Start Small, Learn Fast: How to Explore AI Without a Strategy
Many Swiss companies are in the same place right now. Leadership agrees that AI matters, teams are curious, a few employees already use chatbots on the side, but there’s no AI strategy yet. The instinct is to fix that first: hire consultants, write a strategy paper, run a platform tender, and only then start building.
In my experience, that order is backwards. You can’t write a good AI strategy before you know which use cases actually work in your company, with your data, your people and your compliance requirements. The better path is to experiment first, in a controlled way, and let the results shape the strategy.
Why strategy-first often fails
An AI strategy written without hands-on experience rests on assumptions: which use cases will pay off, which models are good enough, what infrastructure you need. By the time the strategy is approved and a platform chosen, the technology has moved on and the assumptions are outdated.
There’s a subtler problem too. Big upfront platform decisions lock you in before you understand your requirements. Do you need frontier models, or are open models good enough? Does your data have to stay in Switzerland, or is an EU region acceptable? Will usage be steady enough to justify your own hardware? You only find out by trying.
Step 1: Set minimal guardrails, in a week, not months
Experimenting doesn’t mean a free-for-all. Before anyone builds anything, agree on a few basics:
- Which data may be used. Start with public and internal data. Keep client data and personal data out of the first experiments.
- Who approves experiments. One accountable person or a small group is enough.
- A budget cap. A monthly limit keeps experiments from turning into surprise invoices.
A one-page AI usage policy covers this. It’s also a good moment to address "shadow AI": employees are probably already pasting company content into consumer chatbots. Giving them a sanctioned alternative is safer than pretending it isn’t happening.
Step 2: Pick the lowest-friction platform
At this stage the right platform is the one that’s easiest to start with, not the one that looks best on paper.
- If you already use AWS or Google Cloud, start there. Amazon Bedrock and Google Vertex AI give you access to leading models, including Claude, in EU regions under contracts and security reviews your company already has.
- If sovereignty is the top concern, or you have no hyperscaler relationship, Swiss providers like Infomaniak offer pay-as-you-go APIs for open models, hosted entirely in Switzerland, with free test credits and spending limits.
- If most users aren’t technical, a business chat product often shows value faster than building anything custom.
Keep this decision reversible. Most modern LLM APIs are similar enough that switching providers later is realistic, as long as you don’t build deep dependencies on one vendor’s proprietary features in the first month.
Step 3: Collect use cases from the business, not from IT
The best use cases rarely come from the IT department. They come from the people doing the work. Run a workshop with different teams and ask a simple question: where do you spend time on repetitive, text-heavy work?
Typical answers:
- Summarizing long documents, reports or contracts
- Drafting standard replies to customers or suppliers
- Finding information buried in internal wikis, manuals and SharePoint folders
- Extracting data from forms, invoices or emails
- Classifying and routing tickets or requests
You’ll usually end up with a long list. That’s good. The next step is deciding what to test first.
Step 4: Score and pick two or three
Rate each use case on four simple criteria:
- Business value. How much time or money would it save if it worked?
- Data sensitivity. Can you test it without touching personal or confidential data?
- Feasibility. Is the data available and in usable shape?
- Measurability. Can you tell clearly whether it worked?
Pick two or three that score well on all four. A classic first choice is an internal knowledge assistant that answers employee questions from company documents. It’s valuable, low-risk because it’s internal, and easy to demonstrate.
Resist the temptation to start with the most ambitious idea. Early wins build the internal support you’ll need for the harder projects later.
Step 5: Run small pilots with clear success metrics
Give each pilot a fixed time box, typically four to six weeks, and define upfront what success looks like. Hours saved per week? Answer accuracy as rated by users? Adoption: do people keep using it after the first week?
Keep a human in the loop. In early pilots, AI output should support people, not replace their judgment. That reduces risk and gives you honest feedback on quality.
This is also where you learn the things no strategy paper can tell you: whether your documents are in good enough shape for AI to use, where the model struggles with Swiss German or domain-specific terms, and how your compliance team reacts to real, concrete cases instead of abstract ones.
Step 6: Now write the strategy
After a few pilots, you’re in a completely different position. You know which use cases deliver value and which don’t. You’ve discovered your real data quality and compliance issues. You have a sense of your likely usage volumes, which is the key input for the infrastructure question: stay on a hyperscaler, move to a Swiss provider for sovereignty, or eventually run your own hardware for steady, high-volume workloads.
That’s the evidence a real AI strategy should rest on. It will be shorter, more concrete, and far more likely to survive contact with reality than one written in a vacuum.
Two mistakes to avoid
Starting with a big platform tender. Choosing an enterprise AI platform before you know your use cases means buying capabilities you may never need and missing ones you do.
Running pilots without success criteria. Without a defined metric, every pilot ends in "it looked promising." That’s how companies end up with a dozen demos and nothing in production.
The bottom line
You don’t need an AI strategy to start with AI. You need clear guardrails, a low-friction platform, a few well-chosen use cases and honest measurement. Start small, learn fast, and let what you learn shape where you go next.
If your company is at this stage and wants help picking the right first use cases or setting up a pilot, feel free to get in touch.
Tags: AI Strategy, Enterprise AI, Switzerland, LLM
