Data and artificial intelligence
Process automation with AI agents
I automate repetitive processing by combining deterministic rules with language models and explicit guardrails. Automation handles the simple cases and hands ambiguous ones to a human.
What it covers
An AI agent does not replace your business rules. It is a component that reads unstructured information — emails, attachments, poorly filled forms — and turns it into data your existing processes can use. Anything an explicit rule can decide should stay a rule.
The scope is therefore split into three categories: cases handled automatically, cases handled then checked by sampling, and cases always routed to an operator. That split is agreed with the business teams before any build starts.
Guardrails and supervision
Every agent action is limited to an explicit tool perimeter, logged, and reversible where possible. The tracked indicators are effective automation rate, error rate found through checks, and time saved per processed case.
- Cases split into automatic, checked and manual.
- Tools exposed to the agent limited and validated server-side.
- Actions logged with input, output and model decision.
- Sampling checks on sensitive cases.
- Retry queue for cases not handled automatically.
Problems addressed
Teams re-enter data already present in emails or attachments. Incomplete cases consume considerable chasing time. Previous automation attempts break on atypical cases. Nobody can explain why a case was classified a certain way. Workload varies sharply by period with no ability to absorb peaks.
Expected benefits
Reduced processing time on simple cases. Teams focused on genuinely complex cases. Traceable decisions that can be reviewed afterwards. Automation that stops cleanly instead of improvising. Impact measured case by case rather than estimated globally. Progressive scope extension validated by results.
Method and steps
- 1
Task analysis
Observing real processing, measuring time per case category and identifying the missing data that triggers chasing.
- 2
Scope definition
Splitting cases between automatic handling, sampling checks and routing to an operator, validated by business owners.
- 3
Restricted pilot
Implementation on a single case category, with limited tools, full logging and systematic comparison against human decisions.
- 4
Measurement and extension
Analysing deviations, fixing rules, adjusting checks, then extending to other categories once results stabilise.
Deliverables
Task map with current time measurements.
Definition of automatic, checked and manual scopes.
Operational pilot on one case category.
Detailed log of decisions and tools used.
Dashboard of gains and detected deviations.
Progressive extension plan and operations guide.
Technologies used
- Python
- Spring AI
- pgvector
- Redis
- OpenTelemetry
- Keycloak
- LangChain
- API de modèles de langage
Frequently asked questions
Can the agent act directly in our systems?
Only through explicitly authorised tools validated server-side. Irreversible actions, such as sending an email or posting an accounting entry, are prepared by the agent and approved according to your policy.
What happens if the model is unavailable?
Processing falls back to manual mode without blocking: unhandled cases go into a queue operations can inspect, along with the exact state of what has already been done.
How is return on investment measured?
By comparing processing time per case before and after, and by tracking the effective automation rate. Gains are measured on your real cases during the pilot, before any generalisation.
Related case studies
Internal document assistant with guardrails and evaluation
Augmented retrieval over a corpus of standards and technical notes
Demonstration sample - not a real client reference.
Analyse my task
Describe the task and its volume. We will assess what can be automated cleanly and what must stay under human control.