AI · operations platform
Inbound e-mails classified by a model that learns
For a tour operator, an operations platform receives e-mails, attaches them to the right file and sends uncertain cases up to an operator. Every correction trains the model.
Context
A tour operator receives e-mails every day from clients, partners and suppliers: cancellations, changes, complaints, invoices, new requests. Each must be attached to the right file and handled by the right person. Data entry and sorting kept operators away from the urgent files.
A model trained on the team’s decisions
The model learnt from the history of e-mails and the decisions operators took: which file, which category, what next. It was not configured with hand-written rules; it reflects the team’s real practice, and it keeps learning in production.
A confidence threshold, a person who decides
When the model is sure, the e-mail arrives attached to the right file, ready to be handled. Below a confidence threshold, the case goes up to an operator, who decides. That decision becomes a training example: the model learns from every correction.
An architecture that takes the load
Event-driven, serverless architecture on AWS, GraphQL API with server-side authorization, separate environments, deny-by-default authorization. The model’s decisions are logged and can be consulted by the team.
Result
Frequently asked questions
Does the model replace the operators?
No. It does the attaching and the classification; operators decide on uncertain cases and handle the files. Their work changes in nature, it does not disappear.
What happens when the model is wrong?
The operator corrects it, and the correction retrains the model. The confidence threshold is set by management according to the error rate it accepts.
Where is the data?
In a separate environment, under the client’s control, with deny-by-default authorization. Nothing is used to train a public model.
Can the same approach apply to another flow?
Yes: forms, quote requests, inbound documents. The method is the same: a history of decisions, a threshold, human validation.
Related services
A comparable case on your side?
Tell me which process costs you time; I get back within 24 business hours with a first read, no obligation.