Private GPT for CPA, Law and Medical Practices: AI Without Exposing Client or Patient Data
Private AI inside your own AWS, Google or Microsoft cloud account, or on a Mac mini in your office, with our redaction engine built in, so client and patient data never goes into a public chatbot.
What our clients say
Who private GPT is for
CPAs, financial advisors, law offices and medical practices hold confidential client and patient data and often cannot paste it into a public chatbot. Staff sometimes do it anyway. A private setup gives them a safe, approved place to use AI, and keeps your work from depending on a single public AI vendor, whose accounts can be closed without warning.
Rules differ by profession: IRC section 7216 and the FTC Safeguards Rule for tax preparers, GLBA, Regulation S-P and FINRA for financial firms, confidentiality rules for lawyers, HIPAA for medical practices, SOX controls for firms serving public companies. See what each one asks and how a private GPT helps.
Isn’t a ChatGPT or Claude Team plan enough?
Team and business plans are a real step up from free accounts: by default your business data is not used to train their models, and you get admin controls. But they do not stop the round trip. Every prompt and file still travels to the AI company’s servers and back, with your client details in it.
- Prompts and uploaded files go to the AI company’s servers, are processed there and kept in chat history under that company’s retention rules.
- Names, SSNs and account numbers go out as typed. Nothing masks them first.
- Staff can upload whole folders or connect other apps; the record of what was sent sits with the vendor, not with you.
- Team plans generally do not come with a HIPAA Business Associate Agreement, so they should not be used with patient information.
- You still have to judge whether that vendor relationship fits rules like §7216, your engagement letters and client consents.
- Your firm’s AI depends on one vendor account, and accounts can be closed without warning.
- The model runs in your own AWS, Google Cloud or Azure account, under the agreement you already have, or on a Mac mini in your office.
- Our redaction engine masks client identifiers before any prompt, file or web search leaves your app.
- Your logs, your retention rules, your encryption keys.
- Access limited to the folders and staff you approve.
- You choose the models and can switch providers without starting over.
Enterprise plans can add options such as custom data retention. Plans and terms change often, so we compare your current tools with you in the free consultation.
Compliance rules and how a private GPT helps
No AI tool makes a firm compliant on its own. What a private GPT does is remove the biggest gap most firms have today: client data going into tools the firm does not control. Here is how it supports the rules our clients work under.
IRC §7216
Who: tax return preparers.
Asks: no use or disclosure of tax return information without the client’s consent, except as the rules allow.
Helps by: keeping return data inside your account and masking identifiers before anything reaches a model or web search.
FTC Safeguards Rule
Who: tax preparers, accounting firms and other non-bank financial businesses (under GLBA).
Asks: a written information security plan (WISP), access controls, encryption, MFA, logging and oversight of vendors.
Helps by: giving your WISP a documented AI setup: who can use it, what it can see, encryption and logs.
GLBA, Reg S-P and FINRA
Who: financial advisors, broker-dealers and investment firms.
Asks: protect customer information, respond to incidents, and keep required business records.
Helps by: data in your own environment, access logs, and records you control and can retain as your rules require.
SOX
Who: public companies, and firms that audit or support them.
Asks: reliable internal controls over financial reporting, including IT controls: access, change management, review.
Helps by: role-based access, audit trails, versioned and tested rule changes, and “every output is a draft” human review. SOX is about controls; we support them, we do not certify them.
Attorney confidentiality
Who: law firms.
Asks: Florida Bar Rule 4-1.6 and Ethics Opinion 24-1, and ABA Formal Opinion 512: protect client information and understand how an AI tool handles it.
Helps by: no client data with a chat vendor, masking before any model, and a written AI policy for the firm.
Florida FIPA
Who: any business holding Floridians’ personal information.
Asks: reasonable measures to protect personal information, and breach notice within 30 days.
Helps by: less personal data in outside tools, so there is less to expose and less to report.
HIPAA
Who: medical, dental and therapy practices, and their business associates.
Asks: a Business Associate Agreement (BAA) with every vendor that handles patient information, plus a risk analysis, access controls, audit logs and encryption.
Helps by: using only HIPAA-eligible services on AWS, Google Cloud or Microsoft under your BAA with them, or an on-prem Mac mini; redaction set up for patient identifiers; and our team works with test data, signing a BAA with your practice where our work involves patient information.
SOC 2
Who: not a law. Clients and partners often ask vendors for a SOC 2 report.
Asks: proof of security controls, checked by an auditor.
Helps by: building on AWS, Google Cloud or Microsoft, which publish their own SOC 2 reports, with your setup documented for your security questionnaires.
This is general information, not legal advice. We work with your compliance advisor, auditor or attorney, and we do not claim that any setup makes your firm compliant.
Deployed in your cloud account, not ours
The model runs inside your firm’s own cloud account, under your logins, your region and your encryption keys. We usually pick the cloud you already use: Microsoft 365 firms often fit Microsoft Foundry, Google Workspace firms Vertex AI, and firms already on AWS fit Bedrock. All three offer HIPAA-eligible AI services under a Business Associate Agreement.
AWS Bedrock
Models such as Claude, Llama and Amazon Nova, called from your AWS account.- Private VPC endpoints (PrivateLink), so traffic stays off the public internet
- IAM roles and CloudTrail audit trail
- Bedrock Guardrails to mask or block sensitive data
- Encryption with your own KMS keys
Vertex AI
Gemini, Claude and open models from Model Garden, in your Google Cloud project.- VPC Service Controls around your data
- Customer-managed encryption keys and US data location
- Sensitive Data Protection to find and mask PII
- Cloud Audit Logs
Microsoft Foundry
Formerly Azure AI Foundry. OpenAI GPT models and others, in your Azure tenant.- Private endpoints and Microsoft Entra ID sign-in
- Azure AI Content Safety and PII detection
- Azure Monitor logs
- We review abuse-monitoring retention with you
On-prem Mac mini
Open-weight models on Mac mini servers in your office, with no cloud account at all. See how it works.- Prompts and files stay in your building
- Can run fully offline
Each provider says, in its terms, that your prompts and outputs are not used to train its models. Terms and settings change, so we review them with you before go-live and note where any data is kept and for how long.
On-prem private GPT on a Mac mini
For data that should never leave the office, we set up Apple Silicon Mac minis as small AI servers. Their shared memory runs open-weight models well, they are quiet and low-power, and they sit on your network like any other office machine.
Open-weight models
Models such as Llama, Qwen, Gemma, Mistral and OpenAI’s gpt-oss, served with tools like Ollama or LM Studio. You own the setup; no per-message fees.Sized to the job
A Mac mini fits small and mid-size models for drafting, summaries and document search. For larger models or more staff, we add machines or step up to a Mac Studio.Locked down
Disk encryption, staff logins through a private chat page on your network, no access from the internet, and backups and updates managed by us.Cloud plus on-prem
Many firms use both: restricted files stay on the Mac mini, and everyday work goes to your private cloud models, which are more capable.Local models are smaller than the best cloud models. We test them on your real tasks, with sample files, before you commit.
How we safeguard client data
Several layers, so one mistake does not expose a client file.
Private network
The app talks to the model over private endpoints inside your cloud. No public chatbot, no shared vendor account.Who can see what
Single sign-on with your existing logins, roles per team, and the assistant can only read the folders you approve.Encryption
Data is encrypted in transit and at rest, with keys held in your account, so you can revoke access.Redaction before the model
Names, SSNs, account numbers and other identifiers are swapped for placeholders before the prompt leaves your app.Clean logs
Logs keep who did what and when, not the raw client text. Short retention, locked-down access.Written policy
A short AI use policy, an approved-tools list and staff training, so people know what may go into AI.Our redaction engine: what outsiders never see
We built our own redaction engine for accounting, CPA, law and medical work, where client and patient files are full of sensitive details. It is included at no extra cost with every private GPT we set up, runs inside your cloud account or on your Mac mini, and handles three jobs.
File scrubbing
Makes a redacted copy of client files (PDFs, spreadsheets exported to CSV, text) before they go to AI, a vendor or a shared folder. The original is never changed.Safe web search
When staff or the AI need to look something up, names, numbers and client details are masked first, so only the general question leaves your network.Prompt and answer masking
Identifiers are swapped for placeholders before the model sees a prompt, and put back only in the answer shown to the signed-in staff member.The engine is tuned for professional work: it masks who a client is, but keeps what the work needs, such as tax years, form numbers and amounts, so answers stay useful. Here is how a prompt looks before and after.
[CLIENT_1], SSN [SSN_1], about the $4,200 refund on account [ACCT_1].Log redaction
Logs are where client data often leaks: app logs, error messages, monitoring tools and the cloud provider’s own request logs. We set each one up on purpose.
What is logged
User, time, model, document IDs, token counts and redaction counts. Not the raw prompt or file text.Provider request logs
Kept off unless you need them. If on, they go to an encrypted, access-restricted store with the same redaction applied and a set retention period.Tested, not assumed
We run a test set of fake client records through the system, measure what slips past, tune the rules and repeat the test each quarter.No redaction tool catches everything. That is why the design assumes a miss can happen: private network, no raw text in logs, short retention and fail-closed rules.
How a private GPT runs day to day
Lessons from our deployments. The technology is half of it; the other half is firm rules the AI has to follow every time.
A firm rulebook
One set of written rules loads with every chat: what the AI may read, where it may write and what it must never do. Client notes can add facts but cannot override the firm rules.Numbers stay out
The AI never writes SSNs, EINs or full account numbers in chats or files. When it needs to check one, it reports only “matches” or “does not match”.Everything is a draft
Every memo and workpaper is a draft until the responsible CPA approves it. Nothing goes to a client straight from the AI.Playbooks per workflow
Step-by-step playbooks and templates for tasks like return review, tax projections and bank reconciliations, so results are consistent across staff.Totals you can trust
Every total is labelled: taken from the source, re-added and verified, or unverified. The AI never calls a number verified unless it checked it.Originals untouched
Source PDFs are never edited or deleted. Tools that prepare files for the AI keep the originals and record what they changed.Who ran what
A short run log per client: who started work, when, and what was produced, with one staff member per engagement at a time.Tested updates
Rule changes are versioned, tested and reviewed before they reach staff, and client data is never stored with the setup files.What we offer
Start with one piece or take the full setup. Every engagement starts with the free consultation.
AI risk review
Where client data goes today, which tools staff use, and a data map with tiers for what may go where.Private cloud GPT
A private assistant on AWS Bedrock, Google Vertex AI or Microsoft Foundry in your own account, with sign-in, roles and logs.On-prem Mac mini GPT
Open-weight models on Mac minis in your office, set up, secured, backed up and maintained by us.Redaction engine, included
Free with every private GPT setup: file scrubbing, safe web search and prompt masking for client work, tuned for accounting, CPA, law and medical practices.Document search
Ask questions across approved folders, with answers that cite the source document.Firm playbooks
Written rules and step-by-step playbooks for your workflows, such as return review, reconciliations, contract review or visit-note summaries.Policy and WISP support
An AI use policy, an approved-tools list and an AI section for your written information security plan.Training and support
Staff workshops, then month-to-month support with a quarterly review of rules, logs and provider terms.Strategy: how we roll it out
Free consultation
30 minutes on your firm, your data and your rules.Data map and policy
We sort your data into tiers (public, internal, confidential, restricted) and decide which tier may go to which setup.Pilot one workflow
Pick the cloud, build one use case (for example engagement letters or document search) and test it on sample files.Roll out and review
Train the team, watch usage and redaction reports, and review rules and provider terms each quarter. Month-to-month support.Private GPT: common questions
What is a private GPT?
It is an AI assistant that runs in an environment you control, such as your own AWS, Google Cloud or Azure account or a machine in your office, instead of a public chatbot website. You decide who can use it and what data it can see.
Which cloud should we use: AWS Bedrock, Vertex AI or Microsoft Foundry?
Usually the one your firm already uses. All three can run strong models privately in your own account. We compare the models you need, your existing logins and your budget, and recommend one.
Isn’t ChatGPT Team or Claude Team enough?
Those plans don’t train on your business data by default and add admin controls, which is better than free accounts. But every prompt and file still goes to the AI company’s servers and back, with client details unmasked, and is kept under that company’s retention rules. A private GPT keeps the model in your own cloud account or office and masks identifiers first.
Does a private GPT help with SOX?
It can support SOX IT controls: role-based access, audit trails, tested and versioned changes, and human review of every output. SOX compliance depends on your whole control environment and your auditor; we support the controls, we do not certify them.
Is it really private?
It depends on the setup. In the cloud version, the model runs in your account and, under each provider’s terms, your prompts are not used to train their models. On-site stays on your hardware. We document where your data flows and how long anything is kept, so you and your advisor can review it.
What do you redact?
Identifiers such as client names, SSNs, EINs, bank, card and account numbers, dates of birth, addresses, phone numbers and emails, plus anything on your own lists, like client or matter numbers. For medical practices it also covers patient identifiers such as medical record numbers and health plan IDs. Our redaction engine replaces them with placeholders before the model sees a prompt, scrubs them from copies of client files, and masks them in web searches. It keeps details the work needs, such as tax years and form numbers.
Can staff search the web without exposing client details?
Yes. Our redaction engine masks names, identifying numbers and other client details in a search before it leaves your network, so only the general question goes out. Search results are treated as leads to check, not as evidence.
Do the logs contain client data?
They should not. We log who used the assistant, when and which documents, not the raw text. Provider request logs stay off unless needed, and any logs are encrypted with a set retention period.
Can medical practices use a private GPT under HIPAA?
Yes, with the right setup. In the cloud, your practice signs a BAA with AWS, Google or Microsoft and we use only their HIPAA-eligible services. On-prem, patient data stays on a Mac mini in your office. Our team works with test data, and we sign a BAA with your practice where our work involves patient information. HIPAA compliance still depends on your own policies, risk analysis and training.
Does this make my firm compliant?
No tool does that by itself. A private setup helps you meet your confidentiality duties, but compliance depends on your policies, training and how you use it. We work with your compliance advisor or attorney and do not give legal advice.
Cloud or on-prem?
Cloud is usually more capable and easier to maintain. On-prem open-weight models fit data that must not leave your building, though they are smaller. Many firms use both: everyday work in their private cloud, restricted files on-prem.
Can it run on a Mac mini in our office?
Yes. We set up Apple Silicon Mac minis to run open-weight models such as Llama, Qwen, Gemma or gpt-oss, with staff logins, disk encryption and no access from the internet. We size the machine to the models and the number of users, and test it on your tasks first.
How much does it cost?
It depends on the setup and how many people use it. Book the free consultation with the form at the top of this page and we will quote it.
Use AI without exposing client or patient data.
Book a free 30-minute private AI consultation, or call (305) 900-FISH.
