By Jordi Daniëls, founder of AI Board · ·
Implementing AI in an SMB takes four steps, in this order. One: a safe place where AI can reach your company data. Two: findable documents. Three: one decision-maker who uses it daily. Four: only then an engineer for integrations.
The order is not a detail. Start with technology and hiring, and you end up with an engineer and no user. Below, per step: what you do, what it costs, how long it takes and when to bring in help.
What is AI implementation?
Implementing AI means setting AI up so it works alongside you every day, with access to your own company data and a fixed user. The difference from trying AI out: after an implementation, you make real decisions on its answers.
An implementation is not a purchase and not a pilot. It is a way of working that stays in place after the vendor has left. The four steps below build it, each leaning on the one before.
Step 1: a safe place where AI can reach your company data
AI answers stay generic as long as the AI does not know your company. The difference between a toy and a working instrument is access to your own numbers, contracts, minutes and quotes.
That makes the first decision a security decision. Company data does not belong in a public chatbot. Yet it probably already happens: at almost every company, employees use personal AI tools for work (see the source table below).
So the choice is not whether AI touches your data. The choice is whether that happens in a place you manage. That can be a business AI environment with the right permissions, or an assistant running on your own laptop.
Such a place, filled with your documents, clients and numbers, is an AI company brain. You ask it questions in plain language and get answers from your own sources.
Step 2: findable data, not perfect data
The classic advice: first a data warehouse, then AI. For the questions an executive team actually has, that is the wrong order. Modern AI reads your existing files: the spreadsheet, the PDF, the contract, the report.
"AI-ready" in your first year mostly means this: the documents that carry your decisions sit in one findable place. Imperfect is fine. Findable is the bar.
A large data project before first use is, for most companies, the most expensive way to never start. What is and is not required is in AI data readiness for executives.
Step 3: one decision-maker who uses it every day
Every procurement list skips this item, and it is the one that makes the difference. AI does not land through a rollout. It lands through the habit of one person with a real decision on their desk.
That person asks real questions every working day. What is moving in my margin? What do these hundred pages actually say? Which client is quietly ordering less?
A working implementation starts at the top, not with a working group. In a smaller company, that decision-maker is usually you. Rule of thumb: every working day, at least one question you would otherwise put to a colleague or adviser.
AI adoption works as rhythm and ownership, not as a project. The three habits behind it, a weekly rhythm, an owner per role and real company data, are in the SMB AI adoption playbook.
Step 4: only then an engineer for integrations
There comes a point where you genuinely need build capacity. For example when AI has to enter your order flow, touch your ERP or run part of a process on its own.
If you first work with AI on your own data, you know exactly what you are hiring for. The other way round is expensive: an engineer without a daily user is a gym membership without a gym.
These four roles matter most for a smaller company. The last column is about time and effort, not rates: those vary too much to quote without a source.
| Who | When you need them | What it costs in time |
|---|---|---|
| AI engineer: builds applications on existing AI models, such as integrations, workflows and agents | Once AI has to talk to your systems (ERP, CRM), not just your documents | A scoped job per integration; the size depends on your systems |
| Forward deployed engineer: builds the solution and sits at your table | Complex integrations into your core processes | Works alongside you for as long as the integration runs; scarce, so expect a wait |
| Data engineer: connects data sources and keeps them clean | When your data is scattered across many systems and joining them is unavoidable | Ongoing work: sources keep changing, so maintenance never stops |
| AI consultant: advises on strategy, policy and change | For buy-in, policy or a second opinion on a build plan | Advisory hours; usually builds nothing, so your own execution time comes on top |
Machine learning engineers, MLOps engineers and AI architects are rarely needed in a smaller company. Whether to hire a specialist at all, and what to do without a big-four budget, is in hiring an AI specialist.
Do it yourself, outsource or install?
For the first three steps there are three routes. They differ mainly in what you keep once the help has left.
Do it yourself. You pick a tool, prepare your documents and build the rhythm yourself. Little money up front, but evenings of figuring things out. The risk: tools, integrations and passwords pile up until you stop.
Outsource. A consultant or agency delivers a plan, buy-in or an integration. That fits step 4, once you know what needs building. For steps 1 to 3 you often buy a report instead of a habit.
Install. Someone sets up a working AI assistant on your own laptop, grounded in your company, and you work with it the same day. That is the AI Board setup day: have AI implemented in one day.
The facts: one day on site in the Netherlands. You sit in for about two hours yourself. You pay the Claude subscription to Anthropic directly. Updates and support afterwards are optional.
What that looks like for your company is something we discuss in a call.
Installing does not write custom software and does not replace an integration programme with your ERP. That still needs step 4. The trade-off per workflow is in outsource AI or do it yourself.
What does implementing AI cost?
The costs fall into four items. Only the last one is actual building work.
- Licences. A business subscription per user with the AI vendor. The price depends on the plan and the number of users.
- Your own time. Setting up, preparing documents and the daily rhythm. Put this item in your budget explicitly: it decides whether the rest pays off.
- Help to get started. A consultant by the hour, an agency per project, or a one-day installation such as the AI Board setup day. What that looks like for your company is something we discuss in a call.
- Integrations. Engineering work per system, worthwhile from step 4.
Agency rates vary widely, and an amount without a source does not help you. What each item costs, with sources, is worked out in what AI costs your business.
How long does it take?
The honest answer: installing is fast, learning to use it takes weeks. Implementation in one day means the installation is in place. The habit follows after.
- Day one: install. On the setup day the installation runs from 09:00 to 17:00. At the end the assistant runs on your laptop. Integrations that need third-party approval are finished remotely within five working days.
- Week one: fill it. Put the documents that carry your decisions in that one findable place. Numbers, contracts, reports and quotes first.
- Week two onward: the rhythm. One real question every working day. This is where the habit forms, and where you discover which integration is worth an engineer.
- ERP or CRM integrations: their own timeline. An engineer can only estimate those once they know your systems. A fixed deadline up front is a guess.
Implementing AI without data leaks
Two risks deserve your attention: company data that ends up in a public chatbot, and a business tool that sees more than it needs. Three measures help.
- One approved place for AI with company data, as in step 1.
- Permissions that are right: the AI only sees what the user is allowed to see.
- A short AI policy that says what is and is not allowed.
Ask every vendor where your files are stored and where your questions go. Why a local setup changes that answer is in private AI on your laptop.
With AI Board, your company brain and your files stay on your own laptop. If you choose the Claude edition, your questions go via Anthropic. The details are on the security page.
Why most implementations stall
Start with the sobering baseline. The numbers on AI implementation all point the same way.
| What goes wrong | The number | Source |
|---|---|---|
| GenAI pilots with no measurable P&L impact | 95% | MIT NANDA |
| GenAI projects predicted to be abandoned after proof of concept by end of 2025 | at least 30% | Gartner |
| Agentic AI projects predicted to be scrapped by end of 2027 | over 40% | Gartner |
| Companies that abandoned most AI initiatives in 2025 | 42% | S&P Global |
| Companies where employees use personal AI tools for work | over 90% | MIT NANDA |
The MIT figure of 95% is disputed, because the sample is small. Read it as a direction, not a precise measurement.
Notice what is not on that list: that the AI was not good enough. What fails is the last mile between an impressive demo and real work, with real data and a real decision.
The market sees it too. The Dutch daily FD reported in late August that Deloitte, Accenture, Capgemini and McKinsey are hunting AI specialists to cover that mile. Getting AI to work has become the scarce part, not the AI itself.
What the failures have in common: building without a daily user and without access to the real company data. The analysis per cause is in why AI projects fail.
Case study: Teus Louwerens
Teus Louwerens runs three companies alongside each other: Forever Idle, Projaxis and SleevesUp. Before every conversation he pulled the same context together: what was agreed, where things stood, which numbers belong to it.
In late August 2026 he set AI Board up on his own machine. Then he built on it himself: a meeting layer, a task layer, dashboards per company and his own boundaries for what the AI may do.
In his own words (translated from Dutch):
Preparing a meeting takes minutes instead of an hour, and I decide faster because the facts are already there.
His summary: "Working within a day, indispensable within a month." That is the order of this plan: first the place and the data, then daily use, and only then the extensions.
Teus wrote his story himself and was not paid for it. Read the full customer story of Teus Louwerens, including his four lessons.
Ready to start: checklist
Run through this list before you begin. Every line you can answer with yes saves you a detour later.
- You know which three decisions you want to prepare with AI first.
- There is one decision-maker who uses it every working day. In a smaller company, that is usually you.
- You have chosen a place where AI can safely reach your company data.
- The documents that carry your decisions sit in one folder.
- There is a short AI policy on what may and may not go into an AI tool.
- Your laptop is yours and you have admin rights on it.
- Your first real question is already in your calendar.
- You know which integration might need an engineer later, or you know it is too early to tell.
Frequently asked questions
How do you start implementing AI?
Start with a safe place where AI can reach your own documents, and one decision-maker who asks it a real question every working day. No pilot and no data project. Integrations come later, for needs you have proven by then.
What do you need to implement AI?
Four things, in this order. A safe place for your company data. Documents that are findable, not perfect. One decision-maker who uses it daily. And only then an engineer for integrations with systems such as your ERP or CRM.
Should you outsource AI implementation or do it yourself?
The first three steps are mostly a way of working that you carry yourself. Outsourcing pays off at step 4, when AI has to talk to your systems. In between: have it installed, then use it yourself.
Does implementing AI require a big data project?
No, not to begin with. Modern AI reads your existing files, such as spreadsheets, PDFs, contracts and reports. The bar for your first year is findable, not perfect. A data programme is only needed when joining many systems is unavoidable.
Do you need an AI specialist to implement AI?
Not for the first working use. An AI engineer becomes relevant once AI has to talk to your business systems. A machine learning engineer is mainly needed if you train your own models, which rarely applies to a smaller company.