AI tools a 5-person business can use safely

A five-person business can safely use AI for drafting emails and quotes, summarizing calls, answering common questions from your own information, and tidying up bookkeeping tasks. The safety part is simple: pay for a business plan that doesn't train on your data, keep sensitive data out, and have a person approve anything before it reaches a customer. Here's what that looks like in practice.

Where AI actually saves time

Drafting emails, quotes and proposals

This is the easiest win. Give a general-purpose assistant (ChatGPT, Claude, Gemini, Microsoft Copilot) your notes and a past example, and ask for a draft in the same tone. It's good at:

• Turning bullet points into a polite reply.
• Rewriting a stiff email so it sounds human.
• Drafting a quote cover letter from your line items.
• Translating a message and keeping the meaning (have a fluent speaker check anything important).

The draft is a starting point. You check prices, dates, names and promises. AI will confidently fill gaps with things that sound right and aren't.

Summarizing calls and meetings

Many video-call platforms and stand-alone AI note-takers can record a call and produce a summary and action list. Check which features your plan actually includes before you count on one. Two rules:

• Tell people you're recording. Recording laws vary by state and country, and some require everyone's consent. When in doubt, ask at the start of the call.
• Read the summary before you send it. Summaries drop nuance, mix up who said what, and sometimes invent action items.

Answering FAQs from your own information

A chatbot or help-desk assistant that answers from your own documents (price list, policies, service area, hours) can handle the repeat questions. The key words are your own information. Give it a clean, current FAQ document, not your whole hard drive.

Start internally: let your team ask it questions first. Once its answers hold up for a few weeks, consider putting it in front of customers, with a clear way to reach a person.

Bookkeeping help

Accounting platforms are adding AI features to categorize transactions, match receipts and flag duplicates. A general assistant can also explain a report or draft a polite late-payment reminder. Treat all of it as a suggestion. A person reviews every entry, and your bookkeeper or accountant still signs off. Don't paste full bank statements into a general chatbot.

The data rules

Use a business plan, not personal accounts

Free and personal AI accounts often have different data settings from business plans. What the major providers say about their business offerings:

• OpenAI: "By default, we do not train on any inputs or outputs from our products for business users," which covers its ChatGPT business plans and API. Consumer ChatGPT may be used for training unless you opt out.
• Anthropic: its commercial products, including Claude's Team and Enterprise plans and the API, aren't used for training by default. One exception: if a user submits feedback (like a thumbs up or down), that conversation can become eligible. Team and Enterprise owners can turn off feedback with the Rate chats setting.
• Google Workspace: Google says Workspace "does not use customer data for training models without customer's prior permission or instruction," and that covers Gemini in Workspace.
• Microsoft 365 Copilot: Microsoft says prompts, responses and data accessed through Microsoft Graph "aren't used to train foundation LLMs." Copilot only shows data a user already has permission to see, so messy file sharing becomes visible fast. Clean up who can see what before you turn it on.

Read the current terms before you buy. These pages change, and a setting someone flips can change what happens to your data.

What never goes into an AI tool

Even on a business plan, keep these out unless a specific tool has been approved for them:

• Passwords, API keys, login codes.
• Payment card numbers, bank account and routing numbers.
• Social Security numbers, driver's license or passport numbers.
• Health information about customers or staff.
• Anything a client contract or NDA marks confidential.
• Employee reviews, disciplinary notes or salary details.

The FTC's guide for business on protecting personal information boils down to five steps: take stock of what you have, scale down to what you need, lock it, pitch what you don't need, and plan ahead for problems. Pasting data into a new tool is a new place you're storing it.

A person approves anything that goes out

No AI-written email, quote, review reply, social post or chatbot answer goes to a customer without a person reading it first. That's the rule that prevents most damage. Turn off any "auto-send" or "auto-reply" setting until you trust the tool, and keep it off for anything involving money, legal terms or complaints.

Check accuracy every time

AI tools can produce wrong answers that sound completely sure of themselves. NIST's generative AI risk profile calls this "confabulation." Before anything goes out:

• Check numbers against the source: prices, totals, dates, quantities.
• Check names, addresses and phone numbers.
• Check any claim about a law, rule or policy against the official source.
• If the AI says something you didn't tell it, assume it made it up until proven otherwise.

Your one-page AI policy

Write it, share it, and have each person read it. Keep it to one page:

1. Approved tools: list them by name, on the business account. Personal accounts aren't used for work.
2. Never paste: the list above.
3. Approval: a person reviews and approves everything customer-facing before it's sent or posted.
4. Accuracy: check every number, name, date and factual claim.
5. Recording: tell people when a call is recorded or transcribed.
6. Accounts: the business owns every AI account. An admin can add and remove users. When someone leaves, their access goes the same day.
7. Questions: who to ask when you're not sure whether something is OK.
8. Review date: look at the policy again every six months.

Training matters as much as the policy. A short session where everyone tries the approved tool on real (non-sensitive) work does more than a document alone. Our guide to training staff on new software covers how to run one. Once drafting is working, the next step is often automating quotes and bookings.

Closing

Pick one task, one approved tool and one rule: a person approves what goes out. Get that working before adding more. If you'd rather have it done, write to [email protected] or see our AI tools service, where tools are set up to work from your own information, in your name, with your team trained.

Sources

• How your data is used to improve model performance (OpenAI)
• Is my data used for model training? (Anthropic Privacy Center)
• Updates to Consumer Terms and Privacy Policy (Anthropic)
• Generative AI in Google Workspace Privacy Hub (Google Workspace Help)
• Data, Privacy, and Security for Microsoft 365 Copilot (Microsoft Learn)
• Protecting Personal Information: A Guide for Business (FTC)
• Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1 (NIST)