Français
Deutsch
RPA's promise and its structural limits
RPA transformed process automation from 2015 onwards. Tools like UiPath and Power Automate allowed companies to automate repetitive tasks without modifying existing application source code. But ten years on, Swiss CIOs are reaching a bitter conclusion: RPA bots are expensive to maintain, break at the slightest UI change, and can only handle 100% structured processes.
What RPA fundamentally cannot do
- Natural language understanding: impossible to handle "cancel if the customer is VIP and this is their first complaint"
- Exception handling: every variation requires a pre-coded rule
- Unstructured documents: emails, scanned PDFs, images
- Contextual reasoning: unable to arbitrate between two contradictory rules
The AI agent: a fundamentally different paradigm
"RPA asks: 'What is the exact path to the result?' The AI agent asks: 'What is the desired result?'"
| Criterion | Classic RPA | AI Agent |
|---|---|---|
| Exception handling | Pre-defined rules only | Contextual reasoning |
| Unstructured documents | ❌ No | ✅ Emails, PDFs, images |
| Maintenance when UI changes | Full redesign required | Automatic adaptation |
| 3-year maintenance cost | 35–50% of budget | <10% of budget |
Concrete example: supplier invoice processing
RPA approach (2018)
The bot extracts PDF fields at fixed coordinates, checks the purchase order in SAP, validates the amount against a rule table. It works — until the supplier changes their template. Once they do, errors hit 100% of invoices.
AI Agent approach (2025)
The agent receives the PDF by email, understands the content regardless of layout, queries the ERP via API, identifies any discrepancies, drafts a clarification email if needed. If the format changes, the agent adapts without any human intervention.
Transition strategy for CIOs
Phase 1: map existing RPA bots, identify the most expensive to maintain. Phase 2: replace the 3–5 most problematic bots with AI agents, measure the gains. Phase 3: define a clear policy on what stays in RPA and what migrates to agentic.