Private AI for Small Business: Keep Data Safe
Whether it is safe to use AI with company data, what should never go into a public tool, the private deployment options, and what GDPR and the EU AI Act ask of a small business.
Reviewed by Igor Gazivoda, Founder & CEO of StepTo · Updated
What Happens to Data Pasted Into Public AI Tools?
It depends on the product and the plan. The text you paste is sent to the provider, kept for a period, and in some consumer products used to improve the models unless you switch that off, while business and API plans are usually designed so that your data is not used for training. Read the current terms of the exact plan your staff use, because they change.
Data protection is a real reason firms hold back. In Eurostat's 2025 survey of EU enterprises that had considered using AI, the most common reason for not adopting it was lack of expertise (70.89%), followed by lack of clarity about the legal consequences (52.52%) and concerns about data protection and privacy (48.83%). Those are manageable with a deliberate setup, which is what the rest of this page covers.
Are Your Staff Already Using AI Without a Policy?
Probably. In the US Census Bureau's 2025-26 supplement, 64% of AI-using firms reported making no institutional adjustments, such as training staff or building new workflows, and use by individual workers sometimes happens without any formal company decision. The safest response is to approve a tool and a policy, not to ban a tool people already use.
The Census Bureau's working paper CES-WP-26-25 notes that worker task use sometimes occurs without formal firm-level adoption. A ban pushes that use onto personal accounts you cannot see; an approved tool and a short policy bring it into the open.
What Should Never Go Into a General-Purpose AI Tool?
Never paste credentials, customer personal data, employee records, documents under NDA, unreleased financials or source code into a tool that is not approved for that data. If you would not email it to a stranger, do not paste it into an account you do not control.
| Do not paste | Why |
|---|---|
| Passwords, API keys and access tokens | A pasted credential must be treated as exposed and replaced. |
| Customer personal data | Names, emails, addresses and identifiers need a lawful basis and a processor contract before they go to an AI provider. |
| Employee records, pay and health information | These are among the most sensitive categories and often carry stricter rules. |
| Contracts and documents under NDA | Your confidentiality duty to the other party does not pause because the tool is convenient. |
| Unreleased financials and strategy | If it would hurt you on a competitor's desk, it should not be pasted into an account you do not control. |
| Source code and trade secrets | Check the tool's data-use terms and your own IP policy first. |
What Are the Private Deployment Options?
StepTo groups private AI for a small business into three deployment options: a business plan or API under no-training terms, a model running in your own cloud account, and a self-hosted open-weight model. Each trades convenience for control, and the right one depends on how sensitive the data is and who is asking where it goes.
| Option | What it means | It fits | Trade-off |
|---|---|---|---|
| Business plan or API with no-training terms | You use a provider's commercial product under a contract that says your data is not used to train its models and sets how long it is kept. | Most internal drafting, summarising and document questions, once the provider acts as your data processor. | Your data still leaves your systems and sits with the provider for the retention period. You rely on the contract. |
| Model in your own cloud account | The model runs inside the cloud account you already control, in a region you choose, with your access rules and logs. | Sensitive documents, customer data, or a customer or regulator who asks where data is processed. | More setup than a SaaS tool, and you pay for usage in your own account. |
| Self-hosted open-weight model | An open-weight model runs on servers you or your host operate, so prompts never go to an AI provider. | The strictest requirements, or very high volume where usage-based pricing is a poor fit. | You carry the operations, updates and security, and quality can trail the best hosted models. |
No option is automatically "EU-only". Where data is processed is a setting you choose and verify in the contract and the cloud account, not a property of where the engineers sit, so ask every provider for the processing region, the retention period and the list of sub-processors.
What Does GDPR Require When AI Processes Personal Data?
The GDPR applies to AI the same way it applies to any other tool: you need a lawful basis and a stated purpose, a processor contract with the provider, only the data the task needs, a check on where data goes, and an impact assessment when the processing is high-risk. This is general information, not legal advice.
Know the role of the provider
When an AI provider handles personal data on your behalf it is your processor. Article 28 of the GDPR says you may only use processors that give sufficient guarantees, and requires a binding contract setting out what they may do with the data.
Have a lawful basis and a purpose
Using personal data in an AI workflow needs a lawful basis under Article 6 and must fit the purpose you told people about. Using customer data to answer customer questions is a different purpose from using it to train a model.
Send only what the task needs
Data minimisation (Article 5) is the cheapest control you have. Strip or mask names, account numbers and identifiers before sending text, where the task does not need them.
Check where data goes
If the provider or its sub-processors are outside the EU or EEA, the transfer rules in Chapter V apply. Ask where processing and storage happen, and for how long data is kept.
Assess the risk when it is high
Large-scale or sensitive processing, or automated decisions that significantly affect people, can require a data protection impact assessment (Article 35) and limits on solely automated decisions (Article 22).
The full text is in the GDPR on EUR-Lex. For the AI-specific rules, the Article 50 transparency post covers disclosure of AI-generated content.
What Does the EU AI Act Ask of a Small Business Using AI?
For a small business that only uses AI tools, the main points are AI literacy for staff and keeping AI out of high-risk uses such as hiring without proper safeguards. Regulation (EU) 2026/1744 changed the Act in 2026: Article 4 now asks providers and deployers to take measures to support AI literacy, and the high-risk obligations for Annex III systems moved to 2 December 2027.
The amending regulation was published in the Official Journal on 24 July 2026 and is available on EUR-Lex. Rules and dates in this area have moved more than once, so confirm the current position with your adviser before relying on a date.
How Do You Control Who Can Ask the AI What?
Make the assistant respect the permissions your files already have: when someone asks a question, search only the documents that person is allowed to open. An assistant that indexes everything into one pool will happily show payroll to the whole team.
Log who asked what, which documents were used for the answer, the answer itself and the model version, so a wrong or improper answer can be traced. Those logs contain business and sometimes personal data, so they need their own access control and a retention limit.
What Should an AI Acceptable-Use Policy Say?
A small team needs one page that names the approved tools, lists what must never be pasted, keeps a person responsible for outputs and decisions, and says who to tell when something goes wrong. Seven short rules cover it.
- 1Use only the AI tools on the approved list, with company accounts, never personal ones.
- 2Never paste passwords, keys, customer personal data, employee records or documents under NDA into a tool that is not approved for them.
- 3Treat every AI answer as a draft. A person checks facts, figures and quotations before they leave the company.
- 4Do not let AI make or send decisions about hiring, pay, credit or contracts without a named person approving them.
- 5Say when AI helped, if a customer or regulator would reasonably expect to know.
- 6Report a mistaken paste or a wrong answer straight away, to a named person, without blame.
- 7Review the approved list and this policy every six months, because the tools and their terms change.
How Can AI Answer From Your Documents Without Training a Model?
Use retrieval: when a question is asked, the system finds the relevant passages in your documents and hands them to the model as context, so the answer cites your own pages and nothing is trained into the model. Delete a document from the index and it stops appearing in answers.
This is the usual way to build an internal assistant over policies, procedures and contracts, and it is what the private deployment options above host. See who builds RAG systems for what to ask a builder, and AI and high-performance software for the technical side. For other internal processes, see AI for small business operations and AI invoice and document processing.
How Does StepTo Start a Private AI Project?
StepTo starts with a 15-minute call about what data the assistant would touch and where it must stay, then proposes one of its engagement models. Engineers start 2-4 weeks after the first call, and an NDA, a data processing agreement and IP assignment terms are signed before any work begins.
The engagement models are staff augmentation from $4,500 per developer per month, a dedicated team of three engineers from $13,500 per month, and fixed-scope project work from $9,000. The first two weeks are a paid trial on real work, and StepTo replaces an engineer free of charge if the fit is wrong within 30 days. Full terms are on the pricing page.
What Are the Four Steps of an Engagement With StepTo?
An engagement with StepTo starts in four steps, whether you add individual engineers or a dedicated team.
Step 1
First call
A first call with StepTo about the roles, stack and seniority you need.
Step 2
Engineers join in 2-4 weeks
StepTo engineers join your team 2-4 weeks after the first call.
Step 3
Two-week paid trial
The first two weeks are a paid trial on real work, before you commit to the full term.
Step 4
Free replacement within 30 days
If an engineer is not the right fit within the first 30 days, StepTo replaces them at no extra cost.
StepTo staff augmentation starts from $4,500 per developer per month, with a 3-month minimum, weekly billing and a 30-day notice period to scale down; a dedicated team of three engineers starts from $13,500 per month, billed monthly. There is no recruitment fee; full terms are on the pricing page.
What Do Small Businesses Ask About AI and Data Safety?
Is it safe to use ChatGPT, Claude or Gemini with company data?
It depends on the product, the plan and the data. Terms differ between consumer apps, business plans and APIs, and they change, so read the current data-use and retention terms of the exact plan your staff use. As a rule, business and API plans are the ones designed for company data, and a consumer account is not.
Is "private AI" the same as running our own model?
No. Private AI is a goal, not a product: AI where your data is not used to train someone else's model and where you control access, location and retention. You can reach it with a business plan under the right contract, a model in your own cloud account or a self-hosted model.
Do we have to train a model on our documents?
Usually not. Retrieval searches your own documents when a question is asked and gives them to the model as context, so the documents stay where they are and can be removed. Training or fine-tuning is a heavier option that is rarely needed for internal questions.
What does the EU AI Act require from a small business that only uses AI tools?
Mostly little today, but check your uses. The Act's AI literacy duty in Article 4 was replaced by Regulation (EU) 2026/1744 with an obligation to take measures to support the AI literacy of staff. AI used in hiring or worker management is classed as high-risk under Annex III, with those obligations now set for 2 December 2027. This page is general information, not legal advice.
Who is responsible if an AI answer is wrong?
Your company, in practice. The customer or employee who relied on the answer deals with you, not the model provider. That is why the policy keeps a person responsible for anything that leaves the company or affects someone.
Can StepTo set up a private AI assistant for our documents?
Yes. StepTo builds retrieval-based assistants and can deploy them in your cloud account or against a self-hosted model. StepTo engineers start 2-4 weeks after the first call, and an NDA, a data processing agreement and IP assignment terms are signed before any work begins. Fixed-scope project work starts from $9,000.
Ready to Use AI Without Losing Control of Your Data?
Book a 15-minute call to talk through where your data sits and what the assistant should be allowed to see, or send the details and StepTo will get back to you within one business day.
Related insights from the StepTo blog
- Three Quarters of Small Businesses Are Using AI. Why Only 14% Are Getting Real Results.
Goldman Sachs: 76% of small businesses use AI tools; only 14% fully integrate AI into operations. Closing that gap takes implementation, not tools.
- The Build-vs-Call Decision: Why Engineering Teams Are Deploying Small Language Models Instead of Calling the API
A fine-tuned 7B model on your own infrastructure can beat GPT-4 on your domain at a fraction of the inference cost. What the SLM shift actually requires.
Let's talk about your project
Tell us what you're building and we'll get back to you within one business day with a no-obligation assessment.
Office hours
Send us a message
StepTo's Belgrade team replies within one business day.
