Using AI on client work without handing over the client

The problem

You're already using AI. Your confidentiality duty didn't pause.

Lawyers, accountants, and consultants are pasting real matters into AI assistants every day — contracts, financials, client emails. The work is better and faster. But the obligation you carry to the people you serve doesn't switch off when you open a new tab.

Here's the uncomfortable part: anything you paste into a public AI tool is handled under that tool's terms, not yours. Most consumer AI terms permit the provider to retain inputs, use them to train models, and disclose them to third parties or regulators in some circumstances. You can't undo a paste. And "I'll just be careful" is not a method.

What most people get wrong

They protect the tool. The exposure is the input.

Two common mistakes:

• "I turned off training / I'm on the paid tier, so I'm fine." Account settings reduce one specific risk. They don't change what the data is once it's left your machine, and they don't satisfy a professional confidentiality duty on their own.
• "I'll anonymize it later / the AI won't remember." By the time the output comes back, the identifying details are already in. The only safe moment to remove client identity is before you paste.

The fix isn't a better tool or a stronger setting. It's a habit applied at the input — every time, before the paste.

The [Client] bracket method, concretely

Strip identity at the door. Keep the substance.

The method is simple and you can start today, with no software:

Before you paste, replace every identifying detail with a bracketed placeholder. The AI still sees the shape of the problem — it just never sees whose problem it is.

You do the work with the brackets in place. When the AI returns a draft, you swap the real names back in on your own machine, in your own document. The model does the reasoning; it never holds the identity.

This is a discipline you apply by hand. It is not a technical control — Sharpwork can't stop a leak inside a tool we don't run. What it does is keep the most sensitive thing — who the client is — out of the input entirely.

Sharpwork's prompts come pre-bracketed. Every prompt in every kit ships with the placeholders already in place, so the disciplined version is the default version — you fill in your details locally instead of pasting them into the model.

What the law and professional rules already say

You don't have to take our word for it. The duty is already on the books.

A reality check, not legal advice. This is a plain-language summary of where the rules already point. It is not legal or professional advice, and none of these sources mention Sharpwork. Apply your own professional judgment.

Courts have started treating public AI tools as a confidentiality risk.
A 2026 U.S. federal court ruling found that someone who used a public AI tool on their own — without their attorney’s direction — could not claim attorney-client privilege over what they produced, in part because the tool’s terms allowed the provider to retain and share what was entered. The takeaway is plain: pasting a client matter into a public AI tool can forfeit protections you’d assume were automatic.

The legal profession’s ethics guidance points the same way.
The bar’s formal guidance on lawyers’ use of generative AI tells lawyers to protect information relating to a representation when using AI, and to get the client’s informed consent — a real explanation of the risk, not boilerplate — before entering confidential information into a self-learning AI tool.

Accountants are already bound, too.
The accounting profession’s existing duty of client confidentiality doesn’t pause for new tools. Putting client data into a third-party AI tool counts as a disclosure, so the confidentiality rule you already follow applies — this isn’t a new “AI rule,” it’s the existing one.

The common thread: every one of these points the same way — don't put identifiable client information into a tool you don't control. The bracket method is the practical way to honor that and still use AI.

What Sharpwork gives you

The brackets, built in — plus the work, not just the prompts.

• Pre-bracketed prompts for the real tasks in your field, so the confidential-by-default version is the one you already have open.
• Step-by-step workflows that turn AI into finished work — a reviewed draft, a structured analysis — not a generic one-liner you still have to fix.
• The confidentiality method, taught once and baked into every prompt so it becomes a habit, not a thing you remember to do.
• Works with the AI you already pay for — ChatGPT, Claude, Gemini, or Copilot. Nothing to integrate, nothing that sits between you and your tool.
• One-time purchase. No subscription, no login to run AI, no account for us to hold your data in.

Honest limits: what this is NOT

We'd rather you trust the method because we told you the truth about it.

• This is not a technical guarantee. Sharpwork does not run a model, does not sit between you and your AI, and cannot prevent a leak inside a tool we don't control.
• The bracket method is a discipline you apply by hand. It only works if you actually do it, every time, before you paste.
• It does not make AI output "confidential," and it is not legal, accounting, or professional advice. You remain responsible for reviewing and verifying anything the AI produces.
• It reduces the most consequential exposure — client identity in the input — but no method removes every risk of using a third-party tool.

Start with the kit built for the highest-stakes work

The Legal kit is where confidentiality matters most — and where the bracket method earns its keep on day one. Every prompt ships pre-bracketed, mapped to real legal tasks.

See the Legal kit → Browse all five kits →

One leaked client matter costs more than the kit. $197 once, 14-day money-back guarantee.