Skip to content
Small Business Elevate

AI Security Checklist for Small Business: 10 Steps

A 10-step AI security checklist for companies with no IT team, including where to switch off ChatGPT's "Improve the model for everyone" data control.

7 min readReviewed Check this is still true

A clipboard checklist with teal checkmarks and a coral warning icon, overlaid on a security shield, surrounded by a padlock, toggle switch, privacy eye, and key icons

Do three things today, in about an hour: list the AI tools your business uses, decide what must never go into them, and switch off training on your chats.

We come from a cybersecurity background, and this is the order we would work in. Working alone or with one partner? Skip step 6 and keep the policy to half a page.

Do today#

1. Find out which AI tools your team uses#

About 30 minutes.

  • Ask everyone which AI tools they use for work, what for, and what they put in. Say nobody is in trouble, so you get honest answers.
  • Check card statements for AI subscriptions. Ask about browser extensions, AI browsers, AI note-takers and anything connected to work email, calendars or files.
  • Sort the answers into approved, allowed with limits, and not allowed. Approve what you can: a tool blocked on the laptop moves to people's phones, where you see nothing.
  • Give each person their own login, with two-step sign-in by passkey or authenticator app where the tool offers one. Where a tool signs you in through email or Google, as Claude does, protect that account.

In a 2025 survey of 1,000 US workers who use AI on the job, 78% said they use AI tools their employer does not provide.

2. Decide what data may go in#

About 15 minutes. Print this list.

The more it says about a person, the fewer places it can go
  1. 1PublicMarketing copy, published prices, blog drafts

    Any AI tool

  2. 2InternalMeeting notes with no client names, process documents

    Approved tools onlywith training switched off (step 3)

  3. 3ConfidentialClient details, financial records, contracts

    Approved tools only, names and numbers taken out firstTake out names, account numbers, tax IDs and addresses: "Client A owes $4,210"

  4. 4RestrictedID numbers, passwords, health records, card details

    Never goes into any AI tool

The word that does the work is together

  • Maria, on its ownLevel 2
  • Maria with her balance, 60 days overdueLevel 3
  • Maria with her diagnosisLevel 4: never

Bar length shows how far the information may travel. Levels from our checklist; your own rules can be stricter.

Connect work email and files only to approved tools. If an AI agent or AI browser acts for you, keep it signed out of banking and client systems, and read every confirmation before you approve it: OpenAI says the tricks that hijack agents are "unlikely to ever be fully solved".

3. Turn off AI training on your chats#

A few minutes per tool, on every account used for work. It only applies from now on, so do it today.

Where each switch lives

  1. ChatGPTFree, Go, Plus, Pro

    Settings, then Data controls

    Improve the model for everyoneSet to: Off

    OpenAI may train on your chats until you do.

  2. ClaudeFree, Pro, Max

    Settings, then Privacy

    Help Improve our AI modelsSet to: Off

    Claude trains on chats only if you allowed it, at sign-up or in a 2025 pop-up, so check yours.

  3. GeminiPersonal Google account

    gemini.google.com, then Settings & help, then Activity, then click "On" near the top

    Keep ActivitySet to: Turn off

    Keep Activity is on by default for adults.

    Turning it off also disconnects Gmail, Drive and most other connected apps.

  4. CopilotPersonal Microsoft account

    No training switch in the new app, including copilot.com

    Microsoft says your chats and files are not used to train its foundation models.

    Settings, then Personalization

    Allow ads personalizationSet to: Off

    On the same screen, check what "Saved memories" holds.

    On an older app that has not updated: profile icon, then Settings or Account, then Privacy. Switch off "Training on conversation activity" and "Training on voice conversations".

From each vendor's help pages, checked 12 October 2026. Vendors change these screens.

Do not rate answers on client work. A thumbs up or down can send the whole chat for training in ChatGPT and Claude even with the switch off, and Gemini feedback can include your last 24 hours of chats. Memory is a separate setting too: check what each tool has saved.

Business plans (ChatGPT Business, Claude Team, Copilot on a work account) do not train on your chats by default; Gemini on a work account follows your admin. ChatGPT Plus and Claude Pro are personal plans. No plan makes you compliant on its own.

Do this month#

About half a day in all.

4. Confirm payment requests on a number you already have#

About 15 minutes to agree.

  • No payment, bank-detail change or password reset is approved by email, chat, voice or video alone. Call back on a number you already had.
  • Above an amount you choose, two people approve.
  • Agree a code phrase for payments, and ask for it whenever a request is unexpected.

The FBI warns that criminals use AI-cloned voices to impersonate people and get payments, and in 2024 a Hong Kong finance worker sent about US$25.6 million after a video call full of deepfakes.

5. Write a one-page AI policy#

About two hours. Open the AI policy template. Anyone can view it; to fill it in, use File, then Make a copy (needs a Google account) or Download. Work from your step 1 list. It needs four things:

  • Approved tools, by type and plan level. A personal account and a business account of the same product are different things.
  • Data rules: the list from step 2.
  • One rule above the rest: a person reads anything AI wrote before it leaves the business. In Mata v. Avianca, a court fined two lawyers and their firm $5,000 in 2023 over a brief citing cases ChatGPT had made up.
  • Who to call when something goes wrong: one named person and a phone number. Nobody is blamed for reporting.

Everyone reads it and signs it.

6. Train the people who use the tools#

About an hour, then a refresher each year.

  • Walk through the step 2 list with five real examples from your business.
  • Give everyone one sentence to remember: if you would not email it to a stranger, do not paste it into an AI tool.
  • When AI does not know something, it makes up an answer that sounds right. Check names, numbers and prices before anything goes out.
  • Tell one true story. In July 2025, shared ChatGPT chats showed up in search results because people had ticked a box they did not understand.

7. Know what to do when something has already gone in#

Assume it happens once: a contract, a patient list or a password pasted into a personal account.

  1. Tell the person named in your policy the same day, and write down what was shared, in which tool, and when.
  2. Once it is written down, turn off training on that account and delete the chat.
  3. Ask the vendor in writing to delete the data, using the route in its privacy policy.
  4. Ask your lawyer whether anyone must be told. Health, financial and legal data carry notification duties in many states.

8. Review it quarterly#

About an hour a quarter. Between our September and October 2026 checks, three of the four vendors in step 3 changed their help pages or settings.

  • Do the settings in step 3 still match the screen?
  • Is anyone using a new tool? Has anyone left who had an AI account? Remove their access.
  • Are there new rules in your state? Start with the NCSL AI legislation database, updated monthly.

Only if it applies#

Open the ones that describe your business.

9. Ask five questions before a new tool touches customer data

Get the answers in writing.

  1. Do you train models on our data? Search the terms for "training".
  2. Where is our data stored, and in which country? That decides which laws apply.
  3. Has an outside auditor checked your security? Ask for the SOC 2 Type II report.
  4. Will you sign a contract on how you handle our data? A data processing agreement, or a business associate agreement for health data.
  5. What happens to our data when we leave, and how fast is it deleted?

If the tool connects to your email or files, or can act on its own, also ask what it can reach and what it can do without a person approving it. In August 2025, attackers used stolen access from an AI chatbot linked to companies' sales systems to export customer records.

10. Look after your API keys if you have built something on an AI service

An API key is the password software uses to reach an AI service. If you only use AI in a browser or an app, skip this. About 20 minutes.

  • Keep keys out of the code: use environment variables or a password manager.
  • Set a spending limit on every AI platform account. A leaked key runs up your bill.
  • Ask whoever built your AI tool where the keys are stored and who can see them.
  • Replace any key pasted into a chat, a ticket or an email.

In 2025, an xAI key sat in a public GitHub repository for about two months after a researcher flagged it.

If you handle health records, tax returns or client accounts

A business plan is not compliance.

  • Health records need a business associate agreement that covers the exact product you use. OpenAI offers one for ChatGPT Enterprise set up as a regulated workspace, ChatGPT for Healthcare, ChatGPT for Clinicians and its API. It offers none for ChatGPT Business. Anthropic offers one on Enterprise only, not Team, and the account owner has to switch it on. Google covers the Gemini app under a Workspace agreement an administrator accepts. Microsoft covers Copilot in its commercial agreement and says plainly that holding it does not make you compliant.
  • Tax and financial records. If you prepare tax returns, check the IRS rules on client tax information (section 7216) first. Under the FTC Safeguards Rule, firms that handle consumers' financial information may already need a written security plan. Put your AI rules in it.

Checked October 2026.

Next step#

Start with steps 1 to 3 today.

If you would rather have someone go through your accounts, settings and vendor terms with you, and write the policy around how your team works, tell us what you use. The 15-minute AI audit checks this in three questions, alongside four other parts of your business.

Check this is still true

Reviewed 12 Oct 2026

Tools, prices and rules change. Paste the prompt below into any AI assistant you use and ask it to check the parts that date.

See the prompt

I'm reading "AI Security Checklist for Small Business: 10 Steps" by Small Business Elevate, last reviewed 12 October 2026. Please check whether these points are still accurate today. Use the vendors' own current help pages and official sources, and give me links with dates. 1. The settings for model training in each of these four products. The article says: ChatGPT: Settings, then Data controls, then "Improve the model for everyone", set to off. Claude: Settings, then Privacy, then "Help Improve our AI models", set to off. Gemini: gemini.google.com, then Settings & help, then Activity, then click "On" and choose Turn off ("Keep Activity"). Microsoft Copilot: in the new app (from August 2026, including copilot.com) chats and files are not used to train Microsoft's foundation models, so it has no training switch; Settings, then Personalization holds "Allow ads personalization" and "Saved memories". The older app has "Training on conversation activity" and "Training on voice conversations" under Privacy. Is each of those still true, and still worded that way? 2. Do ChatGPT and Gemini still use consumer conversations for training unless the user turns it off? Does Claude still train only when the user allowed it? Does Microsoft still say chats and files in the new Copilot app are not used to train its foundation models? 3. Are ChatGPT Business, Claude Team and Copilot with a work account still excluded from model training by default? Does turning off Gemini's Keep Activity still disconnect Gmail, Drive and most other connected apps? 4. Can a thumbs up or down rating still send the whole conversation for training in ChatGPT and Claude, even with the training setting off? Can Gemini feedback still include the last 24 hours of chats? 5. Does Claude still sign people in through Google or an emailed link, with no two-step setting of its own? Does ChatGPT still offer passkeys? 6. Which of those vendors will sign a HIPAA business associate agreement today, and on which specific plans? Does OpenAI still decline one for ChatGPT Business, and does Anthropic still limit it to Enterprise? 7. Is turning off training still forward-looking only, with no way to withdraw data already used? 8. Are there new US state rules in force since 12 October 2026 that change what a small business must do before putting customer data into an AI tool, or before using an AI chatbot or agent with customers? If you cannot confirm something, say "could not verify" rather than guessing. Keep the answer short.

Share this guideLinkedInX

Read next

All guides

Want a second opinion before you roll something out?

We can review how your team uses AI today, what information goes into it, and who has access, then suggest changes you can make this month.

Tell us what you needSee our AI consulting services for small businesses