On-Prem AI for CPAs and Tax Preparers: Disclosed Data Flow, Explained
· 7 min read · Data & Privacy
Quick answer. For a CPA or tax preparer, W&S discloses exactly where the AI that drafts your client emails and summarizes documents sends that data: a named provider, documented before you sign. On-premise inference is on our roadmap. You own every credential, and a person on your team approves every message before it sends.
Every accountant we talk to asks a version of the same question before anything else: where does my client data go? It is the right instinct. The information that flows through a tax practice is not ordinary business correspondence. It is the most revealing data a person hands to anyone, and a client hands it to you because they trust you to hold it carefully. That is the lens we want to use to look at AI tools, because the answer to “where does the data go” is different depending on how the AI is built.
Why tax data is different
A retailer’s inbox holds order numbers and shipping questions. A tax practice’s inbox holds Social Security numbers, W-2s, K-1s, brokerage statements, mortgage details, and a complete financial picture of a household. Put together, the files on your network describe your clients more completely than almost any other record they will ever produce.
That is why confidentiality is something a firm cares about for its own reasons, long before anyone outside the firm weighs in. A client who sends you their return is trusting you with information they would not share with their neighbor or their employer. When a firm starts thinking about putting AI anywhere near that material, the first question is not “is it fast,” it is “does this change who can see my clients’ information.” A good tool should make that question easy to answer.
Where client data goes with the usual AI tools
Most AI products you can sign up for in a browser are cloud services. When you paste a client document or forward an email into them, that content travels to servers you do not control, run by a company you have no operational visibility into. Some consumer-grade tools may use the things you type as input to improve their own models, which means your client’s data could become part of a system you cannot inspect. By default, OpenAI may use content from personal ChatGPT accounts to improve its models unless you opt out; its business and enterprise tiers are excluded (OpenAI Help Center).
The risk grows when staff reach for those tools without anyone signing off. IBM found that breaches involving shadow AI, the unsanctioned AI tools employees adopt without approval, cost an average of $670,000 more in its 2025 Cost of a Data Breach Report (IBM). Smaller firms feel this acutely: Verizon’s 2025 Data Breach Investigations Report found that 88% of breaches at small and midsize businesses involved ransomware (Verizon).
The deeper issue is that you cannot un-send it. Once a Social Security number or a full financial statement has left your network and reached a vendor’s cloud, you are relying on that vendor’s policies, settings, and good behavior. For a casual draft, that may be a tradeoff a business is willing to make. For a tax client’s most private records, a lot of firms look at that and decide they want a different arrangement. We wrote a whole piece on where your client data goes when you use ordinary AI tools, if you want the longer version.
The on-prem alternative
Disclosure means a named provider, in writing, before you sign, instead of a policy you have to trust. Your client documents are sent to that disclosed provider only to complete the task. If you want the plain-English version of what on-prem AI means, start there.
The reason this is different is that privacy here is a matter of disclosure, not just policy. With a typical cloud tool, your data stays private because a vendor promises it will, with no way to check. With W&S, you get a named provider and a documented path you can verify yourself. This is the core of an AI assistant built for a practice that handles sensitive files.
What the assistant does for a tax practice
W&S sets up an AI assistant around your practice. It handles the routine pile that eats a preparer’s day:
- Drafts client replies in your firm’s voice, so a status question or a document request gets a written answer in seconds instead of minutes.
- Summarizes returns, IRS notices, and long threads into a short readout, so you can see what a forty-message chain or a dense notice actually says without rereading all of it.
- Answers plain questions about your own files and gives you a citation to the source page, so you can click straight to where the answer came from and verify it yourself.
- Handles intake and data entry, pulling details off the documents clients send and putting them where they belong.
- Stops before every send. The assistant drafts, then waits. A person on your team reads the draft and clicks send. It never sends on its own, and that gate is enforced by the system, not by a setting someone could quietly turn off.
Confidentiality is your duty; possession is the tool
Preparers carry real confidentiality obligations around client tax information, and the firms we work with take those obligations seriously. We want to be precise about what on-prem does and does not do here, because the honest framing matters. On-prem AI is not a rule you have to follow, and we are not telling you it makes your firm compliant with anything or that any standard requires it. W&S builds and installs software. We are not your accountant or your attorney, and Edward is not a CPA.
What disclosure gives you is a documented data map. Alongside setup you get a written audit runbook, so the person on your team responsible for confidentiality can verify where the data goes and confirm it for themselves. The duty is yours. The tool gives you a documented way to check the data flow, so you can meet that duty on your own terms instead of taking a vendor’s word for it.
What we verify at install
When we set up your system, we do not just configure it and leave. We verify, and document, that the system behaves the way the pitch says it does:
- Exactly where your data flows. We name the disclosed AI provider and write down the path, so client documents and messages are never used to train anyone’s model.
- Your data is stored encrypted, per client. The assistant keeps your data in an encrypted database separated per client, for the length of the engagement, so it can reference prior context. Ask us to delete it and we do.
- You own the credentials. If you ever stopped working with us, you keep the runbook.
- Support is 24 hour response, next business day resolution target, remote-first. We respond within 24 hours, aim to resolve by the next business day, and handle most of it remotely.
Today, AI drafting runs through a disclosed cloud provider named in your engagement, with the data path documented before you sign. On-premise inference is on our roadmap, and we will tell you plainly which one applies to your deployment.
What it costs
We keep the pricing simple. A first project starts at $1,500, one task at a fixed price with an end date, done in person. A next project starts at $3,000, one time. Monthly help starts at $300 per month, scaled to the size of what you are running.
If you handle tax files for a living, where that data sits is not a detail. It is the whole question. See how this works for accounting firms, or book a short walkthrough and we will look at your real workflow together.
Edward Ahrens is a co-founder of W&S Consulting, where he builds and configures private AI systems for data-sensitive small firms around Atlanta, with disclosed data flow and on-premise inference on the roadmap. He is a software builder, not a CPA or attorney.
AIDataPrivacyAccounting
Want this handled for your office.
The assistant drafts the replies. A person on your team approves every send.
You will talk to Edward. 15 minutes. The call is free.
