What AI process automation really is
A business process is a sequence of steps that goes from an input to a result: an order arrives, gets recorded, fulfilled, invoiced. Classic automation handles the rigid, predictable steps well (“if X arrives, do Y”). AI adds the ability to handle ambiguity: understanding the content of an email, extracting data from a PDF invoice, deciding which category a request belongs to.
Combining the two is what we call intelligent automation: deterministic rules where you need certainty and traceability, AI models where you need interpretation. It does not replace people: it removes the mechanical part and leaves them the decisions.
Where to start: choosing the right process
The first process to automate is not the most complex or the most “strategic”: it is the one with the best ratio of effort saved to risk. Three questions help pick it:
- Is it repetitive and high-volume? The more it repeats, the faster automation pays for itself.
- Are the rules stable enough? If they change every week, stabilise them first.
- Is an error recoverable? Start with processes where a mistake is visible and fixable, not irreversible ones.
The processes SMEs automate first
In practice, a few processes keep coming up as a good starting point:
- Order management: automatic recording, checking and routing, with fewer transcription errors.
- Invoicing and admin: data extraction from documents, reminders, reconciliations.
- Customer support: automatic answers to recurring questions, with escalation to a person when needed.
- Content and catalogues: generating and translating product sheets, SEO optimisation.
- Reporting: aggregating data from multiple sources into a readable dashboard.
Human-in-the-loop: automation that keeps control
Automation that acts on the real world (sends, pays, deletes) should not run free from day one. The model we apply is human-in-the-loop: at first every sensitive action goes through a human confirmation. You observe real behaviour, measure it, fix the edge cases. Only once the system is reliable does autonomy grow, always keeping logs and the ability to step in.
It is the difference between automation you trust and a black box that one day causes silent damage.
Security and compliance: not a detail
When an automated process touches customer data or uses AI models, security is part of the project, not a final add-on. We apply recognised frameworks: OWASP for the web and APIs, the OWASP Top 10 for LLM Applications and MITRE ATLAS for AI-system risks, starting with prompt injection. We also account for the EU AI Act’s obligations. One fact: designing security in from the start costs less than chasing it after an incident.
What it costs and when it’s worth it
There is no single price: it depends on the process, the systems to integrate and the level of autonomy required. The practical rule is to start from a small, measurable perimeter (one process, one clear result), verify the real return, then extend. Custom automation makes sense when the time saved (or the errors avoided) stably exceeds the cost of building and maintaining it.