Five examples of AI use cases

Business problem

Entering supplier invoices takes up skilled time: choosing the ledger account and VAT code requires knowing the supplier, the nature of the expense and internal rules.

User story

“As an accountant, I want to receive for each PDF invoice a suggested ledger account and VAT code, with its rationale, so that I can validate faster without losing control over the coding.”

Input

PDF supplier invoices; the organisation's chart of accounts and coding rules; coding history, if available.

Output

For each invoice: suggested ledger account and VAT code, with a readable rationale. An uncertain suggestion is flagged as such and is never validated automatically: validation stays human.

What would be measured

Rate of accepted suggestions; validation time per invoice, compared with the current process.

Possible architecture

Sovereign AI or protected external AI, depending on how sensitive the data is.

Explore this use case →

Business problem

In a large volume of entries, anomalies are found by sampling or experience. An atypical entry can go unnoticed.

User story

“As an auditor, I want a prioritised list of atypical entries, with an explanation for each, so that I can focus my checks where they matter most.”

Input

A structured accounting export (journal or general ledger) over a defined period.

Output

A prioritised list of entries judged atypical, with the reason for each flag. The analysis and the conclusion remain the auditor's.

What would be measured

Relevance of the alerts, as judged by the auditor.

Possible architecture

Preferably sovereign AI, as accounting exports are complete and sensitive.

Explore this use case →

Business problem

Shared mailboxes receive messages from suppliers, customers and service providers on very different topics. Manual sorting delays handling and some messages end up in the wrong place.

User story

“As an administrative manager, I want emails received from third parties to be classified by topic and routed to the right destination, so that I can cut sorting time without losing any message.”

Input

Emails received in a defined shared mailbox; list of possible destinations (people or departments).

Output

Classification by topic — invoice, reminder, dispute, other — and a suggested destination. A person confirms or corrects before routing.

What would be measured

Classification accuracy; correction rate; sorting time saved compared with the current process.

Possible architecture

Sovereign AI or protected external AI, depending on the content of the messages.

Explore this use case →

Business problem

Writing minutes after every internal meeting takes time, and notes taken during the meeting are often incomplete.

User story

“As an executive assistant, I want a transcript of a recorded meeting and draft minutes, so that I can review and correct instead of writing everything.”

Input

An audio recording of an internal meeting, made with the participants' authorisation.

Output

A transcript and draft minutes (decisions, actions, open points), produced on local infrastructure: the recording is not sent to an external AI service.

What would be measured

Accuracy of the transcript and the minutes; correction time needed.

Possible architecture

Sovereign AI.

Explore this use case →

Business problem

An internal application calls a defined external AI API. The texts sent may contain sensitive data.

User story

“As a CISO, I want a defined type of sensitive data to be detected and masked before each call to the AI API, so that we can use this service while limiting the exposure of that data.”

Input

Requests from a defined application to a defined AI API; type of sensitive data to handle (for example, people's names or case numbers).

Output

Requests transformed before sending, and a log of detections. Pseudonymised data is still personal data: a contract with the provider is still required.

What would be measured

Detection rate; false positives; impact on the quality of responses.

Possible architecture

Protected external AI.

Explore this use case →