Most AI pilots in small and mid-sized companies don't fail because the model is weak. They fail because the foundations underneath are crooked. Before you buy a license, there are three things worth sorting out.
We've seen the same pattern over and over in the last year. A company gets an offer for an "AI customer service assistant", someone runs a demo on a handpicked slice of records, everything looks great. Then deployment day arrives and it turns out the "customer data" is actually three Excel files, with the same customer appearing under four different names and typos in the tax IDs. The model never stood a chance. Not because it was bad, but because nobody knew who was who.
AI isn't magic that fixes mess. AI is an amplifier. It amplifies whatever you already have, including the chaos.
Foundation 1 — Clean data
"Clean data" in SME reality doesn't mean a half-million data warehouse. It means something much simpler: one source of truth per thing. One customer record, one product record, one invoice record. Not five versions of the same file on five drives.
In practice that's consistent formats (dates always in the same format, tax IDs always without spaces), no ghost Excel files circulating by email between departments, and clear rules on who is allowed to change what. It's boring, clerical work. But without it, AI has nothing to start from.
- Identify where your customer, product and order data actually lives.
- Remove duplicates and agree which record is "the real one".
- Set common formats for the key fields (dates, tax IDs, product codes).
- Assign one person to own data quality — even if it's two hours a week.
Foundation 2 — Integrated systems
A typical manufacturing SME has an ERP (often Comarch Optima or Microsoft Dynamics AX), a separate warehouse system, a CRM bought once for a specific project, and a handful of Google Sheets where someone keeps "the thing that had to be ready yesterday". Each system knows a fragment of reality. None of them knows the whole picture.
Until these systems talk to each other, no AI will give you a sensible answer to "how much do we actually earn on customer X". The answer is scattered across four databases that don't know the others exist. Integration — usually through APIs — is the prerequisite for having anything to analyse at all.
- Map it out: which systems you have and what each one really holds.
- Check which expose an API and which need a workaround.
- Decide what is the "source of truth" for customer, product, order.
- Integrate first the two systems between which data is already being copied by hand.
Foundation 3 — Smooth processes
AI amplifies whatever process it touches. If the process works well, AI will speed it up tenfold. If the process is broken, AI will break it ten times faster. That's the whole secret.
A real example. A wholesaler rolled out automatic categorisation of incoming quote requests. The problem: until then, those requests were "categorised" by a sales rep who, along the way, silently filtered out 30% of questions about products the company didn't even stock. The AI didn't do that — it politely categorised everything and dropped it into the queue. The team drowned. The process had been broken for years, and a human was patching it quietly. AI just showed where the hole was.
- Write out the process you want AI to help with — step by step, on paper.
- Mark the steps where "something happens by magic" (meaning someone silently saves the day).
- Fix those steps before you put a model anywhere near them.
- Measure the process before you deploy — otherwise you'll never prove AI did anything.
What next
All of this is boring. Cleaning data, integrations, writing down processes — these are not the things you brag about at a conference. But this is exactly what separates a company where AI actually works from a company where AI is a slide in the board deck.
If you don't know which foundation to start with in your business, instead of guessing, talk to someone who breaks these things down for a living. The first call is free and takes half an hour.
AI won't fix your company. But if the company runs well, AI can make it run a lot better.