AI for Small Businesses: A Practical 90-Day Adoption Plan
A step-by-step plan to adopt AI in a company under 50 people: choose tasks, sort your data, write a one-page policy, run a pilot and check the rules.

Most advice on AI is written for companies with a data team and a legal department. This guide is for a company of fewer than 50 people, including a solo consultant, that wants to use AI tools without creating new risks for itself or its clients.
You get a five-step plan spread over 90 days, a one-page policy you can copy, a worked example for measuring a pilot, and a table of rules to check by country. Laws on AI and personal data differ by country, so the legal points below are examples to verify, not legal advice.
Where small companies stand today
Official surveys show the same pattern: the smaller the company, the less likely it is to use AI.
| Survey | Smaller firms using AI | Larger firms using AI |
|---|---|---|
| EU, 2025 (Eurostat, EU, 2025) | 17% of small enterprises (10 to 49 employees) | 55.03% of large enterprises |
| OECD countries, 2024 or latest year (OECD, 2025) | 11.9% of firms with 10 to 49 employees | 40% of firms with 250 or more |
| US, data to May 3, 2026 (Census Bureau, US, 2026) | Under 20% of firms with 4 or fewer employees | 37% of firms with 250 or more |
These surveys use different questions and dates, so read the gap inside each row and do not compare one row with another. The Eurostat survey covers enterprises with 10 or more employees, so it says nothing about smaller ones. Across all sizes it found that 20.0% of EU enterprises used AI in 2025, up from 13.5% in 2024 (Eurostat, EU, 2025).
Two OECD findings matter more than the adoption rate. In an OECD survey of SMEs in four G7 countries, 50% said their employees lack the skills to use generative AI, and under 30% of SMEs using generative AI reported AI-related training for employees (OECD, 2025). Access to a tool is rarely the obstacle. Knowing what to use it for, and checking what it produces, is.
The 90-day plan at a glance
| Days | What you do | Result |
|---|---|---|
| 1 to 15 | Steps 1 to 3: choose tasks, sort data, write the policy | A short task list and a one-page policy |
| 16 to 45 | Step 4: run a 30-day pilot on two or three tasks | Measured time saved and error counts |
| 46 to 90 | Step 5: train the team, keep or drop each task, add the next ones | A training record and a reviewed policy |
The 90 days are a suggested rhythm, not a rule. A solo consultant can do steps 1 to 3 in an afternoon.
Step 1: Choose tasks before tools
List the recurring tasks in your business, then test each one against four questions. Start with the two or three that pass all four.
| Question | Good first task | Poor first task |
|---|---|---|
| How often do you do it? | Weekly or more | Once a year |
| What happens if the output is wrong? | You catch it in review at low cost | A client, candidate or authority is affected directly |
| Can a competent person check it fast? | Yes, in minutes | Checking takes as long as doing it |
| What data does it need? | Public or non-confidential | Client secrets, personal, health or financial records |
The OECD reports that SMEs typically use generative AI for writing, marketing, online sales and internal process work (OECD, 2025). Common first tasks are first drafts of proposals and emails, meeting summaries, reformatting documents, routine translation and drafting job posts.
Step 2: Sort your data into three lanes
Most incidents with AI tools come from what people paste into them. Decide in advance what may go where.
| Lane | Examples | Rule |
|---|---|---|
| Green | Published content, your own templates, anonymized notes | Any approved tool |
| Amber | Client documents, internal financials, names and contact details | Only tools under a business contract you have read; remove names where you can |
| Red | Material under an NDA that bars third-party processing, health data, passwords, payment details | Never, unless the client agrees in writing and you have checked the law |
Before approving a tool for amber data, read its terms for three points: whether your inputs are used to train the model, where data is stored, and how long it is kept. Then reread your client contracts. A confidentiality clause may already restrict sending client material to an outside service; see the contract clauses checklist.
Step 3: Write a one-page AI policy
A policy that fits on one page gets read. Copy this template and fill in the brackets.
AI use policy
- Approved tools: [list]. Any other tool needs approval from [name].
- Data: green data in any approved tool; amber data only in [tool]; red data never.
- Review: a named person reads and corrects every AI output before it leaves the company. The sender stays responsible for the content.
- Facts: figures, quotes and legal references are checked against the original source.
- Disclosure: we tell clients when [for example: a chatbot answers them, or a deliverable is mostly AI-generated].
- People: no decision about a person (hiring, pay, credit) rests on an AI output alone.
- Problems: report errors or data leaks to [name] within [number] hours.
- Next review: [date].
Step 4: Run a 30-day pilot and measure it
Before the pilot, time each chosen task the usual way for a week. During the pilot, record the time spent including review, and the number of corrections needed.
Example, with assumptions. A four-person agency writes 8 proposals a month at 3 hours each. In the pilot, AI drafting plus human review takes 1.75 hours per proposal. Assumptions: an hour of staff time costs the agency $80, the tool costs $25 a month for one user, and setup plus writing the policy took 6 hours once.
- Time saved: 8 × 1.25 hours = 10 hours a month, worth 10 × $80 = $800.
- Month 1 cost: $25 + (6 × $80) = $505. Net result: $295.
- Later months: $800 − $25 = $775, if quality holds.
Replace every number with your own; the tool price here is an assumption, not a quote. To value an hour of your time, use the day rate calculator. If review time cancels the saving, or errors reach clients, drop the task and test another one.
Step 5: Train people and keep a record
Training can be a one-hour session: how the tools work and fail, your three data lanes, the review rule, and two worked examples from your own pilot. Note the date, the content and who attended.
In the EU this is also a legal matter. The AI Act's AI literacy provision has applied since 2 February 2025, and it is among the rules enforced from 2 August 2026 (European Commission, EU). The Commission's Q&A says organizations should at least build a general understanding of AI, know whether they are a provider or a deployer, and assess the risks of the systems they use; no certificate is required, and an internal record of training is enough (European Commission Q&A, EU, 2025). It also covers people who work under your remit, such as contractors.
Rules to check, by country
No single rulebook applies everywhere. These are starting points from official sources; confirm what applies to you and your clients.
| Where | What to check |
|---|---|
| EU (including France) | The AI Act applies in stages. Transparency rules have applied since 2 August 2026: for example, people must be informed when they interact directly with an AI system (Article 50). Rules for high-risk systems listed in Annex III apply from 2 December 2027 (European Commission, EU). |
| UK | If personal data goes into an AI tool, the regulator expects a lawful basis for each purpose, roles set out in a contract with the tool provider, updated privacy information, and accuracy checks on outputs. A provider outside the UK means a restricted transfer with safeguards (ICO, UK). |
| US | The NIST AI Risk Management Framework is voluntary and written for organizations of all sizes; a revision is in progress (NIST, US). Marketing claims about AI fall under consumer protection law: the FTC announced five enforcement actions over deceptive AI claims and schemes (FTC, US, 2024). State rules vary. |
| India | The Digital Personal Data Protection Rules were notified on 14 November 2025 with an 18-month phased compliance period (Government of India, 2025). Check how they apply to personal data you put into AI tools. |
If you buy from, or sell to, a company that uses AI
Buyers: add four questions to your usual checks. Which parts of the work use AI? Who reviews the output? Which tools will receive our data, and under what terms? Who is liable for an AI error? Put the answers in the contract. Our guide on vetting a consultant or agency covers the rest, and you can compare providers by specialty on Sendbiz.
Providers: answer those four questions before the client asks, in your proposal. Describe what AI does in your process in plain terms, and do not label a service "AI-powered" unless you can show what that means. If you subcontract, pass your data lanes and review rule on to your subcontractors in writing.
Common mistakes
- Buying a tool first. Without a chosen task and a baseline time, you cannot tell whether it helped.
- Pasting client data into a free account. Read the terms and your client contract before any amber data goes in.
- Skipping review. AI output can state wrong figures and invented references with confidence. One named person checks before anything is sent.
- Counting drafting time only. Review and correction time belongs in the measurement.
- Writing a ten-page policy. Nobody reads it. One page, reviewed every six months, works better.
- Training nobody. The OECD figures above show this is the usual gap, and in the EU it is also a legal expectation.
- Assuming one country's rules apply everywhere. If you serve clients abroad, check their country's rules too.
- Overstating AI in your marketing. Claims you cannot support create legal and reputational risk.
Frequently asked questions
How should a small business start using AI?
Pick two or three frequent, low-risk tasks that a person can check quickly, decide which data may go into which tool, and run a 30-day pilot with a measured baseline. Choose the tool after the tasks, not before.
Does a small company need a written AI policy?
A one-page policy is worth having even for a team of two: approved tools, data rules, who reviews outputs and who to tell when something goes wrong. Clients increasingly ask how their data is handled, and the policy gives you the answer.
Does the EU AI Act apply to small businesses?
The AI Act's obligations depend on your role (provider or deployer) and on the type of AI system, and its AI literacy provision has applied since 2 February 2025. Check the European Commission's timeline and guidance for what applies to your case.
Can I put client data into an AI tool?
Only if your contract with the client allows it, the tool's terms fit (training use, storage location, retention) and the personal data rules of the countries involved are respected. When in doubt, anonymize the material or ask the client in writing.
How do I measure whether AI is worth it for my business?
Time the task before the pilot, then record time spent including review, the number of corrections and the tool cost. Keep the task only if the saving survives review time and quality does not drop.
Sources
- 20% of EU enterprises use AI technologies, Eurostat (2025)
- Use of artificial intelligence in enterprises, Eurostat Statistics Explained (2025)
- Large Firms With at Least 20 Employees Biggest AI Users, U.S. Census Bureau (2026)
- AI adoption by small and medium-sized enterprises, OECD discussion paper for the G7 (2025)
- AI adoption by small and medium-sized enterprises, OECD (2025), PDF copy read
- Timeline for the Implementation of the EU AI Act, European Commission AI Act Service Desk
- Article 50: Transparency obligations, European Commission AI Act Service Desk
- AI Literacy - Questions & Answers, European Commission (2025), copy hosted by Sidley
- Previously asked questions (innovation advice), Information Commissioner's Office
Consulting and Freelance Contracts: Clauses That Protect You
Independent Contractor or Employee? The IRS Test Explained