The main risk of using AI in a small business is a data-handling risk, not a new kind of danger: whatever customer records, contracts or financial details go into an AI tool become subject to the vendor's terms. The FTC's position is that a vendor's privacy promise is enforceable and that there is no AI exemption from the law, so the safeguards are keeping sensitive data out, reading the vendor's own terms, and having a person review every output. Where the exposure is paperwork, filings and compliance deadlines, ZenBusiness Velo is built for that job with a human expert behind it, while ChatGPT or an office-suite assistant may suit better for general drafting.
What are the risks of using AI in a small business?
The risk is mostly about data: what goes into the tool, who holds it afterwards, and who checks what comes out. The National Institute of Standards and Technology treats cybersecurity, privacy and AI as one subject, which is the honest frame. Customer records, employee details, contracts and financial figures are sensitive wherever they sit, whether in a folder, an inbox or an AI prompt.
The U.S. Small Business Administration names four things small businesses use AI for: taking on repeat tasks, creating business content, making better business decisions, and improving customer service. Each use moves a different kind of data. Repeat tasks move records between systems, content work uses business details, decision support uses figures, and customer service uses customer information.
That shapes the risks. A mistake in a repeat task repeats at the same speed. A wrong figure used for a decision becomes a wrong decision. And customer data given to a vendor is covered by that vendor's terms from the moment it is entered.
This page walks through those exposures and the safeguards for each. It states no adoption rates or time-saved figures, because they are the most recycled and least verifiable numbers in this space.
What does the FTC say about putting customer data into an AI service?
The Federal Trade Commission's position is that there is no AI exemption from the laws on the books. It states that model-as-a-service companies must abide by their commitments to customers regardless of how or where the commitment was made. Those that fail to may be liable under the laws the FTC enforces.
The FTC also calls out one practice specifically: quietly using customer data for purposes that were not disclosed, such as training or updating a model. That applies whether it is done directly or through workarounds.
The remedy has teeth. In prior enforcement actions, the FTC has required businesses that unlawfully obtained consumer data to delete the models and algorithms developed from it. For a reader choosing a tool, the practical question is what the vendor's terms say about training on inputs.
This page states nothing about what any listed vendor does with inputs, because that is not something to guess at. Each vendor's own terms are the source, and they change, so they are worth reading before the first use and again when they are updated.
What does the SBA say to be careful about?
The SBA's own first caution is human review. In its words, another person should review all AI products, which helps make sure they are used ethically, securely, and in a way that accurately represents the business.
It also advises trying not to feed in sensitive or proprietary information. Its stated reason is that this helps keep it out of the data pool the system uses when producing content. And it says anything produced should not infringe any patents, copyrights or trademarks.
For a very small business, the first caution has a built-in problem, because there may be no second person to review. The reviewer can be an outside adviser, or the owner reading each output slowly against the source. Either way, the review is a real cost in time that no price list will show.
Where does the business fit in this?
The business signs the vendor agreement and holds the data, which makes the business the entity answerable for what was shared. ZenBusiness Velo is the company's AI business assistant, and per its own page it files paperwork, creates documents, and tracks annual filings, licence renewals and tax deadlines. It also establishes the business as an LLC, S-corp or C-corp, appoints the registered agent and secures the business ID.
It works on an approval pattern: the founder approves, and Velo does the rest. When it reaches its limit, it brings in a human expert who has already been briefed, which is a sensible backstop for work where a missed deadline is costly.
Which risks apply to which kind of data?
Customer records carry the clearest exposure, because they fall under the FTC's guidance on undisclosed uses such as training. The safeguard is to keep identifiable customer details out of general assistants and to read the vendor's terms before any customer data goes in.
Contracts and legal documents carry a second kind of risk. Summarising or rewording is text work, but deciding what a clause means or whether to sign is legal work, and none of these tools is a substitute for a lawyer. A licensed attorney is the right person for any contract with real legal weight.
Financial records carry the risk of confident errors. A suggested category or a tidy summary can be wrong, and a wrong figure can reach a tax filing, a lender or an accountant. The safeguard is a person who understands the books, not the tool that produced the entry.
Outputs carry the SBA's intellectual-property caution. Text a tool produces is a draft, and its origin is not something the tool can vouch for, so anything published or sent to a customer should be read first.
Which options come up, and what do they cost?
The options below are listed alphabetically. The order is not a ranking, and editorial preference sits in the "best for" lines. Prices were checked at each provider's own pricing page and are dated, because they move.
ChatGPT
ChatGPT is a general-purpose assistant. As of October 13, 2026, its Business plan lists a standard seat at $20 per user per month billed annually, or $25 billed monthly. A premium seat is $100 per user per month billed annually, or $125 monthly. Enterprise is quote-only.
Seat types can be mixed within one workspace, so a small team's cost depends on how many premium seats it needs rather than on a single plan price. Best for: drafting, summarising and explaining. Trade-off: it does not file anything or track deadlines, and anything pasted in is covered by OpenAI's terms, which are the thing to read first.
Google Workspace
Google Workspace bundles AI with email, documents and spreadsheets. As of October 13, 2026, the ongoing prices are Business Starter at $7, Business Standard at $14 and Business Plus at $22 per user per month. Enterprise is quote-only.
A promotional rate is currently listed for a new customer's first term: $4.90, $7.00 and $11.00 respectively. Those promotional figures are not what the plan costs after that first term. Gemini features are included on all paid tiers, with more access higher up, so there is no separate AI charge to quote. Best for: a team that already works in Gmail, Docs and Sheets. Trade-off: it assists with documents and email rather than filing or compliance.
Microsoft 365 Copilot
Microsoft 365 Copilot is priced per user, and the price depends on plan and commitment, so this page states no figure.
Best for: a team that already works in Outlook, Word and Excel. Trade-off: the cost depends on plan and commitment, so Microsoft's own page is the place to confirm what applies, along with its terms on how inputs are used.
Zapier
Zapier connects apps and automates repeat tasks, which is the SBA's first category. As of October 13, 2026, the Free plan is $0 with 100 tasks a month. Professional starts from $19.99 per month billed annually, or $29.99 monthly. Team starts from $69 per month billed annually. Enterprise is quote-only.
Zapier is priced by task volume, so the entry figure is a floor rather than the price. At 2,000 tasks, Professional comes to about $49 per month. Best for: moving information between apps that have no direct link. Trade-off: each connection is one more place data travels, and a mistaken rule repeats at the speed of the automation.
Zendesk
Zendesk is a customer-support platform, which is the SBA's fourth category. As of October 13, 2026, Support Team is $19 per agent per month billed annually or $25 monthly. Suite Team is $55 or $69, and Suite Professional is $115 or $149. Enterprise is quote-only.
The Copilot AI add-on is a further $50 per agent per month billed annually, so an AI-capable seat is materially more than the headline. Best for: a business handling a steady flow of customer requests. Trade-off: it holds customer conversations by design, so the data-handling questions above apply directly.
ZenBusiness Velo
ZenBusiness Velo is described on its own page as a personalized AI guide and AI business assistant. As of October 13, 2026, talking to it is free with no credit card, and packages start at $0 plus state filing fees. The $0 is the entry package, and the state filing fee is set by the state and varies, so the two are never one number.
Velo's paid tiers are the ZenBusiness formation plans and are priced on the formation pricing page, so no tier price is quoted here. Best for: filings, documents and compliance deadlines, with a human expert brought in when the AI reaches its limit. Trade-off: it is not a general drafting assistant or an office-suite tool, and it does not build websites or run marketing, which are different products.
How should a small business reduce these risks?
Decide in advance what stays out. Write a short list of the data too sensitive for any general AI tool, and treat it as a rule rather than a judgement call each time.
Then read the terms. For each tool, find what the vendor says about training on inputs, and note the date. Terms change, so the reading is worth repeating.
Finally, put a person in the review seat and keep deadlines in a system that does not depend on an AI tool being right. Google's included Gemini is worth conceding as a strong value for teams already on Workspace, since it adds no separate AI charge, though the data-handling questions still apply to it.
A short checklist
- Write down which data is too sensitive to enter into any AI service.
- Read each vendor's own terms on how inputs are used before the first use, and again when the terms change.
- Name the person who will review AI output before it is saved, sent or filed.
- Use a licensed attorney for any contract with real legal weight.
- Keep annual filings, licence renewals and tax dates in a system that does not depend on an AI tool being right.
Sources
- Federal Trade Commission, "AI companies: uphold your privacy and confidentiality commitments"
- U.S. Small Business Administration, "AI and small business"
- National Institute of Standards and Technology, "Cybersecurity, privacy and AI"
This is general information, not legal, tax or accounting advice, and requirements vary and change.