Knowing which clients are waiting for a reply
Follows client conversations across 139 mailboxes: who is waiting for a reply, for how long, and which messages point to an unhappy client.
- Role
- Lead developer
- Client
- Accounting firm, Paris
- Year
- 2026
- mailboxes followed
- 139
- messages backfilled in a day
- 175,000
- AI budget per month during the pilot
- €1
Context
Around 5,000 client emails arrive every month. Nobody could tell which ones were still unanswered, or where a relationship was getting tense.
What I built
- 01Real-time mailbox sync through Microsoft Graph
- 02Every client message paired with its reply, with response times per team
- 03AI detection of messages from unhappy clients, under a cost ceiling
- 04An API for the BI team
Key decisions
A year of history in a day
Backfilling twelve months across 139 mailboxes, about 175,000 messages, took a single day: 1,500 messages a minute without one throttling error from Microsoft's API.
One euro a month
The AI analysis was priced before it was built: €1 to €10 a month for the whole firm. The pilot runs capped at one euro, and the GDPR questions were raised before go-live, not after.
Outcome
The firm can finally see which clients are waiting and which are losing patience: across twelve months of history, 7% of the emails analysed signalled discontent.