How to Find the Right AI Use Case

Most AI projects in SMBs don't fail because of technology. They fail because someone picked the wrong problem to solve. Here are the 5 criteria we run through at Cogswell IT before any project – no buzzword bingo, with real examples from auto shops, medical practices and restaurants.

You're sitting in a webinar, someone shows you a demo, and suddenly you think: "We need that too." Three months later you've spent 18,000 euros on a tool no one uses. Sound familiar? You're not alone.

Here's the truth: artificial intelligence works in SMBs – but only where the problem fits. Not every process is an AI use case. And not every impressive demo survives the reality check in a small auto shop in Bensheim or a doctor's office in Heppenheim.

So you don't have to guess, let's walk through the 5 questions you should ask yourself before spending the first euro. If all five turn green, the conversation is worth having. If three are red: walk away.

1. Does the process repeat often enough?

AI pays off through scale. A task you do three times a year doesn't get automated – you just do it. A task you do three times an hour is a candidate. Rule of thumb: we're talking at least 100 repetitions per month, usually more. Below that, the effort for setup, testing and maintenance is almost always higher than the savings.

Important: "repeating" doesn't mean "identical". Every call into your practice is different – but the structure is the same (greeting, request, time suggestion, confirmation). That's exactly what makes it automatable.

Example: A dental practice in Bensheim takes 60 to 80 calls per day. 70% of them are appointment requests, reschedulings or prescription refills. That's a clear use case for a voice agent – not because it's technically spectacular, but because the pattern repeats thousands of times per year. A tax advisor with three complex clients per month? Entirely different story.

2. Is there enough data – and at what quality?

"We have so much data" is what we often hear. When we take a look: Excel files with three different column names for the same field, scanned PDFs without OCR, five customer management systems running in parallel. That's normal – but it means: before AI can do anything useful, we need to clean up.

What you need depends on the use case:

• Generative AI (text, replies, summaries): a few example documents plus public models often work.
• Classification (invoices, sorting emails): minimum 200 to 500 clean examples per category.
• Forecasting (demand, maintenance, churn): 12 to 24 months of historical data – consistently captured.

Example: An auto shop wanted to predict which customers were likely to come in for inspection within the next 60 days, so they could proactively reach out. Sounds great – but the shop management software didn't have consistent vehicle data before 2022. Result: we delayed the project by six months while clean data was being captured. Had we started earlier, the model would have been useless.

3. Is the error tolerance acceptable?

No AI system is 100% correct. Period. The question isn't whether errors happen, but what happens when. Some processes can handle a 5% error rate, others can't tolerate 0.1%. Before you start, you need to answer this honestly.

Three simple questions:

• If the AI decides incorrectly – what does it cost? (Time? Money? Lost customer? Liability?)
• Is the error immediately visible, or does it surface only weeks later?
• Is there a human in the loop who can correct it before damage is done?

Example: An Italian restaurant in the region uses a voicebot for table reservations. If the bot occasionally mishears a name ("Miller" instead of "Müller"), that's a slight eye-roll at the host stand – no drama. A pharmacy, on the other hand, that wanted to use AI for dispensing medication? We turned that down. Wrong dosage isn't an eye-roll – it's a career-ending event.

4. Does the ROI actually add up – honestly calculated?

This is where the biggest lies happen. AI projects are often sold with the savings potential of a full-time employee who gets "completely replaced". In reality, AI rarely replaces entire roles – it takes over parts of tasks. That's still valuable, but the math looks different.

Honest calculation:

1. How many hours per month does your team spend on the task?
2. What share of that is actually automatable (more like 50–80% than 100%)?
3. What does an hour cost including overhead?
4. What setup and operating costs hit you over 36 months?

Example: A trades business with 15 employees wanted AI for quote generation. Status quo: 8 hours per week, one employee. Realistic savings with automation: 5 hours. At 35 euros per hour, that's roughly 9,100 euros per year. One-time setup 6,000 euros, operation 80 euros per month. Break-even after month 9, clear profit after that. That's a good case. Had we promised 80% savings, the customer would have been disappointed after three months – and we would have torched our reputation.

5. Who's accountable when the AI gets it wrong?

This is the question left out of 90% of sales presentations – and the question that comes up in 100% of projects sooner or later. If the chatbot gives a customer wrong information, who's liable? The vendor? You? The employee who didn't double-check?

Three things need to be settled before project start:

• Who's legally liable? – Check the contract clauses with your AI provider. Does your vendor demand a risk exclusion for "all AI outputs"? Be careful.
• Who controls internally? – Do you need a human to sign off on critical decisions? For diagnoses, legal advice, financial recommendations: yes, always.
• How are errors documented? – Logging, feedback loop, regular review. Otherwise the system learns nothing, and neither do you.

Example: A medical practice software uses AI to pre-prioritize patient requests. We agreed: anything classified as "urgent" still routes through a human before escalation. Anything classified as "standard" runs automatically – but with weekly random sampling. Responsibility stays with the physician, the tool supports. That's how AI works in regulated environments – not differently.

What do we do with this list?

At Cogswell IT, we run through it with every customer – before we write a quote. Sometimes after 30 minutes we say: "Don't do it, the math doesn't work." We lose revenue on that. We earn it back because customers return when they genuinely have a good use case.

If you're thinking about an AI project right now and unsure whether it's the right problem to tackle: drop us a line. We listen, ask questions, and tell you honestly if it fits. More on our approach is on our AI consulting page.