Change the work before the tool. Map the one process you want AI in, bring its data into one place, name an owner and give them time to learn, write a one-page usage policy, pick one metric and measure it before and after, and train the team that does the work. Leave the rest of the company alone until that process shows a result.
When factories got the electric motor, many did the simplest thing: they took out the steam engine and put an electric motor in its place. Everything else stayed as it was. One long shaft running through the whole building, every machine belted to it. The productivity gains took decades to show up.
They came when factories were rebuilt. Electric motors could be small and sit right next to each machine, so the floor could be laid out around the flow of work rather than around the central shaft. The economist Paul David described that lag in a 1990 paper, "The Dynamo and the Computer".
I made the same point on the REWIRED podcast: "You cannot implement AI without rewiring your own processes." Most companies today are doing what those factories did. They buy a subscription, lay it over yesterday's work and wonder why nothing moves. I went through the seven usual reasons in why AI is not delivering results.
So what do you change, concretely? Below is the list I would put on the table on a Monday morning.
Why isn't buying an AI tool enough?
Because the tool works with what you give it. If the process is chaotic and the data is scattered, AI produces chaotic results, only faster.
On the podcast I put it this way: "Trash in, trash out. It's a maximizer." AI amplifies what it finds. A clear process gets faster. A confused one gets confused at scale.
What should I change first? The Monday list
- Map the process before you choose the tool. Take a single process, say quoting or supplier invoice intake. Write the steps on one sheet: who does what, with which data, where it gets stuck, what exceptions come up. If two colleagues describe it differently, you have found your first problem, and it is not an AI problem.
- Bring the process data into one place. AI works well when all the information sits in one source. If part of it lives in the ERP, part in someone's spreadsheet and part in a phone call to the warehouse, no model can join it up. The Romanian specifics are in the next section.
- Name an owner and give them time to learn. Someone has to answer for the process once AI is in it. Put a few hours a week in their calendar for testing, not "when they have time". I wrote separately about who that person should be in how to structure your company for AI.
- Write a one-page usage policy. Which tools are approved, on which accounts (company, not personal), what data never goes into them, and who checks an output before it reaches a client. If the tool itself is still undecided, see ChatGPT vs Copilot for business.
- Pick one metric and measure it first. Time per quote, invoices processed per day, errors per month. Write down today's value. Without it, three months from now you will only have impressions.
- Train the team, not only the enthusiasts. The people who do the work need to know what AI can do, where it goes wrong and what they may give it. Beyond common sense, Article 4 of the EU AI Act has required since 2 February 2025 that companies using AI systems take measures on their staff's AI literacy; what that means in practice is in AI Act Article 4: what SMEs need to do.
- Decide what you will not change. This matters as much as the other six. See below.
How do I get my business data ready for AI?
Start from the data you already hold in a structured format. You may have more of it than you think, and in Romania the tax system has already done part of the work.
- E-invoices. In Romania, since 2024, B2B invoices between companies established in the country go through the RO e-Factura system as XML (Ministry of Finance guide, read 26 September 2026). Your incoming and outgoing invoices already exist as structured data. For any process around purchasing or suppliers, that is the cleanest place to start.
- Your accounting software and ERP. Whatever you run (SAGA, SmartBill or a full ERP in Romania; something else elsewhere) can export data. Ask your accountant or vendor which exports you have and how often you can run them. A regular export beats a spreadsheet rebuilt by hand every month.
- Spreadsheets built for display. This is the tricky part. The nicely formatted table with merged cells and colours reads well for a person, but a program cannot read it. Keep a plain version: one row per record, one column per field, no merged cells.
- Knowledge in people's heads. If a step in the process means, say, "I ring Maria in the warehouse", that step cannot be automated until the information is written down somewhere.
If the process needs a technical integration (an assistant over your internal documents, for instance), Sapio has a separate guide on getting your company data ready for AI.
Where should I not change anything?
Anywhere you lack a clear process, an owner and a metric. And in a few places where change costs more than it returns.
- The org chart. Do not reorganise the company for AI before you have a result. Change a process, not a structure.
- Rare processes. What happens twice a month is not worth automating. What happens hundreds of times a day is. I wrote more on this in what is actually worth automating.
- Decisions about people. Hiring, appraisals, dismissals. AI can help prepare them, but the decision stays with a person, and the AI Act treats many of these uses as high-risk.
- Relationships built on trust. The long-standing client who wants to speak to a particular person does not need a chatbot.
- Licences for everyone. Do not buy seats for every employee on day one. Start with the team that runs the chosen process.
This article is not legal advice. If you use AI in decisions about people, check your obligations with a lawyer.
How long before I see a result?
For one well-chosen process, think in quarters, not weeks. For the whole company, longer.
The good news is that a company does not have to rebuild an entire factory floor. You can rewire a single flow, quoting for instance, and measure the difference. A smaller company has a real edge here: a process is easier to change at 60 people than at 6,000, where every step has several stakeholders who would rather nobody touched it. In Romania the gap is especially visible. In 2025, only 5.2% of Romanian enterprises with at least 10 employees used AI, against 20% across the EU (Eurostat, 11 December 2025). Whoever puts their processes in order now starts ahead of most.
Checklist: is your business ready for its first AI project?
| Question | If the answer is no |
|---|---|
| Have we chosen one process and mapped it step by step? | Pick the process that costs you most every week and write its steps on one sheet. |
| Is its data in one place, in a format a program can read? | Start from e-invoices and accounting exports; flatten display spreadsheets. |
| Is there an owner, with time in the calendar? | Name one person and block a few hours a week. |
| Do we have a one-page AI usage policy? | Write it: approved tools, company accounts, banned data, who checks outputs. |
| Have we recorded today's value of one metric? | Measure it for a week before anything changes. |
| Has the team that runs the process been trained? | Train on their own tasks, not a generic course. |
| Do we know what we will not change? | List it, starting with the org chart and rare processes. |
If you have more than two "no" answers, do not choose the tool yet. Fix those first. It goes faster than it looks.
I often think about the engineers who stood for decades next to the central shaft, with a brand-new electric motor, wondering why nothing had changed. The motor was never the problem. The whole factory had been designed around a single shaft, and nobody had asked whether it still needed one. What is the central shaft in your company?
If you want to work through this list for one process in your company, the details are on my AI consulting page; for your team, see my AI workshops.