95% of all AI pilots deliver no measurable impact (MIT, 2025). Not because AI doesn’t work, but because companies start wrong. They buy AI tools without a plan. They start pilots without an end goal. They write policy documents that no one reads.
Meanwhile, 91% of organizations that do use AI structurally report revenue growth. The difference is not in budget or technical knowledge. It is in approach.
These are the 7 steps that work. No theory, we use them ourselves. Where we are still figuring it out, we say so.
The 7 steps in brief:
- Start with one AI task, not a strategy
- Deploy AI for low-risk back-office work
- Give your team explicit permission to experiment
- Share prompts that work in a shared document
- Make three simple agreements about AI use
- Only measure after the first month, start with hours
- Make AI the standard way of working, not an experiment
1. Start with one AI task, not a strategy
Sound familiar? You read an article about AI in the evening. You think: “we need to do something with this.” The next morning the daily grind begins and it fades into the background. Same ritual the next week.
82% of small businesses think AI is not relevant to them (OECD, 2025). At the same time, the average employee who uses AI saves 5.6 hours per week. Managers even 7.2 hours.
The problem is not that you know too little about implementing AI. The problem is that you don’t start.
What you can do tomorrow: Choose one task you have to do this week anyway. Do it yourself first. Then do it with ChatGPT or Claude. Compare the result. That’s all.
Not ten tasks. Not an AI strategy. One task, twice.
2. Deploy AI for the work no one wants to do
Most organizations start with the wrong thing. They want to deploy AI for customer contact, marketing, strategy. The sexy stuff.
The data says something else. The biggest AI ROI for organizations lies in back-office work: invoicing, reporting, data entry, summaries of long documents. The work everyone puts off.
Why? Two reasons:
- Low risk. If AI botches an internal summary, no one notices but you. If AI botches a customer email, you have a problem.
- Directly measurable. “This took 2 hours, now 20 minutes” is easier to prove than “our marketing is 12% more effective.”
AI tools start at 0 to 20 euros per month per user. Less than an hour of freelancer rate. The investment is not the subscription, it is the time to learn.
A practical example from ourselves
When building ibgids.nl we had to categorize thousands of company profiles: which company does pentesting, which does GRC, which does awareness training. Manual work no one felt like doing. AI classification did it in a fraction of the time. Not perfect, we had to correct manually, but the most boring part of the work was gone. That was the moment we understood where AI delivers the most.
3. Give your team permission to experiment with AI
This is where it stalls at most organizations. Not at the technology. At the culture.
31% of employees actively sabotage AI initiatives (Built In, 2025). Among younger employees that is even 41%. Not because they are lazy, but because they feel threatened. A third of your team thinks AI undermines their value.
The solution is not a mandatory AI mandate from above. That makes it worse. The solution is three things:
- Make it explicit. “You may use AI for X. Not for Y.” Clarity reduces fear. But watch which tools: 12% of AI skills turn out to be malicious.
- Involve the skeptics. Their concerns point to blind spots your enthusiastic colleagues overlook.
- Appoint an ambassador. One person who helps colleagues, not the most technical, but the most enthusiastic. Give that person 2 hours per week.
48% of employees would use AI more often if they received formal training (McKinsey, 2025). Not a three-day course. Training in the workflow: “Here, try this with your next report.” Gen Z employees are your secret weapon here: 62% already actively coach older colleagues in AI tools.
What we noticed when making our own team AI-native: the biggest breakthrough came not from the techies, but from the people who were most skeptical. As soon as they saw that AI made their work easier instead of redundant, they became the biggest ambassadors. The key was letting them choose which task they wanted to try, not prescribing it.
4. Learn to write AI prompts by stealing
You don’t have to become a prompt expert. You don’t have to be creative. You only have to copy what works.
Start a shared document, Google Doc, Notion, doesn’t matter, where your team shares prompts that worked well. No perfection. The rough versions are precisely what help.
“I used this prompt to summarize a client meeting and the result was good” is more valuable than a three-hour prompt-engineering course. Why? Because it has context. It comes from a colleague who does the same work as you.
A team that shares prompts learns exponentially faster than individuals figuring it out alone. And it costs nothing except the discipline to keep it up.
We maintain an internal prompt library. Not fancy, just a shared document with prompts that work. The most used one: a prompt for analyzing security reports and extracting the five most important risks. It took someone fifteen minutes to perfect, and now it saves everyone who uses it an hour per report. The beautiful thing: once your team sees that sharing pays off, it happens by itself.
5. Build AI guardrails for your organization, not bureaucracy
You don’t need a 40-page AI policy. You need three agreements:
- What can and cannot go into company data? Public AI tools (free ChatGPT) are fine for general tasks. Customer data, financial information, contracts? For that you use a business account with data protection. For sensitive data you can also consider a local LLM.
- Who checks AI output before it goes to a customer? Always someone. Always.
- What do we do when it goes wrong? Not if, but when. A simple escalation path.
That is enough to start. You build the rest when you need it.
BCG calls this the 10-20-70 rule: spend 10% of your energy on technology, 20% on data and tooling, and 70% on people and processes. Most companies flip that ratio. They fail. What goes wrong without policy is described in our article on vibe coding security.
Our own rules fit on half an A4. No customer data in free tools. All AI output that goes out is checked by someone. And if something goes wrong, we know who to call. That was enough to start. We have tightened it twice since then, but the foundation has stayed the same. Start simple, expand when you run into something.
6. Measure what AI delivers (but not too early)
The temptation is great to ask after two weeks: “What does it deliver?” That is too early.
60-80% of every AI project goes to preparation and learning (PwC, 2025). The first month is an investment in understanding, not in results. If you want to calculate ROI on that, you will be disappointed.
Our advice:
- Month 1: Don’t measure. Just learn and experiment.
- Month 2-3: Track where AI saves time. Simple: “This took X hours, now Y.”
- Month 4+: Quantify. How many hours per week? Which tasks? What do people do with the freed-up time?
Don’t measure in euros. Measure in hours. Measure in: “we didn’t do this, now we do.” The euros come by themselves once the hours add up.
We measure it simply when implementing AI: tracking hours per task, before and after AI. No dashboard, no tool, a spreadsheet. After three months we saw a pattern: research tasks went from hours to minutes, but creative writing stayed almost the same; though it did get better. That insight helped us focus on where AI really makes a difference instead of forcing it in everywhere.
7. Make AI the standard, not the experiment
There comes a tipping point. The moment someone on your team says: “Have you run this through AI yet?” instead of “Shall we try AI?”
That is the moment AI stops being an experiment and becomes a way of working.
You can’t force this. It happens by itself if steps 1 through 6 work. But you can accelerate it:
- Use AI yourself. If you as a director don’t use AI, your team won’t either. Companies where leaders actively use AI perform 3x better (McKinsey, 2025).
- Buy, don’t build. MIT research shows that buying AI tools succeeds 3x more often than building them yourself. Use AI features in your existing software (CRM, accounting, helpdesk) instead of developing your own tools.
- Think in roles, not tools. The real shift is not “we use AI” but “our role is changing.” From executor to someone who steers AI. The difference between typing a prompt and designing a workflow in which AI does 60% of the work.
We are not all the way there ourselves yet. For some tasks AI is already the standard, for research, classification, code reviews. For other things we are still experimenting. The difference from a year ago: no one asks anymore whether we use AI. The question now is always how. That is the tipping point.
What we don’t know yet
We are honest about this: we haven’t figured everything out.
AI agents, autonomous AI that performs tasks independently, are rapidly becoming mainstream. Gartner predicts that 40% of all business software will contain AI agents by the end of 2026. But there is a world of difference between letting the AI run and building responsibly with AI agents. What that concretely means for a company with 20 employees? We don’t know exactly yet.
What we do know: organizations have a catch-up advantage. No legacy systems, no complex governance, short lines. You can become AI-native faster than a company with 10,000 employees. The adoption gap with large companies has almost disappeared.
And want to objectively assess whether an AI system really does what it promises? The AI Periodic Table gives you the language and framework to do so.
The question is no longer whether you, as an organization, start with AI. The question is whether you do it smartly, or whether you belong to the 95% that keeps piloting without results.
Start with step 1. One AI task, twice. The rest follows.
Need help?
Want to know what becoming AI-native looks like for your business? We look together at where AI delivers the most and which steps make sense for you. No sales pitch, just an honest conversation.
Sources: MIT Sloan, Why So Many AI Pilots Fail, McKinsey, The State of AI 2025, BCG, AI at Scale, OECD, SME AI Adoption, PwC, AI Business Survey 2025, Gartner, AI Agent Predictions, Built In, AI Workplace Resistance