The complete guide to using AI safely in a small business
A complete, plain-English guide to using AI safely in a small business: the main risks from standards bodies and security agencies, a one-page policy, data rules, output checks, vendor questions, prompt injection defenses, human approval for AI agents, training and incident planning.

Artificial intelligence tools have moved from curiosity to everyday equipment in small businesses. Staff use chatbots to draft emails, summarize meetings, write job ads, analyze spreadsheets and answer customer questions. Done well, this saves real time. Done carelessly, it can leak customer data, produce confident but wrong answers, create legal exposure and open new doors for attackers. The goal is not to avoid AI, but to use it with the same care you would give any powerful tool that touches your business.
This guide is a complete, plain-English walkthrough of using AI safely in a small business. It explains the main risks as described by standards bodies and security agencies, then turns them into practical steps: a simple policy, rules about what data may go into a tool, how to check outputs, how to choose and review vendors, how to protect against misuse by attackers, and how to keep humans in charge when AI acts on your behalf. It is general guidance, not legal, privacy or security advice for your specific situation, so involve a qualified professional for regulated data or high-stakes decisions.
In this guide
- Why AI needs its own safety habits
- The main risks, in plain English
- Start with a short AI use policy
- Decide what data may go into AI tools
- Check outputs before you rely on them
- Choose and review AI vendors
- Defend against prompt injection and misuse
- Keep humans in charge of AI that acts
- Be careful with people decisions and customer-facing AI
- Train staff and prepare for incidents
- Common mistakes
- Where WorkBOS fits
- Frequently asked questions
Why AI needs its own safety habits
Most AI tools are easy to start using and hard to see once they spread. One employee pastes a client contract into a free chatbot to summarize it. Another connects an AI note-taker to calendars and email. A third uses a browser extension that reads every page. None of this involves a purchase order or an IT review, which is exactly why it needs a lightweight set of rules.
Authorities have been clear that existing good practice still matters. The UK National Cyber Security Centre's AI and cyber security guidance tells leaders to build security in from the start, to treat security as a business priority rather than a purely technical matter, and to prepare a response plan for AI failures and serious incidents. The U.S. Cybersecurity and Infrastructure Security Agency says in its AI materials that fundamental security practices still apply to AI systems. So the foundation is familiar: know what data you hold, limit who can see it, keep software updated, and have a plan for when something goes wrong.
Risk-based thinking, not fear
Not every AI use carries the same risk. Using a tool to brainstorm blog titles is very different from using one to screen job applicants or draft legal terms. The U.S. National Institute of Standards and Technology (NIST) offers a voluntary AI Risk Management Framework, released as version 1.0 in January 2023 and organized around four functions it calls Govern, Map, Measure and Manage. NIST describes it as voluntary guidance to help build trustworthiness into how AI products and systems are designed, developed, used and evaluated. You do not need to adopt a framework formally, but its logic scales down well: decide who is responsible, understand where AI is used and what could go wrong, check how it performs, and act on what you find.
NIST has also published a Generative AI Profile (NIST AI 600-1, released July 26, 2024) as a companion to that framework, aimed at the specific risks of generative AI tools such as chatbots and image generators. It is detailed and written for professionals, but it is a useful reference if you want to go deeper.
The main risks, in plain English
Here are the risks small businesses most often meet, grouped so you can match each with a control later in this guide.
1. Data exposure
Information you type into an AI tool may be stored, reviewed by the provider, used to improve its models, or exposed through a bug or a breach, depending on the product and your plan. The OWASP project, which publishes widely used security guidance, lists sensitive information disclosure as one of the top ten risks for applications built on large language models, noting that sensitive data can be exposed through the model and through the application around it. You can read the list on the OWASP Top 10 for LLM Applications page.
2. Wrong, biased or made-up answers
Generative AI can state false information confidently. The NCSC describes this as "AI hallucination" and also notes that models can be biased, often go along with leading questions and may absorb incorrect or controversial material from training data. OWASP lists misinformation as another top risk. The practical meaning is simple: never treat AI output as fact until a competent person has checked it.
3. Prompt injection and manipulation
Prompt injection means crafting input that makes a model behave in ways its owner did not intend. The NCSC explains that this can lead to offensive output, disclosure of confidential information or unintended actions in systems that accept model output without checks, and that the risk grows as models pass data to other applications and services. OWASP ranks prompt injection first in its list. If your AI tool reads emails, web pages or documents from outside, someone can hide instructions in them.
4. Too much autonomy
OWASP calls out "excessive agency", meaning systems that are given more autonomy or capability than they need. An AI that can send emails, move money, delete files or change records without review is a larger risk than one that only suggests.
5. Supply chain and vendor weaknesses
The tools you use depend on models, plug-ins, data sources and hosting providers you do not control. OWASP lists supply chain weaknesses among the top ten. The NCSC urges organizations to understand their data, model and software supply chains and to question suppliers about security.
6. Privacy and legal exposure
Personal data brings obligations. In the U.S., the FTC's data security guidance for businesses tells companies to collect only the sensitive personal data they need, to protect it and to dispose of it securely, and the FTC has said it expects companies to keep marketing claims truthful, including claims about AI. In the UK and EU, data protection law applies to personal data used in AI systems, and the UK Information Commissioner's Office publishes guidance on AI and data protection, including a risk toolkit and advice on explaining AI-assisted decisions to the people affected. If you serve customers in other regions, check the rules that apply.
7. Cost and availability
OWASP also lists unbounded consumption: AI processing that can use resources without limit. For a small business that usually shows up as a surprise bill or a tool that stops working. Set spending caps and monitor usage.
Start with a short AI use policy
You do not need a thirty-page document. A one- or two-page policy that staff can remember is far more effective. Keep it plain and put it in your employee handbook or shared drive. Our employee handbook basics guide shows where it fits.
What to include
- Purpose and scope: who the policy covers (staff, contractors) and which tools.
- Approved tools: a short list of tools staff may use for work, and how to request a new one.
- Data rules: what may and may not be entered into each type of tool (see the next section).
- Output rules: AI output must be reviewed by a person before use, and certain content (legal, medical, financial, hiring) needs specialist review.
- Disclosure rules: when to tell customers or colleagues that AI helped, and how.
- Accounts and access: use company accounts with multi-factor authentication, not personal accounts, for work.
- Incident reporting: what to do and who to tell if someone pastes the wrong thing or an AI behaves oddly.
- Consequences and review: the policy is reviewed twice a year and breaches are handled fairly.
Name an owner
Assign one person to own AI decisions: maintaining the approved list, answering questions and reviewing new requests. In a small business that may be the owner or an operations lead. The NCSC's questions for leaders start with this: who is accountable for AI security?
Make the safe path the easy path
If your approved tool is clunky, staff will use whatever they find. Pick tools that fit the work, give people access and show them how to use them well. Our guides on AI for small business: practical ways to use it safely and writing AI prompts that give useful business results help with adoption.
Decide what data may go into AI tools
This is the most important rule set you will write. A simple traffic-light system works well.
| Level | Examples | Rule | |---|---|---| | Green: public or low risk | Published web content, generic questions, ideas for marketing copy, public product info | Allowed in approved tools | | Amber: internal business information | Internal plans, draft pricing, non-sensitive project notes, anonymized examples | Only in business-grade tools with clear data terms, and with minimum necessary detail | | Red: sensitive or regulated | Customer personal data, employee records, health data, financial account numbers, passwords and keys, legal files, unreleased financials, trade secrets | Do not enter unless the tool is specifically approved and contracted for it and a responsible person has signed off |
Practical ways to reduce exposure
- Remove identifying details. Replace names, emails, account numbers and addresses with placeholders before pasting.
- Use the minimum. Paste only the paragraph you need, not the whole document.
- Never paste secrets. Passwords, API keys, security codes and private links do not belong in any prompt.
- Prefer business plans over free consumer tiers. Business or enterprise plans often have clearer terms on data retention and whether your inputs are used for training, but read the terms rather than assuming.
- Turn off history and training where you can. Many tools provide settings for this. Verify them and record the choice.
- Watch connections. Tools that connect to email, calendars or drives can see everything those accounts can see. Grant the least access that works.
The FTC's guidance for businesses points the same way: collect only what you need, protect it with a written security program and dispose of it securely. Our guide to AI and customer data privacy goes deeper on customer information.
Know your promises
If your privacy policy, contracts or customer promises say you do not share data with third parties, pasting customer data into an outside AI service may break that promise. Check before you adopt a tool, and update your privacy notice when your practices change. Avoid quietly changing your terms to allow broader data use without clear notice to people.
Check outputs before you rely on them
Because AI can be fluent and wrong at the same time, review is not optional.
A simple review checklist
- Facts: verify names, numbers, dates, quotes and citations against a primary source. Fabricated sources are a known problem.
- Logic and math: recompute figures; do not trust totals the tool produced.
- Tone and fit: does it sound like you and suit the audience?
- Compliance: does it make claims you cannot support, or promise things you cannot deliver?
- Confidentiality: does it include or imply any information that should not be shared?
- Bias: does it make assumptions about people based on traits that should not matter?
- Sign-off: has a named person approved it before it is sent or published?
Match the level of review to the stakes
Light review is fine for internal brainstorming. Customer-facing content, financial figures, legal text, health information and anything that affects a person's job, pay or access needs careful review by someone qualified. Where stakes are high, treat AI as a drafting assistant, not an authority.
Keep a record
For important uses, note which tool produced the draft, who reviewed it and what changed. A short log helps when something is questioned later, and it supports honest measurement of whether AI is actually saving you time and money, a topic covered in how to measure AI time and money savings.
Be honest in your marketing
Claims about what AI can do for your customers are advertising claims. Keep them accurate and supportable, and do not overstate what a tool does. Do not use AI to generate fake reviews or testimonials.
Choose and review AI vendors
Before you sign up for a tool or connect it to business data, ask the vendor a short set of questions. Write the answers down.
Vendor questions
- Data use: are my inputs and outputs used to train models? Can I opt out, and is that the default?
- Retention: how long is data stored, and can I delete it on request?
- Location and sharing: where is data processed, and which subprocessors can see it?
- Security: what controls protect accounts and data, such as encryption, multi-factor authentication, access logs and independent audits?
- Access controls: can I limit who in my organization uses it, and what they can connect?
- Incident handling: how will you tell me about a breach, and how fast?
- Admin visibility: can I see usage, set spending limits and export my data?
- Contract terms: what do the terms say about confidentiality, liability and changes to terms?
- Exit plan: how do I get my data out if I leave?
The NCSC's leadership questions include whether you understand your data, model and software supply chains and can question suppliers on their security. You do not need to be an expert; you need to ask and keep the answers.
Start small and review
Pilot a new tool with a small group and low-risk data for a few weeks. Check what it did well, what it got wrong and what surprised you. Re-review vendors at least annually, because terms, features and ownership change.
Watch for "AI inside" features
Existing software now ships AI features that may be switched on by default. When a tool you already use adds AI, check its settings and notices, and update your approved list.
Defend against prompt injection and misuse
Attackers are also using AI, and attackers can target your AI tools.
Prompt injection in plain English
If an AI assistant reads an email, web page or file, hidden text in that content can tell it to do something else, such as reveal information or take an action. The NCSC and OWASP both flag this as a key risk. You cannot fully remove it today, so the safest design is to limit what the AI can do if it is fooled.
Practical defenses
- Limit tool permissions. Give AI the minimum access it needs, read-only where possible.
- Separate trusted and untrusted content. Be cautious when an assistant processes external emails, web pages or uploaded files.
- Require approval for actions. Do not let an AI send messages, move money, change records or delete data without a person confirming.
- Treat output as untrusted input. OWASP lists improper output handling as a risk: do not pass AI output straight into other systems or code without checks.
- Log and review. Keep logs of what the AI accessed and did, and review unusual activity.
- Keep secrets out of reach. Do not store credentials in prompts, notes or documents the AI can read.
Expect better phishing
Attackers can use AI to write more convincing messages, so poor spelling is no longer a reliable warning sign. Strengthen verification instead: confirm payment and bank detail changes by calling a known number, use multi-factor authentication everywhere and teach staff to pause on urgent requests. See spotting and handling phishing and multi-factor authentication and password practice.
Be wary of deepfakes and voice cloning
Fake voices and videos can impersonate executives or customers. Agree a verification routine for sensitive requests, such as a call-back to a known number or a code word, and train staff to use it.
Basics still matter
NIST's small business cybersecurity corner offers quick-start guides and topic pages on phishing, multi-factor authentication, ransomware, cloud security and responding to an incident. Most AI-related breaches begin with ordinary weaknesses such as reused passwords and excessive access. Our cybersecurity basics and backup guide cover them.
Keep humans in charge of AI that acts
A growing class of tools, often called AI agents, do more than answer questions: they can draft and send emails, update records, schedule meetings or run multi-step tasks. That is powerful, and it is where "excessive agency" risks arise.
Principles for safe agents
- Start in suggestion mode. Let the agent propose actions for a person to approve. Move to more autonomy only for low-risk, reversible tasks after it has earned trust.
- Least privilege. Give each agent only the access its job needs, and separate agents by role.
- Human approval for consequential actions. Payments, external messages, deletions and changes to customer or employee records should need a named person's approval.
- Clear limits. Set caps on spending, volume and scope.
- Audit trail. Record what the agent did, why and who approved it.
- Kill switch. Know how to switch an agent off quickly.
- Regular review. Check logs, errors and near misses each month.
Our guide on AI agents and human approval explains how to set these up in practice.
Decide what AI should never do alone
Write a short "never alone" list: hiring or firing decisions, medical or legal advice, financial transfers, changes to contracts, and responses to complaints or regulators. A person handles these, with AI limited to support.
Be careful with people decisions and customer-facing AI
Hiring and performance
Using AI to screen applicants, rank candidates or assess employees can introduce bias and legal risk. Rules on automated employment decisions are emerging in several places and differ by location, so take advice before relying on them. If you use AI as a drafting aid for job descriptions or interview questions, have a person review everything and keep decisions with people. Structured, consistent processes, as in our guides to structured interviewing and fair performance reviews, are safer and fairer than opaque scoring.
Chatbots and customer-facing tools
If you put an AI assistant in front of customers:
- Tell people they are talking to an automated system.
- Limit what it can say and do, and give it only approved information.
- Provide an easy route to a human.
- Test it with awkward and hostile questions before launch.
- Monitor conversations for wrong answers, promises you cannot keep and data leaks.
- Make sure it cannot reveal other customers' information.
The ICO's guidance emphasizes giving people clear explanations of how AI assists or makes decisions that affect them, which is good practice wherever you operate.
Content and intellectual property
Check the terms of any tool for ownership of outputs and for any restrictions on use. Be cautious about publishing AI-generated text or images as if they were wholly original, and avoid prompts that ask a tool to copy a specific person's work. Where copyright or trademark matters to you, ask a lawyer.
Train staff and prepare for incidents
Short, repeated training
Run a 30-minute session when you introduce the policy, and a ten-minute refresher each quarter. Cover the traffic-light data rules, review expectations, how to report problems and a few real examples. Use scenarios: "A customer sends you a contract and asks for a summary. What do you do?" People learn more from practice than from slides.
Plan for incidents
Decide in advance what you will do if someone pastes sensitive data into the wrong tool, an AI sends something wrong to a customer, or an account is compromised. A basic plan includes:
- Stop the activity and contain the access.
- Record what happened, when and what data was involved.
- Contact the vendor and use deletion or support channels.
- Assess whether personal data was involved and whether notification duties apply. In the U.S., breach-notification rules vary by state and sector.
- Fix the cause: update the policy, settings or training.
- Review what you learned at the next meeting.
The FTC's guidance for businesses includes a data breach response guide, and the NCSC advises having a response plan for serious incidents involving AI tools. A tabletop exercise once a year, where you talk through a scenario, costs little and exposes gaps.
Keep records of your decisions
Record which tools are approved, the risk level you assigned, the vendor answers you collected and any exceptions. When a customer, auditor or insurer asks how you manage AI, you will have an answer.
A 30-day starter plan
- Days 1 to 5: List the AI tools staff already use, including free ones and built-in features. Ask openly; the goal is visibility, not blame.
- Days 6 to 10: Write the one-page policy, the traffic-light data table and the never-alone list. Name an owner.
- Days 11 to 15: Pick approved tools, set up business accounts with multi-factor authentication and review data and training settings.
- Days 16 to 20: Ask vendors the key questions and record the answers. Remove or replace tools you cannot vouch for.
- Days 21 to 25: Train the team with scenarios and share the incident reporting route.
- Days 26 to 30: Start a log for important AI uses, set spending limits and schedule a review for next quarter.
Common mistakes
- No policy, so every person invents their own. A short rule set beats silence.
- Pasting customer data into free tools. Most data incidents begin with a well-meaning shortcut.
- Trusting output without checking. Fabricated facts and numbers reach customers this way.
- Giving AI broad access. Wide permissions turn a mistake or an attack into a disaster.
- Letting agents act unsupervised. Keep approval on anything consequential.
- Skipping vendor questions. Ask about data use, retention and security before connecting anything.
- Ignoring built-in AI features. New features arrive switched on inside tools you already own.
- Overstating what AI does in marketing. Keep claims truthful.
- Treating it as a one-time project. Tools, terms and threats change, so review regularly.
Where WorkBOS fits
WorkBOS is a business operating system that keeps projects, CRM, invoicing, HR, chat, drives and automations in one place, with optional AI agents. Those agents act only with human approval, which matches the keep-humans-in-charge principle above, and roles and permissions plus an audit log let you control who can see what and review what happened. Keeping records in one system also makes it easier to apply a single set of access rules rather than managing many separate tools. See the features page, the pricing options (including a free plan for up to five users with no card), the docs or the Academy for details. Whatever tools you choose, the policy, the data rules and the review habits in this guide matter more than any single product.
Frequently asked questions
Q: Is it safe to put customer information into ChatGPT or similar tools? A: Treat customer personal data as red-level by default. Enter it only if the tool is approved for that purpose, its data terms are clear, and a responsible person has signed off. Otherwise remove identifying details or leave the data out. Check what your privacy notice and contracts promise customers.
Q: What is the biggest AI risk for a small business? A: For most, it is staff sharing sensitive data with tools that have not been reviewed, followed closely by relying on wrong output without checking it. Both are controlled with a short policy, clear data rules and a review habit.
Q: What is prompt injection and should I worry about it? A: It is a way of tricking an AI tool, often by hiding instructions in content it reads, into revealing information or doing something unintended. It matters most when an AI can read outside content and take actions. Limit permissions, require approvals for consequential actions and treat AI output as untrusted until checked.
Q: Do I need a formal AI policy if I only have a few employees? A: Yes, but keep it to a page or two. Even tiny teams share data, use free tools and make quick decisions. A short policy sets expectations, protects customers and gives you something to point to later.
Q: How do I know whether a vendor uses my data to train its models? A: Read the terms and privacy documentation, check the account settings and ask the vendor in writing. Business and enterprise plans often have different terms from free consumer ones. Record the answer and re-check when terms change.
Q: Can I use AI to screen job applicants? A: Be careful. AI screening can introduce bias, and rules on automated employment decisions are developing and vary by location. Use AI only for drafting support, keep decisions with trained people, use consistent criteria and seek legal advice before relying on automated scoring.
Q: Should customers be told when AI is involved? A: When customers interact directly with an automated assistant, tell them and offer a route to a human. For AI-assisted content, disclosure expectations vary by context and law, but honesty is a safe rule. Never present AI output as a person's firsthand experience or as a real customer review.
Q: What standards or frameworks can I look at without becoming an expert? A: NIST's AI Risk Management Framework and its Generative AI Profile are good references, OWASP's Top 10 for LLM Applications lists common technical risks, and the UK NCSC and ICO publish accessible guidance on security and data protection. You can borrow their questions for your own policy.
Sources are linked in the text above: NIST, OWASP, UK National Cyber Security Centre, UK Information Commissioner's Office, CISA, Federal Trade Commission.
Run it all from one place
Projects, clients, finance, people and AI employees in one system. Start free, no card needed.
Start free →