When Is Vision AI Worth It?

Not every visual problem needs AI. Sometimes a good sensor suffices, sometimes the data foundation is too thin, and sometimes costs exceed benefits. This lesson gives you a decision matrix for informed build-vs.-buy decisions.

Decision Matrix

When Vision AI Makes Sense ✅

Criterion	Indication for Vision AI
Volume	> 1,000 visual inspections/day
Variability	Many product variants, no fixed template
Accuracy	Human inspectors achieve < 95% accuracy
Speed	Cycle times < 1 second required
Consistency	Subjective assessment by different inspectors problematic
Environment	Dangerous or inaccessible inspection environment
Data	Enough example images (min. 100 per class) available

When Vision AI Does NOT Make Sense ❌

Criterion	Indication against Vision AI
Volume	< 50 inspections/day (manual is cheaper)
Simplicity	Problem solvable with simple sensors (light barrier, scale)
Data	Too few examples, defects too rare or too variable
Cost	Error costs low (< €1,000/year)
Change	Inspection object changes weekly (constant retraining needed)
Physics	Defect not visually detectable (e.g., internal material defects without X-ray)

Build vs. Buy

Option 1: Cloud API (Buy)

Google Vision, AWS Rekognition, Azure CV — ready-made APIs for standard tasks.

Advantages: Quick start, no ML expertise needed, scales automatically Disadvantages: Data leaves the company, ongoing costs, limited customization Cost: €1–5/1,000 images Ideal for: Prototyping, standard OCR, generic classification

Option 2: Platform (Buy + Customize)

Roboflow, Landing AI, Clarifai — no-code/low-code CV platforms.

Advantages: Custom models without programming, visual labeling, one-click deploy Disadvantages: Vendor lock-in, monthly license costs, less control Cost: €500–5,000/month Ideal for: SMBs without data science team, fast iteration

Option 3: Custom (Build)

PyTorch + own team — full control, maximum customization.

Advantages: Full control over data and model, no vendor dependency, optimized performance Disadvantages: Requires ML engineers, longer development time, maintenance effort Cost: €50,000–300,000 initial + team Ideal for: Differentiation through CV, critical applications, high volume

Decision Helper

Is CV a core product or competitive advantage?
  → Yes: BUILD
  → No:
    Are standard APIs sufficient?
      → Yes: CLOUD API
      → No: PLATFORM

Cost Planning

Total Cost of Ownership (TCO) — 3-Year View

Cost Type	Cloud API	Platform	Custom
Setup	€2,000	€10,000	€100,000
Ongoing (year)	€12,000	€36,000	€80,000
3-year TCO	€38,000	€118,000	€340,000
Break-even vs. manual	Immediately	6 months	18 months

Hidden Costs

Often underestimated:

Labeling: 50–200 hours for initial dataset
Edge hardware: €5,000–30,000 per location
Integration: PLC connection, ERP connectors, alerting
Retraining: Models need updating when products change
Support: Who helps when AI is wrong? Who monitors?

Pilot Strategy

The 30-60-90 Day Plan

Day 1–30: Proof of Concept

Problem definition and data review
Collect and label 100+ example images
Quick prototype with cloud API or platform
Go/no-go decision based on accuracy

Day 31–60: Pilot

Refine model, collect more data
Integration into existing workflow (semi-automated)
Define and measure KPIs
Gather stakeholder feedback

Day 61–90: Evaluation

Calculate ROI based on pilot data
Create scaling plan
Make final build vs. buy decision
Build business case for rollout

Golden rule: Start with the most expensive visual problem in your company. The ROI must be obvious — then the budget follows automatically.