AI 2 min read
AI automation: how to choose processes with real ROI
A framework to prioritize AI automations for SMBs: volume, variability, risk, integrations, data, and return measurement.
In this article +
AI automation works when it is applied to a process that already exists, hurts often, and has a verifiable output. It fails when the project starts with the model instead of the workflow.
The question is not “what can we do with AI”. The better question is: “which repetitive decision consumes time, creates errors, or blocks people who should be doing higher judgment work”.
Short answer
Prioritize processes with high volume, semi-stable rules, accessible data, and an output that is easy to validate. Avoid starting with rare processes, politically sensitive workflows, or processes with no clear owner.
Selection matrix
| Criterion | Positive signal | Risk signal |
|---|---|---|
| Volume | Happens daily or weekly | Happens a few times per year |
| Variability | There are exceptions, but repeated patterns | Every case requires new judgment |
| Data | Data is in email, CRM, ERP, PDFs, or APIs | Data lives in informal conversations |
| Validation | A person can approve or correct the output | Nobody can say whether it is right |
| Integration | There is an API, export, or structured access | Only a manual UI with no exportable data |
| Risk | A mistake is correctable | A mistake has high legal or financial impact |
Decision diagram
flowchart LR
A[Candidate process] --> B{High volume?}
B -- No --> X[Do not prioritize now]
B -- Yes --> C{Accessible data?}
C -- No --> D[Prepare data first]
C -- Yes --> E{Verifiable output?}
E -- No --> F[Define human criteria]
E -- Yes --> G[Two-week pilot]
G --> H[Measure savings and errors]Use cases with common return
| Process | Possible automation | Success metric |
|---|---|---|
| Incoming invoices | Extract supplier, amount, date, and concept | Minutes saved per invoice |
| Sales emails | Classify urgency and draft a response | Time to first response |
| Support tickets | Categorize, summarize, and route | Average assignment time |
| Weekly reporting | Consolidate sources and generate summary | Preparation hours |
| Internal onboarding | Create tasks, access, and reminders | Manual steps removed |
Minimum architecture
A serious pilot does not need a huge platform. It needs traceability.
flowchart TD
A[Input: email, PDF, CRM, or form] --> B[n8n workflow]
B --> C[Data validation]
C --> D[AI model or rules]
D --> E[Human review]
E --> F[Action in final system]
F --> G[Logs and metrics]How to measure whether it is worth it
| Metric | Simple formula | Why it matters |
|---|---|---|
| Time saved | Cases per month x minutes saved | Translates AI into operational capacity |
| Error avoided | Errors before - errors after | Justifies controls and validation |
| Cycle time | Case start to case closure | Measures real process speed |
| Human intervention | % of cases requiring correction | Shows whether the automation is mature |
Rules to avoid empty projects
- Do not automate a process nobody understands.
- Do not use AI where a deterministic rule is enough.
- Do not launch without logs, human review, and rollback.
- Do not measure “tokens used”; measure time, errors, and speed.
- Do not sell full autonomy when the risk requires human approval.
Applied AI done well is less spectacular than a demo, but much more profitable: it turns repetitive work into a supervised flow.
