"We need AI" has become a reflexive response to business problems. Sometimes that's right. Often it isn't.
After working on AI projects across utilities, e-commerce, and enterprise software, we've developed a framework for answering this question honestly — before spending money on a solution that might not fit the problem.
The First Question: Is It Actually a Prediction Problem?
Machine learning is fundamentally a tool for making predictions under uncertainty. Before reaching for AI, ask: is the core of my problem a prediction?
Examples of prediction problems AI handles well:
- "Given this image, what meter value does it show?" (perception/classification)
- "Given this customer's history, will they churn next month?" (forecasting)
- "Given this text, is it a complaint or a compliment?" (classification)
- "Given these sensor readings, is this machine about to fail?" (anomaly detection)
Examples that look like AI problems but aren't:
- "Our invoicing process is slow" — this is a process problem. Automation (scripts, RPA) usually solves it better than ML.
- "We don't know our inventory levels" — this is a data integration problem. A proper database and reporting system comes first.
- "Customer service takes too long" — often a staffing/process problem. AI can help at the margin, but not if the underlying workflow is broken.
If you can write a clear rule for the decision, you probably don't need ML. Rules are faster to implement, cheaper to maintain, and easier to debug.
The Data Question
AI requires data. Not just "some data" — the right kind of data, in sufficient quantity, with reasonable quality.
For supervised learning (the most common type), you need labeled examples: inputs paired with the correct output. Ask yourself:
Do you have historical data where the outcome is known?
- For churn prediction: past customers + whether they churned
- For defect detection: past images + labels of which ones had defects
- For demand forecasting: historical sales + the actual demand that materialized
How much data? Rules of thumb:
- Simple classification (2-3 classes): 1,000+ examples per class minimum
- Complex detection (many object types): 10,000+ labeled examples
- Time series forecasting: 2-3 years of historical data for seasonal patterns
Data quality matters more than quantity. 5,000 well-labeled examples beat 50,000 noisy ones. Budget for data cleaning and labeling — it's usually 50-70% of the total project effort.
The ROI Question
AI projects have high upfront costs and lower marginal costs at scale. This means they make sense when:
- The task scales — you're doing the same prediction thousands or millions of times
- The current solution is expensive — manual labor, high error rates, or slow decision cycles
- Accuracy matters — decisions with real consequences (not just "nice to know")
A framework for rough ROI estimation:
Annual value = (Current cost per decision) × (Annual volume) × (Error rate reduction)
+ (Speed improvement value) × (Annual volume)
Rough project cost = €30K–€150K for a custom model (data, development, deployment)
+ €5K–€20K/year maintenance
If the annual value doesn't exceed project cost within 2-3 years, the business case is weak. Be honest about this before starting.
When to Start Simpler
Some problems are better solved with non-ML approaches:
Rule-based systems — when decisions follow clear logic that experts can articulate. Faster to build, transparent, easy to audit. Use ML when the rules are too complex or numerous to maintain manually.
Traditional statistics — for forecasting with small datasets. A well-tuned ARIMA model often beats a neural network when you have less than 2 years of monthly data.
Classical algorithms — for optimization, routing, scheduling. These are well-understood, computationally cheap, and explainable. Don't use ML for problems that are fundamentally combinatorial optimization.
Better UX/process design — sometimes the "AI" solution is just making it easier for humans to do their job well.
The Maturity Check
Before investing in AI, your organization needs certain foundations:
- Data infrastructure: can you reliably collect, store, and access the data needed? If you're manually exporting Excel files to feed into a model, that won't scale.
- Technical capacity: do you have (or can you hire/contract) people who can build, deploy, and maintain ML systems? A model that nobody can maintain is a liability.
- Process readiness: is there a clear owner for the AI system's outputs? Who acts on the predictions? What happens when the model is wrong?
Organizations that skip these foundations often end up with impressive demos that never make it to production.
A Practical Decision Framework
Use AI when:
- ✅ The problem is genuinely a prediction problem
- ✅ You have (or can collect) sufficient labeled data
- ✅ The ROI calculation is positive within 2-3 years
- ✅ The task scales to justify development costs
- ✅ Your organization has the infrastructure to use the outputs
Be skeptical when:
- ❌ The problem could be solved with clear rules
- ❌ You don't have labeled historical data
- ❌ The task volume is low (< 1,000 decisions/month)
- ❌ You don't have someone to maintain the system
What We Tell Clients
When a company comes to us saying "we want AI," our first response is always a set of questions: What specific decision do you want to automate or improve? What data do you have? What does success look like, and how will you measure it?
The answers determine whether AI is the right tool. Sometimes it is — and we build something. Sometimes it isn't — and we say so, because a failed AI project wastes more than just budget.
Thinking about an AI investment? Talk to us — we'll tell you honestly whether it makes sense for your situation.