client identifiers.

Prohibited use: Anything involving a client's name, account details, financial information, medical information, legal strategy, tax return data, or any other information that could identify the client or constitute a disclosure under your professional standards.

Tier 2: Enterprise AI With Data Protection Contracts

ChatGPT Enterprise, ChatGPT Business (formerly ChatGPT Team), Claude for Enterprise, Microsoft 365 Copilot, and similar enterprise tiers.

What happens to your data: Still transmitted to the provider, but with contractual protections against training, retention, and unauthorized use. Typically includes a Data Processing Agreement (DPA) or, for healthcare, a Business Associate Agreement (BAA).

Approved use: Client data is permitted IF you have verified the following: (a) a signed contract with zero-retention and no-training clauses, (b) the contract addresses your specific regulatory obligations (IRC §7216, HIPAA, Reg S-P, state laws), (c) your engagement letters disclose AI use, and (d) you have logged usage for any audit trail your insurer or regulator may ask for.

Verification required: “We do not train on your data” is different from “We do not retain your data.” Both protections are needed. The contract should explicitly address both.

Tier 3: On-Device AI

Applications that run the AI model entirely on your local hardware. The model file lives on your computer; inference happens on your CPU or GPU; data never leaves the device.

What happens to your data: Nothing leaves your device. There is no transmission, no retention by a third party, and no possibility of subpoena-able vendor records.

Approved use: The strongest fit for any client work. With no third-party disclosure, the vendor risks addressed in Tiers 1 and 2 don't arise — though the device itself, and anything the tool saves on it, still needs your normal safeguards. On-device AI converts the compliance question from “does my vendor agreement satisfy these requirements?” to “is there a vendor at all?” — and when the answer is no, the position is far simpler to defend.

Step 5: Update Your Engagement Letters

Your engagement letters should disclose AI use and clarify the scope of your professional responsibility. The exact language depends on your profession, but the core elements are consistent.

Sample disclosure language (adapt for your jurisdiction and engagement type):

“In providing services under this engagement, [Firm] may use artificial intelligence (AI) tools to assist with research, drafting, analysis, or related tasks. [Firm] uses AI tools that are subject to written data protection agreements that prohibit the AI vendor from using your information for model training and that require deletion of inputs after processing, OR [Firm] uses AI tools that operate entirely on [Firm]'s local hardware so that your information is not transmitted to any third party. [Firm] reviews all AI-generated content before providing it to you and remains fully responsible for the work product delivered.”

For tax engagements, add: “[Firm] will not disclose tax return information to any AI service provider in a manner inconsistent with IRC §7216.”

For attorneys, ABA Opinion 512 requires informed consent before using self-learning AI tools with client information. The disclosure should be sufficiently specific that the client can make an informed decision.¹²

Step 6: Write a Firm AI Usage Policy

Your AI policy does not need to be long. It needs to be clear and enforceable. At minimum, it should specify:

1. Approved tools by tier — list the specific tools approved for Tier 1, Tier 2, and Tier 3 use. Forbid all others without written approval.
2. Data classification — define what counts as confidential client data (it should include anything that identifies a client or reveals matters relating to the representation).
3. Allowed and prohibited uses by tier — make the rules concrete (e.g., “Tier 1 tools may not be used to summarize, analyze, or draft anything involving a client's name, account, or matter.”).
4. Verification and review — all AI-generated content must be reviewed by a qualified professional before it is used in client work or delivered to a client.
5. Audit trail — record what tools were used for what tasks. An insurer offering AI coverage may ask for this; your regulator may too.
6. Incident response — define what to do if confidential information is accidentally submitted to a non-approved tool (notification, mitigation, documentation).
7. Quarterly review — the AI landscape moves fast. The policy should be reviewed and updated at least quarterly.

Step 7: Train Your Team

Policies that are not understood are not followed. Train at least once a year, and again whenever your approved tool list changes.

Effective training covers: the three risks (compliance, privilege, insurance), the tier system and what tools fall in each tier, what data triggers tier restrictions, how to use the approved Tier 2 and Tier 3 tools for everyday work, what to do if a mistake happens, and how to log usage for the audit trail.

Training documentation matters. Maintain attendance records and assessment results. If a regulatory action or insurance claim arises, “everyone knew” is not sufficient; “everyone completed the training on this date with this score” is.

Step 8: Audit Quarterly

Repeat Step 1 every quarter. Survey staff. Review usage logs. Check for new AI tools that have appeared in browser histories or expense reports. Update the policy based on what you find.

In Deloitte's 2026 survey of 3,235 business and IT leaders, only 21% said their organizations have a mature governance model for agentic AI, while 74% expect to be using AI agents at least moderately within two years.³⁰ Professional services firms that close that gap now will have a defensible position when the next round of regulatory scrutiny or insurance underwriting arrives.

The Shortcut: When On-Device AI Solves the Whole Problem

Steps 1 through 8 are the comprehensive answer. The shortcut is recognizing that on-device AI — Tier 3 in the framework above — converts most of the problem into a non-problem.

When AI runs entirely on your Mac or workstation, there is no third party receiving client data. That eliminates the IRC §7216 disclosure question. It eliminates the Reg S-P vendor contract question. It eliminates the ABA Opinion 512 informed-consent question (the standard turns on disclosure to AI tools with self-learning training capabilities; on-device models do not train on your inputs). It removes the fact the Heppner court relied on: a third-party platform whose terms let it collect and disclose what you typed. And on the insurance side, it gives you a far simpler usage profile to describe to an underwriter: in-house software with no external transmission, rather than cloud AI with potential third-party exposure.

This is not a claim that on-device AI is appropriate for every workflow. Large document review at scale, cross-firm collaboration features, and certain specialized research tasks may still require Tier 2 enterprise tools with the right contracts in place. But for the vast majority of day-to-day professional work — explaining a clause to a client, summarizing a financial statement, drafting a routine letter, analyzing a confusing IRS notice — on-device AI provides a level of compliance certainty that no contractual guarantee can match.

The 2026 regulatory and insurance landscape did not so much create new obligations for professionals as clarify that existing obligations apply just as fully to AI tools as they have always applied to any other technology that touches client data. The firms that build sound AI governance now — grounded in the actual architecture of the tools they use, not just the marketing — will be positioned to use AI competitively while their less prepared peers face client losses, regulatory exposure, and uninsured liability claims.

Step 1 is to find out what is actually happening in your firm. Start there.