AI Email Inbox Agent for a Logistics Company
A logistics company received 200+ emails daily: inquiries, confirmations, complaints. We built an agent that reads, classifies, and replies to 80% of emails automatically.
- 80%auto-replied
- −3 hrmanager work / day
Industry context
In logistics a shared inbox works as a dispatch desk: rate requests, confirmations, documents, route changes and complaints all arrive in one stream from different parties at once. The problem is not the number of emails but that the cost of each is different and unpredictable. A rate request answered within an hour is a won load; the same request noticed in the evening is lost to whoever replied first. A delay complaint left for a day turns into a client conflict. So the real job of automation here is not to answer everything but to stop the important from drowning in the routine.
The challenge
The shared inbox became a bottleneck: managers spent hours sorting mail, important requests sank, and response time stretched into days.
Our approach
- 01
Email classification
We taught the agent to sort mail by type: inquiry, confirmation, complaint, other.
- 02
Reply templates
We assembled vetted answers to common emails for auto-generation.
- 03
Human in the loop
Non-standard and risky mail is handed to a manager with a summary.
How the solution works
- 01Gmail APIAn email lands in the shared inbox
- 02GPT-4oClassifies it and drafts a reply
- 03n8nRuns the logic and sends
- 04AirtableKeeps the register of emails and statuses
Via the Gmail API the agent reads incoming mail, GPT-4o classifies and drafts replies, n8n runs the logic and sending, and a registry of mail and statuses lives in Airtable. 80% of routine mail is closed without a human.
Why this stack
The Gmail API provides access without migrating to another system: the team keeps working in their familiar inbox while the agent reads and writes in the same threads. That removes the biggest risk in this kind of project: resistance from a team that does not want to change tools. GPT-4o classifies and drafts replies, because in email the boundaries between types are blurred: one message can be a confirmation and a complaint at once, and rules cannot capture that. n8n runs the logic and, crucially, the order: classify first, verify next, send only then. Airtable holds the registry of mail and statuses. Without it you cannot answer the simple question “what is currently unanswered”, which was the original problem.
Where projects like this usually break
A reply that became a commitment
In logistics an email with a price or a deadline is effectively an offer. An agent that quotes a rate on its own creates a commitment the company may not be able to meet. Anything containing a price, a deadline or a guarantee must pass through a human, even when it looks routine.
A thread without memory
Correspondence stretches for weeks and resumes after pauses. An agent looking only at the last message replies out of context, greeting a client on day ten of an ongoing conversation. Full thread context is mandatory, otherwise automation looks worse than silence.
Auto-replying to a complaint
The fastest way to lose a client is answering an emotional complaint with a correct template. The classifier has to recognize tone, not just topic, and such emails must go to a human as top priority, with no attempt to reply.
Results
| Before | After |
|---|---|
| Managers spent hours sorting email | GPT-4o sorts the mail, 3 fewer hours of work a day |
| Important requests drowned in the inbox | Every email and its status is in the Airtable register |
| Replies took days | 80% of emails get an automatic reply |
Classification + auto-replies for common mail within 2 weeks.
Who this fits
This fits companies with a shared inbox receiving dozens of structurally repetitive emails a day: logistics, wholesale, service companies, travel. The main signal that it is time to automate is not volume but loss: if requests regularly sit unanswered until the next day, you are already paying for it. If instead every email is unique and needs negotiation, automate only classification and prioritization and leave replies to people. And note: if the company has no internal agreement on what counts as urgent, that has to be defined first. An agent will execute a rule but will not invent one.
Frequently asked
Does the agent reply on our behalf without review?
Only mail with an unambiguous answer closes automatically: shipment status, company details, routine procedural questions. Anything containing a price, a deadline or a commitment is drafted by the agent but sent after human approval.
Do we need to change our email system?
No. The agent connects to the existing mailbox through an API and works inside the same threads, so the team keeps its familiar interface. That is a deliberate choice: projects requiring a migration take far longer to adopt.
What about emails in other languages?
The model handles multiple languages, so classification does not depend on the language of the email. Replies are written in the same language the sender used. For international logistics that is a baseline requirement, not an option.
How do we know when the agent is wrong?
The Airtable registry stores every email, the type assigned to it and what was done with it. That makes it possible to review a sample periodically and see where classification drifts before an error turns into a complaint.