AI & automation 
Where should you start with AI automation?
For a Belgian SME, a useful first AI pilot focuses on a frequent task, accessible data and an easily checked result: classifying a request, extracting a document or drafting a response. Measure the current process first, then test one workflow with human review. Rules-based automation is sufficient when inputs and decisions are structured; AI is worth evaluating when text, documents or wording vary.
Three checks before choosing a tool
- 01Start with a process, a business owner and a measured baseline.
- 02Test errors, exceptions and review effort alongside successful examples.
- 03Connect a pilot to existing tools after establishing its usefulness.
01
AI, automation or an agent: what does your company need?
Process automation connects steps and software. For example, a validated form creates a CRM contact and a sales task. Explicit rules and structured inputs may require no AI model at all.
AI can help interpret a freely written request, classify a document or prepare a summary. Its output needs evaluation: plausible answers can be wrong. An AI agent can use several tools to pursue a goal, making clear permissions and an action scope particularly important.
| Need | Approach to consider | What to check |
|---|---|---|
| Move a form submission into a CRM | Rules-based automation | Fields, duplicates and recovery |
| Interpret variable messages or documents | An AI step in a workflow | Accuracy on representative examples |
| Find internal information | Search and an assistant with sources | Authorised documents and inspectable references |
| Take actions across several tools | An agent with limited actions | Permissions, approval and action logs |
02
10 tasks a Belgian SME can examine for a first AI pilot
These ideas do not assume that AI is profitable in your company. Compare the current workflow, a rules-based solution and AI assistance. Low volume or difficult data may make a different improvement more useful.
If your organisation works in French, Dutch and English, evaluate each language on its own examples. A French demonstration does not establish Dutch extraction, classification or response quality.
| Task | Input → proposed output | Control and indicator |
|---|---|---|
| 1. Qualify incoming enquiries | Free-text email → category, need and draft CRM entry | Confirm ambiguous cases; measure routing errors |
| 2. Draft support responses | Question + documentation → draft with references | Review before sending; measure total time and corrections |
| 3. Extract document information | Order or form → structured fields | Check amounts and references; measure field accuracy |
| 4. Find internal information | Question → answer linked to an authorised document | Check sources and access; measure useful answers |
| 5. Summarise meetings | Notes or authorised transcript → proposed decisions and tasks | Validate owners and deadlines; measure review effort |
| 6. Prepare quotations | Brief + approved catalogue → draft and questions | Human approval of prices and commitments; track rework |
| 7. Prepare sales follow-ups | CRM history → proposed message and next step | Review before sending; observe replies without assuming causation |
| 8. Classify case documents | Incoming files → proposed type, folder and missing items | Confirm filing; measure classification errors |
| 9. Draft reporting commentary | Calculated indicators → commentary and questions | Trace figures to sources; check invented explanations |
| 10. Draft content translations | Source text → working French, Dutch or English version | Language and business review; measure corrections per language |
03
Prioritise a use case beyond the demonstration
Choose a task someone can explain, including its exceptions and expected outcome. Gather common and difficult examples, keeping some aside for evaluation. Set success criteria before tuning the solution to the sample.
Candidates worth examining combine adequate frequency, repetitive input or search and straightforward human review. Low volume, unavailable data or difficult-to-reverse consequences suggest starting elsewhere. Keep a simpler deterministic approach in the comparison.
Five questions for the decision
- 01
Frequency
How often was the task actually performed last month?
- 02
Current effort
How much time does it take, including rework and checks?
- 03
Available data
Can examples, documents and access be obtained legitimately?
- 04
Acceptable error
Who detects a wrong answer and restores the usual process?
- 05
Ownership
Who approves criteria, monitors the pilot and decides what follows?
04
Data, languages and integrations: prepare a pilot
Map the tools actually used: email, CRM, ERP, document storage and forms. Check access, formats, duplicates and permissions. A demonstration on an exported file does not establish that everyday integration will work with your software.
Select the information needed and who may access it. The European Commission explains that pseudonymised data allowing reidentification remains personal data. Check providers, retention and access before sharing real records; precise obligations depend on the processing.
Document company terminology and languages. Sales categories, abbreviations and product names need shared definitions. Provide an explicit fallback when documents are missing, sources conflict or a request falls outside scope.
Connect AI to your organisation
Sources consulted
05
Keep control of actions and errors
Start in proposal mode: the assistant classifies, extracts or drafts, and a person confirms. Limit access, keep usable logs and plan how to disable a step without stopping the business. A confidence statement generated by the model is not evidence of accuracy.
Approval can precede sending a message, updating a record or making a commercial commitment. n8n documents a review step that pauses tool execution until a person decides. Test the chosen mechanism within the complete workflow.
Also test unreadable documents, malicious requests, missing information and unavailable services. An incoming document is material to analyse, not authority to change rules or access unrelated records.
Control an assistant using business tools
Sources consulted
06
Measure actual value with a before-and-after calculation
Response generation time does not measure process improvement. Include data preparation, review, rework, exceptions and supervision. Track quality and tool costs alongside time. Available capacity is not automatically a cash saving.
This fictional calculation assumes 300 comparable cases each month. Before: 12 minutes per case. After: 8 minutes including individual review and corrections, plus 5 hours of monthly supervision. None of these figures is a Gaspa client result.
| Situation | Calculation | Total time |
|---|---|---|
| Before | 300 × 12 minutes | 60 hours |
| After, processing and individual review | 300 × 8 minutes | 40 hours |
| Additional monthly supervision | 5 hours | 5 hours |
| After, total | 40 + 5 hours | 45 hours |
| Theoretical capacity released | 60 − 45 hours | 15 hours per month |
07
Budget and deployment: advance through explicit decisions
Separate understanding the task, testing examples, integration and everyday operation. Costs depend on data preparation, connected software, evaluation and monitoring. An individual subscription does not necessarily cover a shared process with permissions and integrations.
At each stage, decide to continue, simplify or stop based on results. Before launch, identify ownership, alerts, support and rollback to the earlier process. Connecting a pilot to a CRM does not by itself establish reliability.
08
What Gaspa’s references allow you to verify
Alba Golf shows Gaspa’s work on a mobile product with AI-assisted video analysis, coach and player spaces, APIs and a back office. It illustrates delivered scope, rather than demonstrating administrative productivity gains for SMEs.
Match references to the process and constraints you need to address. Prepare examples, state what must remain subject to human approval and choose an observable outcome. This creates a concrete starting point for the discussion.