How to Implement AI in Your Business: 7 Steps
Most AI projects don't stall on the technology, they stall on the order of things. These seven steps take you from 'we should do something with AI' to a solution your team uses every day.
Implementing AI in an SME: changing how work gets done, not an IT project
Almost every business owner I talk to wants to implement AI in their business. It usually starts the same way: someone buys a ChatGPT or Copilot subscription, a few enthusiastic colleagues start using it, and three months later hardly anyone uses it in a structured way. The tool is there, but the work hasn't changed.
That's rarely down to the technology. It's down to the order. If you start with the tool, you then go looking for a problem that fits it. If you start with the work, you find the places where AI genuinely saves time and errors, and you pick the right solution for them. This plan follows the second route. It's based on how I approach AI implementations at SMEs, and on what I see go wrong when it's done the other way round.
Step 1: Start with the work, not the tool
First, make visible where the time in your business goes. Not in a big process model, but practically: which tasks come back every week, take a lot of hours and follow a recognisable pattern? Think of drafting quotes, answering emails, writing reports, retyping data between systems or sorting customer questions.
Ask the people who do the work. They know exactly which tasks frustrate them and where things get stuck. An hour at a whiteboard with three colleagues often yields more than a month of analysis. Need inspiration? My overview of AI use cases for SMEs lists nine concrete examples with the time saved per week, and applying AI in your business shows per department where AI pays off most.
Step 2: Pick one workflow and make it your model line
The biggest mistake when implementing AI is starting too broad: a bit of AI everywhere, nothing really changed. Pick one workflow from your list and redesign it from start to finish. I call that a model line: one flow, say from enquiry to quote, that is fully redesigned, runs live and has one owner, even when the work crosses several departments.
A model line is not a pilot. A pilot tests a tool next to the existing work and often stalls there. A model line replaces the old way of working for that one flow and shows what the work looks like when it's right. That convinces the rest of your organisation more than any presentation, and your team learns the craft along the way.
A good candidate for your first model line:
- Happens often: daily or weekly, not once a quarter.
- Follows a pattern: the same kind of input leads to the same kind of output.
- Can absorb an AI mistake, because a person still checks the result.
- Is measurable: you know what the flow costs today in hours, lead time or errors.
- Has an owner who can decide about the whole flow, not just their own part of it.
Step 3: Sort out the basics: data, privacy and agreements
Before your team starts putting company data into AI tools, a few things need to be in place. Which data may go in and which may not? Are you using business accounts, where your data isn't used to train models? How do you handle personal data under the GDPR? Using ChatGPT safely in business shows how to set that up in practice.
Write the agreements down in a short AI policy. It doesn't have to be a legal document: one or two pages on what is and isn't allowed, who is responsible and when a person needs to check the output. Don't forget the AI Act either: since February 2025, businesses must take measures to support the AI literacy of staff who work with AI. A good implementation covers that from the start.
Step 4: Choose the right approach: assistant, automation or agent
Not every problem calls for the same solution. Roughly speaking there are three levels, and it pays to choose deliberately:
An AI assistant for your team
ChatGPT, Microsoft Copilot or Claude, set up properly with clear instructions for your work. Quick to start and well suited to writing, summarising and analysing. Which one fits you? I compare them in ChatGPT vs Copilot vs Claude.
An automated workflow
A fixed sequence of steps that runs automatically, such as reading an incoming request, summarising it and adding it to your CRM. Reliable and good for repetitive work with a predictable flow.
An AI agent
A system that chooses its own steps to reach a goal: looking up information, deciding what needs to happen, preparing a draft and escalating to a person when it's unsure. More powerful than a fixed workflow, but it needs more design and oversight. I explain the difference in agentic AI vs AI agents vs generative AI.
Step 5: Redesign the flow and put it live
Don't automate the old process. Walk through the flow with the people who do the work first, and cut what doesn't belong: double entry, waiting for approval, retyping between systems. Only then build the AI support, for the flow as it should run. For what that looks like in a complex process, read the case don't automate the waste in Lead to Order.
Put the first version live in the real work, not in a test setup next to it. Keep a human in the loop: AI makes the draft, a colleague checks and approves it. In the first weeks you quickly see where it works, where it fails, and which instructions or integrations are still missing.
Measure from day one: lead time from start to finish, hours per week, and how often the result needs changing. Agree up front which outcome means scale up, adjust or stop. Then you make that call on numbers, not on enthusiasm.
Step 6: Train your team and make it part of the work
An AI solution nobody uses delivers nothing. Adoption is therefore not a side issue but half of the implementation. Make sure everyone knows why you're doing this, show how it makes their work easier, and give them room to practise. The six AI skills for professionals show what people need to work with it well.
Appoint a point of contact per team who picks up questions and shares good examples. Read more about the leadership that goes with it in leading AI adoption in your organisation.
Step 7: Measure, improve and scale to the next workflow
Is the model line running and hitting the numbers you agreed up front? Then record what it delivers and fix what still chafes. Next, pick the next workflow from your list in step 1. Each flow after that goes faster, because the foundations (agreements, accounts, know-how in your team) are in place and your people learned the work on the model line.
That way AI grows into your business step by step, instead of as one big project that never gets finished.
What does implementing AI cost?
That depends mostly on the approach from step 4. A well set-up AI assistant mainly costs licences and a few days of configuration and training. Your own agent connected to your systems takes more building. Common models are a one-off project, a monthly subscription or a workshop to get started.
In what does AI implementation cost? I lay out the pricing models, the hidden costs and a simple payback calculation.
Do it yourself or get help?
With an AI assistant for your team you can start on your own, especially with this plan. Outside help pays off when you want to connect AI to your own systems, when you want an agent built, or when you simply can't make it happen internally next to the day job. At the companies where I implement AI, that last one is often the real reason: the idea is already there, the time isn't.
If you go that way, my guide on choosing an AI implementation partner helps: agency or freelancer, the questions for the first call and the red flags.
Common mistakes when implementing AI
- Starting with the tool instead of the work that needs to improve.
- Starting everywhere at once, so nothing really gets finished.
- Running a pilot next to the old work that never becomes the real way of working.
- No agreements on data and privacy, so people improvise.
- Automating the old process one-to-one, including all the detours that don't belong in it.
- Not measuring, so after three months nobody knows whether it pays off.
- Not bringing the team along, so the solution goes unused.
Key takeaways
- Implementing AI starts with the work that needs to improve, not with the tool.
- Start with one model line: one workflow redesigned from start to finish, running live, with one owner.
- Agree on data, privacy and AI literacy up front in a short AI policy.
- Choose deliberately between an assistant, a fixed workflow or an agent.
- Cut the waste first, measure from day one, and agree up front when you scale.
Want to implement AI without losing months?
I help SMEs go from first use case to a working solution: setting up tools, building agents and training your team, so you can carry on yourself afterwards.
See AI implementation