Client confidentiality and generative AI can work together when three things hold: your client contracts allow the use, the tool runs on a business plan whose data terms you have actually read, and your team knows what never goes into any tool. Most agencies cover the third by instinct and the first two by assumption.
The assumptions are where it breaks. Not in a dramatic leak, but in client material quietly leaving the agency under terms nobody checked. This guide covers the three controls in order. It describes working practice, not legal advice: how your contracts apply is a question for your counsel.
Where confidentiality actually breaks
An illustrative example, not a real agency. A 35-person healthcare marketing agency serves a regional hospital system, a maker of orthopedic devices and a group of urgent-care clinics. Its operations director asks each team one question: which AI tools are you using on client work?
The answers come back in three groups. The account teams use the AI assistant built into the agency's own office suite. Two strategists use personal chatbot accounts they pay for themselves, because those were faster to set up. A designer uses a free browser tool to rewrite headlines. Nobody has done anything they thought was wrong.
Meanwhile a copywriter is holding the transcript of a patient interview for the hospital's patient-story campaign, and wants a first draft by Friday. Where that transcript goes next is the whole question. The answer depends on the contract, the account, and the content, in that order.
Step 1: Read the client contracts before you pick a tool
Your obligations come from what you signed, not from what the AI vendor promises. Pull the master services agreement, any separate NDA and the current SOW for each client, and look for five things:
- What counts as confidential. Many definitions cover anything the client discloses, marked or not. Assume briefs, research, interview material and unreleased plans are inside it unless the contract says otherwise.
- What you may use it for. Wording that limits confidential information to performing the services.
- Who you may share it with. Limits on disclosure to employees and contractors, and any requirement for client consent before using subcontractors or third-party services.
- Health and personal information terms. If you signed a business associate agreement or a data processing addendum with a client, read it as closely as the NDA.
- Anything that mentions AI. Search for "artificial intelligence", "machine learning" and "generative". Some clients now add their own rules.
Then give each client a one-line answer: permitted, permitted with conditions, not permitted, or unclear. Where it is unclear, you have two good options. Ask your counsel how the clause applies, or ask the client directly and get the answer in writing. A short note to your client contact saying which tool you would like to use, on which plan, for which kind of work, often settles it faster than a legal reading. For the hospital in the example, that note matters more than any vendor's terms.
Step 2: Put client work on a business plan, and know its terms
Consumer and business versions of the same tool can carry different data terms. Here is what three vendors say about their business plans on their own pages. Checked on November 20, 2026. Terms change, so check again before you rely on them, and remember that your actual terms are the contract and any data processing addendum you sign.
- OpenAI. Its Enterprise privacy at OpenAI page (updated January 8, 2026) says that by default it does not use business data from ChatGPT Business, ChatGPT Enterprise and its API Platform for training. The same page says it uses data from versions of ChatGPT for individuals, and that it can sign a business associate agreement for API Platform customers.
- Microsoft. Its Microsoft Learn page, Data, Privacy, and Security for Microsoft Copilot (dated July 9, 2026), says prompts, responses and data accessed through Microsoft Graph are not used to train foundation models, and that Copilot only surfaces organizational data the individual user has at least view permission for.
- Google. Its Generative AI in Google Workspace Privacy Hub (last updated October 6, 2026) says that, for qualifying Workspace editions, your content is not human reviewed or used for generative AI model training outside your domain without permission, and that Gemini only retrieves content the user has access to.
Two practical conclusions for an agency. First, personal and free accounts are out for client work, whatever the tool. The two strategists in the example were not careless; they were on the wrong plan. Move everyone to accounts the agency owns and administers.
Second, an assistant built into your office suite sees what each user can see. In an agency, that is a client-separation problem before it is an AI problem. If the device maker's launch folder is shared with the whole company, the assistant can surface it to anyone working on the clinic account. Tidy the shared drives first: one folder per client, opened only to the people on that account.
Step 3: Decide what never goes into any tool
Business terms reduce the risk. They do not make every input acceptable. For a healthcare agency, the list that holds on every plan looks like this:
- Patient names, dates of birth, conditions, or any detail that could identify a patient, including in interview transcripts and photo captions.
- Material from any client whose one-line answer is "not permitted" or still "unclear".
- Unreleased product, clinical or regulatory information, such as a device maker's launch plan before it is public.
- Customer and patient lists, CRM exports and survey responses with names attached.
- Another client's material in the same conversation. One client per conversation or project.
- Passwords, access tokens and login links, for anyone's systems.
Write the list on one page and name the person who answers questions about it. A rule people have to guess at gets guessed at under deadline.
Back to the transcript
With the three controls in place, the copywriter's Friday draft is straightforward. The hospital's one-line answer says AI drafting is permitted with conditions, confirmed in writing by the client contact. The transcript is de-identified first: names, dates and anything specific to the patient's care replaced with placeholders. The draft is made in the agency's own business account, in a project used only for the hospital. The real details go back in afterwards, in the agency's own document, and the account lead checks every factual claim before the client sees it.
Nothing here slowed the work much. What changed is that every step now has an answer someone wrote down. Proposals need the same written answers, and using AI to write proposals sets out who reviews what.
Where the AI Opportunity Review fits
Confidentiality is part of deciding which workflow is worth testing at all. The AI Opportunity Review, a $1,500 engagement over 10 business days, reviews data suitability, access and tool requirements before recommending a workflow. Data suitability is one of four things each candidate workflow is scored on, next to benefit, effort and operational risk. Deferring AI may be the recommendation.
If a pilot follows, it uses existing commercial tools where suitable, with connections, permissions and software costs confirmed before you commit. The standard pilot is deliberately narrow; sensitive workflows need separate scoping. What you get from a $1,500 AI Opportunity Review walks through the deliverables.
The same rule applies to scoping itself. Describe your workflow in general terms. Do not submit client names, credentials or confidential documents in the form.
GradeCircle builds and runs six products of its own: AfroSynergy, CallerBridge, HealthSquire, Heyzl, Lexfora, and LukeAPI, two of them in beta. The assumptions in our models come from operating those systems, not from a framework. See what we run.
If your agency writes proposals every month, scope your proposal workflow: five questions, and you see a preliminary fit summary before sharing contact details.



