NetworkDefender
NetworkDefender
Someone in your office is using AI right now. Maybe it is the bookkeeper summarizing a messy trial balance, or the paralegal cleaning up a first draft, or you at ten at night asking a chatbot to make a client email sound less annoyed. None of that is a problem by itself. The problem is what happens to the text after it gets typed in, and whether the tool it went into is one your firm actually chose.
This covers how widespread AI use already is, what genuinely goes wrong with client and company data, why the free version of a tool and the business version are different products, what Wisconsin law and your professional rules already require of you, and a checklist you can work through this week. AI is useful. The goal is keeping it useful without handing your clients' information, or your own firm's, to a system you do not control.
Adoption stopped being an open question sometime last year. The 2025 U.S. Chamber of Commerce and Teneo survey put small business AI use at 68 percent, up from 47 percent the year before among firms of 10 to 100 employees. Wisconsin is not behind on this. The Governor's Task Force on Workforce and Artificial Intelligence reports the state ranks first nationally in the employment weighted share of businesses using AI, and local business surveys have owners naming AI as their single biggest area of opportunity.
The number that should actually worry you is a different one. In the same Chamber research, roughly 77 percent of small businesses using AI have no written policy governing it. That gap has a name: shadow AI, meaning staff using tools through personal accounts the firm never approved and usually does not know about. A TELUS Digital survey of 1,000 U.S. employees in January 2025 found 57 percent of generative AI users had entered confidential company information into public tools, and 68 percent were doing it through personal accounts.
People do this because it works, not because they are careless. If nobody has told them where the line is, they draw it themselves, one paste at a time.
Here is what happens to a document after it goes into a consumer AI tool.
It may become training data. On free and personal paid tiers, what you type can be used to train the provider's next model, and it sits on the provider's servers in the meantime. Samsung is the case everyone cites. In 2023, engineers pasted confidential source code and internal meeting notes into ChatGPT, and the company banned the tool internally once it concluded the data was not coming back. Nobody hacked Samsung. Employees handed the material over trying to work faster.
The terms can change underneath you. On August 28, 2025, Anthropic updated its consumer terms so Free, Pro, and Max conversations would be used for model training unless the user opted out, with the setting defaulting to on, and stretched retention from 30 days to as long as five years for anyone who did not. Existing users had until October 8 to decide. Anthropic had been the privacy forward option up to that point. Whatever a consumer product promises you today can be revised in a product update, which is thin ground to build confidential work on.
Retention and subprocessors. Even where a vendor is not training on your input, it usually keeps prompts for some period, and it leans on other companies further down its supply chain that also handle the data. Every one of those is another party holding client and company information you remain answerable for.
The output can be confidently wrong. Generative AI invents case citations, figures, and sources that read perfectly. Lawyers around the country have been sanctioned for filing briefs built on cases that never existed. If a fabricated number lands in a tax memo or a coverage opinion, it is your mistake, not the software's.
And the same technology is being pointed back at you. KnowBe4 found 82.6 percent of phishing emails in its 2025 report contained AI generated content, rising to 86 percent in the 2026 edition, and Oxford research it cites clocked AI written phishing at a 54 percent click rate against 12 percent for conventional attempts. Voice and video deepfakes have moved into wire fraud too. A finance employee at the engineering firm Arup approved 15 transfers worth roughly 25.6 million U.S. dollars in 2024 after sitting through a video call in which every other participant was synthetic. Firms that move client money, whether through trust accounts, real estate closings, or premium remittances, are the natural target for that kind of thing.
This is the distinction most owners miss, and it matters more than anything else here. The consumer app and the business app can share a name, a logo, and an interface while operating under completely different contracts.
Consumer tiers, meaning free plans and personal subscriptions like ChatGPT Plus, personal Gemini, or Claude Pro, may use your input for training unless you go hunting for the setting. They retain it longer and give you no meaningful contractual protection. They are not suitable for client or company data.
Business and enterprise tiers commit in writing not to train on your content:
Two things follow from that list. First, on business associate agreements. A BAA is the written contract HIPAA requires before a vendor can handle protected health information. Microsoft, Google, and Azure OpenAI will sign one for qualifying enterprise services. Consumer tiers will not, and no toggle inside a personal account changes that. If your agency writes health lines or your firm touches medical records, this is the deciding factor.
Second, buying the business tier is necessary but does not finish the job. Microsoft 365 Copilot answers from whatever the user already has permission to open. If your folder permissions are loose, as they are in most small firms running on the same file server they set up a decade ago, Copilot will cheerfully surface payroll or partner compensation to whoever asks it. The tool has to be configured, not just purchased.
Generic AI safety advice skips this part, and it is where the real exposure for a Wisconsin professional firm sits. Which rules apply depends on the kind of firm you run, so the ones below are grouped by profession. No law bans AI, but several of the rules you already follow require you to control where your data goes.
Start with what Wisconsin does not have. There is no broad state law governing business use of AI. The two AI specific statutes on the books are narrow. One covers synthetic media in political ads (2023 Wisconsin Act 123), the other non-consensual deepfake intimate images (2025 Wisconsin Act 34), and neither touches your practice. A 2026 bill on AI deepfake harassment and fraud, Senate Bill 815, stalled in committee. So the rules that actually bind you are the ones you already answer to.
For every firm, Wisconsin's breach notification law (Wis. Stat. § 134.98). This is the baseline, and it applies to almost everyone. If unencrypted personal information in your possession is acquired by someone unauthorized, you have to notify the affected people within 45 days. Here is where AI comes in. Paste a client's name next to a Social Security number, financial account number, or driver's license number into a consumer AI tool, where the provider and its subprocessors keep it, and you have a real argument that an unauthorized disclosure just happened. GLBA covered financial institutions can follow the federal notification path instead, as our post on Wisconsin CPA requirements explains. Most general practice law firms cannot.
For accounting firms, GLBA, the FTC Safeguards Rule, and the IRS WISP. The Safeguards Rule (16 CFR Part 314) treats an AI vendor like any other service provider. You have to oversee it by written contract, run a risk assessment, and keep an inventory of where customer data goes. So the moment client data goes into an AI tool, that tool has to show up in your risk assessment, your data inventory, and your written information security plan. Two more rules stack on top. The AICPA Confidential Client Information Rule (ET § 1.700.001) requires client consent before you disclose confidential information, and putting it into a third party tool counts as a disclosure. For tax work, IRC § 7216 adds criminal exposure for disclosing taxpayer information without consent, and AICPA SSTS No. 1.3 requires reasonable efforts to protect what you share.
For insurance agencies, the Wisconsin Insurance Data Security Law (Wis. Stat. §§ 601.95 through 601.956). Since November 1, 2023, WIDS has required due diligence on and oversight of any third party vendor handling nonpublic information, and an AI vendor is squarely one of them. It also runs on a much tighter clock. You get three business days to notify the Office of the Commissioner of Insurance after a breach, far tighter than the 45 days under the general statute. The NAIC's model bulletin on insurer use of AI, which Wisconsin is expected to move toward, would tighten vendor governance further.
For law firms, SCR 20:1.6(d) and SCR 20:1.1. Two Supreme Court rules apply. The confidentiality rule requires reasonable efforts to prevent the unauthorized disclosure of client information, and the competence rule, through its technology comment, expects you to understand the risks of the tools you use. ABA Formal Opinion 512, issued July 29, 2024, puts it plainly: before putting client information into a self learning AI tool, a lawyer has to evaluate whether it could be exposed to or accessed by others. Wisconsin has not issued its own formal AI opinion, but its ethics counsel treats the existing rules, including SCR 20:5.3 on supervising nonlawyer help, as enough to govern AI use.
And federally, the backdrop behind all of it. There is no comprehensive federal AI law binding a firm your size. But the NIST AI Risk Management Framework and its generative AI profile (NIST AI 600-1) have become the yardstick regulators point to, so building against it now tells the same reasonable efforts story your other obligations already ask for.
One thread runs through all of it. None of these rules ban AI. Every one asks the same four things. Know where your client and company data goes, choose your vendors on purpose, write down why, and be able to move fast when something slips. Staff quietly using a consumer chatbot fails all four.
None of this takes an enterprise budget. It takes deciding, and then writing the decision down.
For firms carrying the heaviest confidentiality duties there is a further option worth knowing about. Private or tenant isolated AI means the model runs inside an environment you control, or in a dedicated space walled off by contract, so prompts and documents never reach a shared public model at all. If you handle privileged litigation files, sensitive transactional work, or real volumes of health information, this is the cleanest way to satisfy the reasonable efforts and vendor oversight standards without giving up the productivity. Most firms do not need it, since a properly configured business tier covers them. Where confidentiality obligations run highest, it takes the third party question off the table entirely. NetworkDefender is building private, secure AI options for exactly this kind of firm.
Banning AI outright does not work. It just moves the activity onto personal phones and personal accounts, where you have no logs, no contract, and no idea what left the building. The firms that get this right treat AI like any other vendor with access to client and company data. They pick business tier tools with real terms, switch off training, write down the rules, teach their staff, and keep the paperwork showing they thought it through.
That is the same discipline Wisconsin's breach law, the Safeguards Rule, WIDS, the AICPA standards, and SCR 20:1.6(d) already expect from you everywhere else in your practice. Handle it once and the productivity is yours to keep.
The firms that handle this well are not the ones with the biggest technology budgets. They are the ones that govern their AI the way they document their client work.
NetworkDefender Security helps Wisconsin CPA firms, law firms, and insurance agencies adopt AI safely. We write AI acceptable use policies, configure Microsoft 365 and Google Workspace tenant controls, run vendor due diligence, and fold AI into your existing security program and incident response plan. Contact us to schedule a security review.