Implementing AI in your business: the step-by-step plan

An opportunity without a number is an opinion. So give every step in the process three numbers: how long does it take, with how many people, and how often does it happen?

An example. Order entry takes 40 minutes each time, 2 people work on it and it happens 80 times a month. That's 6,400 minutes, over 106 hours a month. At a labour cost of € 45 an hour, this process costs € 57,600 a year. Now you have something to talk about.

Measure what goes wrong too. How often is an order wrong? How long does a customer wait for an answer? What does a mistake cost if it only surfaces at delivery? Many businesses forget error costs and lead time, even though they often deliver more than the hours saved.

Look at the organization as well

Alongside the processes, we measure the organization's AI maturity on seven dimensions: strategy, data, technology, people, processes, governance and culture. The weakest one usually decides where you start. A brilliant opportunity is pointless if the data is spread across five Excel files on someone's laptop.

Want a first indication? Take the free self-scan on this site: seven questions, two minutes.

With a list of opportunities on the table, it's tempting to do everything at once. Don't. Put the opportunities on an impact-effort matrix: what does it deliver, and how much effort does it take? Start with a process that delivers a lot for little effort. Big projects come later, once the foundation is in place.

Calculate like an accountant, not a salesperson

• Never 100 percent. There's always checking and exception handling left. We calculate with at most 90 percent saving per step.
• Three scenarios. Cautious, expected and ambitious. You decide on the expected one.
• A brake when in doubt. Data missing, or no examples of good work? Then cap confidence at 40 percent.

Back to order entry. Say AI takes over 60 percent of the work in the expected scenario. That saves you € 34,560 a year. If building and running it costs less than half of that, you have a business case.

Choose the right build layer

Not every opportunity needs to be built. We distinguish four layers:

1. 01Prompts and knowledge: working better with ChatGPT, Copilot or Claude
2. 02Skills and assistants: your own assistants that know how you work
3. 03Workflow automation: processes that run by themselves, often in n8n
4. 04Custom builds: AI agents and software for what no standard tool can do

Your team can learn layers 1 and 2 itself. Layers 3 and 4 you build with a partner. The cheapest solution that works wins.

AI is only as good as what you put in. That's not a cliché, it's the reason most pilots stall. Before you build, look at three things.

1. Where is the data?

Which system holds the customers, the items, the prices? Is there an API or an export? Is the data clean, or does every customer have three spellings? AI can only put an order into the ERP properly if the item codes are right.

2. Do you have examples of good work?

If you want AI to draft quotes like your best account manager, you need a pile of good quotes for it to learn from. Without examples, AI guesses at your standard. That's why, when examples are missing, we calculate with at most 40 percent confidence.

3. What are the ground rules?

Agree which tools are allowed, which data can go in and where a human needs to check. Use business licences and data processing agreements, and choose EU hosting where you can. Writing those rules down is part of our AI consulting work.

At NinA, data stays in the EU, on the client's own infrastructure or on servers with a European provider. We work to ISO 27001 and NEN 7510 (certification in preparation). Azure OpenAI is an option if you already work in Microsoft.

Now it's time to build. But small: one process, one team, real data. No demo with made-up examples, because those always work.

Start with 'human approves'

We distinguish four roles for human and AI. The work stays with the human. AI assists: the human does it, AI provides a first draft. Human approves: AI does the work, a human checks before it goes out. Or fully AI, with spot checks.

Almost every new process starts at 'human approves'. The employee sees what AI suggests and approves or adjusts it. That way you learn where it works and where it doesn't, and trust grows on the basis of what you see. Only after months of it going well does a step move to fully AI. A conversation with an angry customer never moves.

Connect to what's already there

AI only really works when it's inside your own systems: Exact, AFAS, SAP, Microsoft Dynamics, Salesforce, HubSpot, Microsoft 365, Google Workspace, your WMS or TMS. A standalone chatbot on the side is just one more screen. We connect a lot through n8n, and build custom where needed.

Think about production from day one

A pilot isn't the finish line. Agree upfront who owns it, where it runs, who monitors it and what happens when a system changes. Otherwise, three months from now, you'll have a working pilot that nobody dares switch on.

This is the step that gets skipped most often, and the one that decides whether it works. AI projects rarely fail on technology. They fail on the work floor: nobody has time, nobody knows what's running, and after six months the team is back to the old way.

Be honest about what it's for

People worry about their jobs, even if nobody says so out loud. Be clear: AI takes over tasks, not people. The goal is to grow without hiring ten extra people, and less retyping for the people already there.

Train on your own work

A generic AI demo is forgotten within a week. Training with your own quotes, your own customer questions and your own systems sticks. A half-day workshop gets a team to the same level.

Create ambassadors

Find one or two people per team who are keen. They test first, hear the complaints and help colleagues. Involve the people who do the work today as well. They know the exceptions that appear in no process diagram.

The law asks for it too

Since 2 February 2025, the EU AI Act (Article 4) has asked organizations to take measures for the AI literacy of the people who work with AI on their behalf. Training on your own work is a logical way to do that. Keep a record of who has had which training. More on what the law asks is on our AI training page. If you're unsure about your obligations, ask a lawyer.

The first process is running. Now the real work begins: keeping it running, showing what it delivers and picking up the next process.

Measure every month

How many hours did it save? How many fewer errors? How quickly does the customer get an answer? Compare it with the numbers from step 2. Show it to the team and to leadership. Nothing helps adoption like a colleague saying: 'I spend four hours a week less on invoices.'

Sort out maintenance

AI models get updated, APIs change, a vendor modifies its system. Someone has to monitor it, update it and handle backups. If you don't, the process falls over after six months and nobody knows why. At NinA, that's covered in the partnership or a maintenance contract.

Scale up along the roadmap

Then the next process on the list. That's how you move through the four phases of AI adoption: from loose prompts to assistants, to workflows and eventually to agents that plan their own steps. Most organizations are in phase one. The value is in phases three and four.

After twelve months, redo the baseline. Where are you now, and what's worth the most next year?