Last autumn, a Swiss medical device distributor called me. They had just completed a six-month AI project. The goal was to use machine learning to predict order volumes and automatically plan procurement. The results: the model ran, produced outputs, and nobody trusted them. Procurement still used their own spreadsheets. The project was quietly abandoned.
The technology worked. The vendor delivered what was promised. But the underlying order data was inconsistent — some records incomplete, some duplicated, some still in a legacy system that never got migrated. The AI was learning from noise. And when people can't understand why the AI recommends what it does, they stop using it.
This is not an isolated story. I've seen the same pattern repeat across manufacturing, distribution, and medtech in Switzerland. The failure mode is almost always the same: AI on top of unreliable data.
Why Most B2B AI Projects Fail
AI tools don't fail because they're bad. They fail because they amplify whatever is already in your data — including the problems. If your Salesforce and ERP have different customer IDs for the same account, an AI model will treat them as two customers. If order status fields are inconsistently filled by different sales reps, the model learns inconsistency.
There's a useful analogy: imagine asking a new analyst to forecast next quarter's revenue using a spreadsheet where 30% of rows have missing dates, product names are spelled differently across entries, and some deals are counted twice. No matter how skilled the analyst, the output will be unreliable. AI is the same — except it processes thousands of rows and looks authoritative while doing it.
The result is what I call confident garbage: predictions that look precise, update automatically, and are wrong in ways that aren't obvious until they cost you something real.
What Actually Works: Three Patterns I've Seen Succeed
Not all AI projects fail. The ones that succeed share a common characteristic: they start with process clarity, not with AI capability. Here are three patterns that consistently deliver results in Swiss B2B operations.
1. Automate repetitive decisions first, then make them intelligent
The most successful AI deployments I've seen didn't start as "AI projects." They started as automation projects. A distributor near Basel automated their order routing: orders above a certain value got flagged for review, orders below threshold went straight to fulfillment. No machine learning — just rules.
After six months of clean, consistent data flowing through that system, they had something valuable: a reliable training dataset. Then they introduced a model to predict which flagged orders were actually low-risk and could skip manual review. Adoption was immediate because the people using it already trusted the underlying process.
The principle: automate the routine before you optimize it with AI. Each automaton step cleans and standardises your data. By the time AI enters the picture, it has something solid to work with.
2. Use AI where humans are already the bottleneck
AI adds the most value where human capacity is genuinely limiting. A Swiss manufacturer I work with had a skilled operations manager who spent every Monday morning reviewing 40-60 purchase requests, checking each against stock levels, supplier lead times, and budget. She was good at it — but it took her half a day, and urgent requests had to wait.
We built a system that pre-processes each request: pulls current stock, checks supplier data, flags requests that meet standard criteria as auto-approved. She now spends 45 minutes reviewing the 8-10 exceptions that genuinely need judgment. The system didn't replace her — it removed the part of her job that was pure data lookup.
The result was immediate and trusted because the scope was narrow, the data sources were known, and she could verify every recommendation the system made before she chose to rely on it.
3. Make AI outputs explainable to the person acting on them
A key difference between AI tools that get adopted and those that don't: can the person using it understand why it's recommending something? "The model suggests re-ordering SKU 4471" is not enough. "Stock is at 12 units, average weekly usage is 18, and lead time from your supplier is 3 weeks" is actionable and verifiable.
In B2B operations, the people making decisions are experienced. They have judgment built over years. They will not — and should not — delegate critical decisions to a system they can't interrogate. The AI tools that stick are the ones that make their reasoning transparent enough that an experienced professional can agree, disagree, or override with confidence.
Before You Invest in AI: A Practical Diagnostic
If you're considering an AI project in your operations, run through these questions first. They take 30 minutes and will save you months of wasted budget.
- Can you export a clean, consistent dataset for the process you want to improve? If the answer involves significant manual cleanup, your data foundation needs work before AI does.
- Do the people who will use the AI output trust the current underlying data? If they already work around the data in Salesforce or their ERP, an AI layer won't fix that distrust.
- Is the current process documented well enough that a new employee could do it correctly? If not, you're not ready to automate it, let alone make it intelligent.
- What decision does the AI need to support — and who makes that decision today? The clearer and narrower the decision, the higher the chances of a useful AI application.
- What happens if the AI is wrong? If the cost of a wrong recommendation is high and hard to detect, start with AI as a second opinion rather than an autonomous actor.
The honest answer: Most Swiss B2B companies are not AI-ready yet — and that's fine. The companies that will benefit most from AI in the next 3 years are the ones investing now in clean processes, reliable integrations, and consistent data. That's the foundation. AI is the accelerator you add once the foundation holds.
Where to Start If You Want Real Results
Pick one process that is manual, repetitive, and involves data that already exists in your systems. Map every step. Identify where data moves between systems by hand — copy-paste, email, spreadsheet. Fix those handoffs first. Build the automation. Run it for three months.
At that point, you'll have clean, structured data for a real process. You'll know where the exceptions are. You'll know which decisions require human judgment and which don't. And you'll have built organisational trust in automated systems — which is, in practice, the hardest part of any AI project.
The companies I've seen succeed with AI didn't buy the most advanced tools. They did the unglamorous work first: clear processes, reliable data, honest integrations. Then they made those processes intelligent.