Skip to content
AI Compliance Tool
Check my AI risk
PlatformSolutionsPricingRegulationsFree toolsGuides Partners Sign in

Responsible Use of ChatGPT and Other AI Assistants at Work

Practical rules for staff using ChatGPT and similar assistants: data rules by category, enterprise vs consumer accounts, verification habits, disclosure and training.

By AI Compliance Tool editorial team · Published 2026-10-14 · Reviewed 2026-10-14 · Reviewer: Pending independent review · 9 min read

Key takeaways

  • Sort information into four categories (public, internal, personal, restricted) and give each a clear rule for whether it may go into an assistant and on which account tier.
  • Consumer and business account tiers usually differ in whether your inputs may be used to improve models, how long they are kept and what contract protects you. Read the terms for your tier; do not assume.
  • Treat every output as a draft. Verify facts, figures, names and citations against a source you can name before you rely on them.
  • Disclose AI involvement where a person is interacting with it or is affected by a decision; silence is the riskier default.
  • Short, role-based training with a completion record is the cheapest control you have and is what the EU AI Act Art. 4 literacy measures look like in practice.

Assistants such as ChatGPT and its competitors are now in nearly every small business, usually without anyone deciding. The risk is not that staff use them; it is that they use them without rules for what goes in, what comes out and who is told.

This guide gives you a working standard you can adopt as-is or fold into your acceptable-use policy. It describes vendor terms in general categories only; always read the actual terms for the tier you are on, because they vary and change.

Data rules by category

The simplest workable approach is four categories with a rule for each.

CategoryExamplesConsumer accountBusiness or enterprise account (with reviewed terms)
PublicPublished web copy, press releases, public regulations, your own marketingAllowedAllowed
Internal, non-personalProcess notes, draft procedures, product descriptions, anonymised dataAllowed only if the inventory entry permits itAllowed
Personal informationCustomer, applicant, tenant or employee names with any details; call transcripts; CVsNot allowedAllowed only where the inventory entry records a purpose, the terms exclude training on your inputs, and retention is known
RestrictedHealth, financial-account, biometric or children’s data; credentials and keys; unreleased financials; third-party confidential documents; anything a contract prohibits sharingNot allowedNot allowed unless a specific written approval exists for that tool and data type

Why personal information is the dividing line: under PIPEDA, an organisation may collect, use or disclose personal information only for purposes a reasonable person would consider appropriate (s. 5(3)), needs knowledge and consent (Principle 4.3), must safeguard it according to sensitivity (Principle 4.7) and remains accountable for information transferred to third parties (Principle 4.1.3). Alberta and BC PIPA contain parallel consent and reasonable-security duties. Quebec Law 25 adds a mandatory privacy impact assessment before acquiring an information system involving personal information (s. 3.3) and an assessment before communicating information outside Quebec (s. 17). In the EU, GDPR applies in full. Canada’s privacy regulators’ joint generative AI principles (guidance, December 2023) say the same in plain terms: have legal authority, limit collection, be open about AI use.

A useful test for staff: “If the person this is about were watching me paste it, would I be comfortable?” If not, it is not public or internal.

Enterprise vs consumer accounts

Account tiers differ in ways that matter. Rather than state any vendor’s current terms (they change), check these points for the tier you actually use and record the answers in the inventory entry:

  • Training on inputs. Does the vendor use your prompts and uploads to train or improve its models? Can this be turned off, and is it off by default on your tier?
  • Retention. How long are conversations and files kept? Can an administrator set a shorter period or delete on demand?
  • Contract. Is there a business agreement or data-processing addendum that names the vendor as your processor, lists subprocessors and commits to incident notification?
  • Administration. Can you see who has accounts, remove leavers, and export or delete a user’s data?
  • Location. Where are inputs stored and processed? Quebec s. 17 and GDPR transfer rules make this a real question.
  • Features that reach out. Does the assistant browse, run code, connect to your email or files, or take actions? Each of these is a permission to record.

A business tier with reviewed terms turns a consumer tool into something you can account for. It does not by itself make any use lawful; purpose, consent and safeguards still have to be yours.

Example: Orillia Fastening Supply, a 35-person distributor, found that six staff were on personal consumer accounts and were pasting customer purchase histories to draft quotes. The fix was a business-tier plan with training on inputs disabled, a 30-day retention setting, leaver removal added to the HR checklist and a one-line entry in the inventory recording the purpose (“quote drafting from customer order history”). Consumer accounts were restricted to public and internal content.

Verification habits

Assistants produce fluent text that is sometimes wrong. The habits below are the output half of responsible use.

  1. The output is a draft. Whoever sends, publishes or files it owns it.
  2. Check the five things that are most often wrong: numbers, dates, names, quotations and citations. Open the source. If the assistant cites a case, regulation or statistic you cannot find, treat it as fabricated.
  3. Do not ask the assistant to verify itself. “Are you sure?” produces confidence, not accuracy.
  4. Legal, financial, medical and safety statements get a human expert, not a second prompt.
  5. Know what the assistant cannot see. It does not know your contract, your customer’s history or this morning’s price change unless you gave it that, and giving it that may be a data-rule problem.
  6. Watch for automation bias. After a few weeks of good drafts, people stop reading. Build in a spot-check: a manager reviews a sample of AI-assisted customer replies monthly and records the result.

For public claims, the verification bar is higher. The FTC enforces Section 5 against unsubstantiated AI and performance claims (Operation AI Comply, September 2024; Workado order, August 2025). If an assistant drafts “our tool is 99% accurate” into your web copy, you need the evidence, not the draft.

Disclosure

Disclosure has two different triggers. Keep them separate.

People interacting with AI. If a customer or member of the public is talking to an AI system rather than a person (a chatbot, a voice agent), say so at the start. EU AI Act Art. 50(1) requires this from 2 August 2026 unless it is obvious from context. Utah Code 13-75-103 requires disclosure when a consumer clearly asks, with a safe harbour for systems that disclose at the outset and throughout. Canada’s regulators’ generative AI principles recommend disclosure as a matter of openness. The sensible default everywhere is to disclose.

People affected by AI-assisted decisions. If an assistant helps screen applicants, score customers or decide eligibility, the rules about automated decisions apply: Quebec s. 12.1 notice for decisions based exclusively on automated processing; GDPR Arts. 13(2)(f) and 22; Ontario ESA s. 8.4 statements in publicly advertised job postings for employers with 25 or more employees; Illinois HB 3773 employee notice; Colorado SB 26-189 notices from 1 January 2027.

Internal and editorial use. If a person drafts with an assistant and reviews and sends under their own name, the rules above are generally not triggered. Your internal standard should still say what is acceptable: for example, AI-assisted drafting is fine; presenting AI-generated analysis as independent expert work is not; synthetic images of real people require written consent.

Training

Training is what turns the rules above into behaviour, and the record is what proves it.

  • Who. Everyone who uses an assistant for work, including contractors and agencies that touch your data.
  • What. Forty-five minutes: the four data categories with real examples from your business; which account tier to use for what; the verification habits; the two disclosure triggers; how to report a mistake without blame.
  • When. On joining and annually; and a short refresher when a tool gains new capabilities (browsing, file access, actions) or you enter a new jurisdiction.
  • Record. Date, attendees, materials version. In the EU this supports the Art. 4 AI literacy measures expected of deployers; the Omnibus text (Regulation 2026/1744) states this does not require guaranteeing any specific level of literacy for any individual, but you should be able to show what measures you took.

A one-page quick reference pinned in your chat tool, listing the four categories and the five things to verify, does more than a long policy nobody opens.

A short standard you can adopt

We use AI assistants for drafting and analysis. Public and internal information may be used on approved accounts. Personal information may only be entered on the business account listed in the AI inventory for a recorded purpose. Restricted information is never entered. Every output is a draft; the person who sends it checks the facts. We tell people when they are talking to an AI system and when AI helps make a decision about them. Everyone completes the AI briefing yearly, and mistakes are reported the same day without blame.

That is seven sentences. Most of the value is in applying them consistently and keeping the records that show you did.

Sources

Frequently asked

Can staff use a personal ChatGPT account for work?
Only if your policy and inventory permit it for public and internal information. Personal and restricted data should not go into an account the business does not control, because the business cannot see the terms, manage retention or respond to a request from the person whose data it is.
Does pasting a customer email into an assistant break privacy law?
It can. A customer email usually contains personal information, and sending it to a third-party service is a use and disclosure that privacy law expects you to have a basis for and safeguards around. Business-tier tools with contractual protections and a documented purpose are the usual answer; consumer tools are not.
Do we have to tell customers we used AI to write a reply?
There is no single rule. If the customer is interacting with an AI system rather than a person, disclose it. If a person reviewed and sent a drafted reply, most rules do not require a label, but an honest internal standard and a disclosure where the content could mislead is good practice.
Do we have to train staff?
In the EU, the AI Act Art. 4 requires providers and deployers to take measures to support the AI literacy of staff using AI systems. Elsewhere there is no general training mandate, but training records are the most practical evidence that your data rules are real.

This guide is general information, not legal advice. Laws and dates are drawn from our verified regulatory dataset and link to official sources; check them for changes after the review date.