Guide · AI in business

Introducing AI in your company: getting started for SMEs

Introducing AI in a company works for SMEs in 6 steps: find a process, clarify data, choose a tool, write an instruction, set rules and measure the result. An AI tool is quickly bought; the first step is the hard part. Here is the path with the templates I use myself, and everything to do it yourself is on this page.

Last changed 2 October 2026

Section on my own use of AI updated.

In short
  • Start with a process, not with a tool: frequent, a lot of text, clear rules.
  • Decide beforehand which data may go into which AI. A data traffic light is enough for that.
  • A saved instruction with fixed limits works more reliably than a spontaneous chat, and a person approves.

The starting point: many tools, little plan

Around 30% of Austrian companies with 10 or more employees use AI; among those with 10 to 49 employees, it is 26% (Statistics Austria and Eurostat, survey 2025). Most companies without AI have not yet considered it at all (77%). Among the specific obstacles, they most often name a lack of expertise (15%), ahead of data protection concerns and legal uncertainty (11% each). And only 10% of companies with 10 or more employees have an AI policy.

This matches what I hear in conversations: there is no plan for where to start, and nobody to set it up cleanly from a technical point of view. That is exactly what this guide is for, written for SMEs without their own IT department.

Step 1: Find the first workflow

Don't start with the question “Which AI should we buy?”, but with “Which process costs us time every week?”. A good first process meets 4 conditions:

  • It comes up often: several times a week, not once a quarter.
  • It consists mainly of text: writing, summarizing, answering.
  • It has clear rules: There is a list of services, a template or a familiar structure.
  • The result can be checked: A person can see in 2 minutes whether it is right.

In almost every small company, 3 processes meet all 4 conditions: writing quotes, answering customer inquiries and creating minutes. Pick one of them. Not all 3 at once.

Step 2: Define the data traffic light

Before anyone on the team uses the first instruction, you decide which data may go into which AI. One page is enough:

LevelWhich dataWhere to
Greenpublic information, your own texts without personal dataevery tool
Yellowinternal documents, quotes, customer inquiriesonly on the business plan with a data processing agreement, EU data processing and training on your data switched off
RedHealth data, client and patient files, access credentials, bank and card data, ID copies, personnel files, salaries and job applicationsnot into a cloud AI, only into a local or specially secured solution

The traffic light is half of an AI policy. It takes the uncertainty away from the team, and you don't have to answer every question individually.

The color says which data may go into which tool.

Step 3: Choose the tool

Your instruction works in every tool that can store a permanent instruction. The feature has a different name depending on the provider (manufacturer information, accessed on 30 September 2026):

ToolThe instruction is hereData in the EU
ChatGPTProjects with project instructionsavailable for Enterprise, Edu and the programming interface (API)
Microsoft CopilotAgents in Copilot ChatEU Data Boundary for EU customers
Gemini (Google Workspace)Gems (according to Google, to become Skills)Data region Europe, depending on the workspace edition
ClaudeProjects with project instructionsaccording to Anthropic, stored in the USA
Le Chat (Vibe) by MistralProjects with instructionsHeadquartered in France, hosting in the EU by default

My recommendation: use the tool you already have in the house. If you work with Microsoft 365, start with Copilot; if you use Google Workspace, start with Gemini. For yellow data, always use the business plan, sign the data processing agreement and check in the table above whether your data stays in the EU. I myself work with Claude every day and run automations on my own server in the EU.

Step 4: Build the first instruction

The difference between “just asking ChatGPT” and a process is the saved instruction. It works by the same rules every time, which is why the quality stays stable. Every good instruction has 5 parts: role, task, input, result and limits. The limits make the difference: they keep the AI from inventing prices or promising dates.

An instruction has 5 parts; the limits keep the AI from making up prices.

My instruction for quotes follows. Replace the placeholders in square brackets and save it as a project instruction, agent or Gem:

Instruction for quote assistant
ROLE
You are the quote assistant of [Firmenname], a company in the [Branche] industry. You write quote texts for this company's customers.

TONE
[e.g.: informal “you”, friendly and matter-of-fact, short sentences, no filler phrases]

SERVICE LIST (the only source for services, units and prices)
[per line: service | unit | net price | short description in 1 sentence]

TASK
I give you keywords from a customer conversation. From them you create the text for a quote.

WHAT I DELIVER TO YOU
Keywords on: customer (name, location), occasion, wishes, quantities and dimensions, desired time frame, special features.

OUTPUT FORMAT
1. Subject: 1 line.
2. Introduction: 2 to 3 sentences that refer to the conversation.
3. Service description: 3 to 5 sentences in everyday language.
4. Items: table with No. | Service | Description | Quantity | Unit | Unit price net | Total net, below it the grand total with a verification calculation.
5. Next steps: 2 to 3 items.
6. Please review (internal only): open questions and everything that isn't in the list of services.

LIMITS
- Prices, units and services come exclusively from the service list. Never invent a price, a service or a discount.
- Include only line items that follow from my keywords. What might be technically necessary but isn't in the keywords, you name in section 6 as a question.
- If my keywords mention something that isn't in the service list, you don't write a line item for it but name it in section 6.
- Don't promise any appointments, deadlines, guarantees or discounts.
- Calculate totals step by step and show the verification calculation.
- If any information is missing, first ask all follow-up questions together (at most 5) and wait for my answer. If all the information is there, you write the quote right away.

Test the instruction on 3 real quotes from the past weeks before the team uses it. Whatever goes wrong becomes a new line under LIMITS. You build the instructions for inquiries and minutes by the same pattern.

Step 5: Write down rules and bring the team along

An instruction that only you know remains a tool for you. To turn it into a process in the company, it takes 3 things:

  • A one-page AI policy: the data traffic light, the permitted tools and the rule that a person checks every result before it leaves the company.
  • A short training session: Everyone who uses the instruction sees it in action once and tries it on a case of their own. An attendance list records who has been trained. This also documents the measures for AI literacy that Article 4 of the AI Regulation requires.
  • One responsible person: They collect what doesn't work and add the limits to the instruction.

Step 6: Measure, then expand

Measure the time for the workflow before the AI takes it over, and 4 weeks after. Without this number you don't know whether it is worth it, and you can't show anyone on the team why you are doing it.

Only when the first process is in place does the next one get added. The step after that is the connection with e-mail, calendar and storage: then an inquiry lands automatically with the inquiry assistant, and the draft is ready in the mailbox. In my own company I use AI daily for research, drafts and analyses. A second AI checks the results, and I give the approval myself. How an agent works in the process, I explain in AI agents explained simply.

Common mistakes when getting started

  • Starting with the tool instead of the workflow: Licenses are bought, but nobody knows what they're for.
  • Everyone tries things on their own: That saves a few minutes here and there, but doesn't change any workflow.
  • Instructions without limits: Then the AI invents prices or promises appointments.
  • Customer data in a private account: Without a business plan, a contract and data processing in the EU, customer data doesn't belong there.
  • Everything at once: 5 workflows at the same time usually end with none of them finished.
Do it yourselfYou have everything you need.

Everything on this page can be implemented without me. I add new tools and changes here, and every Sunday there is a tip about it in the newsletter.

To AI Sync
Have it implementedAI workshop in your own company

I come to your company, from Eisenstadt to Mattersburg and Wiener Neustadt, and build the first instructions with your team on your real documents, including data traffic light, template for the AI policy and training record.

Learn more

Frequently asked questions

How do I get started with AI in my business?

With a single workflow that comes up often, moves a lot of text and has clear rules, such as quotes, replies to inquiries or minutes. For that, you build a saved instruction with fixed boundaries, decide which data may go in, and measure the time before and after.

Which AI tool is right for a small business?

The tool you already have in house or that fits your data. ChatGPT, Microsoft Copilot, Gemini, Claude and Le Chat (Vibe) from Mistral can all store saved instructions. For internal documents, you use the business plan with a data processing agreement and data processing in the EU.

May I enter customer data into ChatGPT?

Not in a personal account. Customer data is yellow data and belongs in a business plan with a data processing agreement, data processing in the EU and training switched off. For ChatGPT, data storage in Europe is available for Enterprise, Edu and the API according to OpenAI; for the Business plan it is not documented. Health data, client and patient files, access credentials, bank and card data as well as personnel and application documents do not belong in a cloud AI.

Do small businesses need an AI strategy?

Not a thick strategy, but 3 decisions: which process first, which data may go into which AI, and who approves. That fits on one page and is the core of every AI strategy for small businesses.

Do employees have to take AI training?

The law does not prescribe a specific training. Under Article 4 of the AI Act (EU) 2024/1689, every company that uses AI professionally takes measures to support the AI literacy of the employees who work with it. This has applied since 2 February 2025, and since 27 July 2026 in the version of Regulation (EU) 2026/1744, which no longer requires a specific level of competence. A short training with an attendance list is a simple proof of this.

Changes to this post

  • 2 October 2026Section on my own use of AI updated.
  • 1 October 2026First version.
Christoph Harrison
Christoph HarrisonAI consultant and network technician, Eisenstadt

Founder of Harrison Networks and Lumen Media. Uses AI daily in his own company and implements it with SMEs in AI advice.

Contact

Let's talk.

About AI in your business, your network or your visibility. A short conversation is usually enough to see where things get stuck.

Christoph Harrisonavailable for new projects