Cents per email: running AI on a freight inbox

A freight sales desk lives in its inbox. Customers send enquiries, vendors reply with rates, and threads go back and forth all day. When we put AI on that inbox for Quote Desk, the first bill told us something simple: most of the money was going on email that was never going to be business.

Where did the money go?

We measured the production pilot up to 23 September 2026. By then the AI had read 1,517 emails, about 9.1 million input tokens.

• 81% of the token spend went on emails whose verdict was "other": newsletters, notifications, internal chatter.
• About half the spend sat in threads that were re-read every time someone replied, and never produced business.
• 1,113 of 1,685 attachment reads were repeat copies of 572 distinct files.

None of this is a model problem. It is a question of what you ask the model to read.

Why not filter out the noise?

The obvious fix is a rule: skip emails that look like noise. We tested that idea against the inbox and it failed. In a hand-labelled sample, 66 reply threads that looked like operational noise were carrying real vendor rates. A rule that skips them loses money quietly, which is worse than spending it.

So the principle became: cheap signals decide how much model to spend, never whether to look.

What does Quote Desk do instead?

A cheap screen before a full read

Every email gets a short first look, about 2,066 input tokens against about 5,680 for a full read, roughly a third of the cost. The screen can clear an email only at 95% confidence or higher. Across 185 clears in production, none had a subject that mentioned rates, a quote, an enquiry or an RFQ.

Memory for threads and senders

A thread that has been read is not read again from the start when someone replies. Sender reputation lowers how much is spent on a sender who never sends business, and resets on that sender's first business email.

Half-price batches for what can wait

Doubtful mail, and mail that arrives while its owner is away, goes through the provider's batch interface at half price. When slow batches once caused double billing, we added an overdue fallback that pauses batching and cancels slow batches.

A spend limit that pauses instead of failing

When the provider spend limit is reached, reading pauses. Mail is not dropped and nothing fails; it waits until spend is available again.

How do we know accuracy held?

Every cost change is checked against a hand-labelled set of 50 real emails with attachments. A change ships only if there is no loss of accuracy on money fields: rates, currencies, validity dates. Fields the AI is unsure of still go to a person before they are saved.

What it means for you

If you are putting AI on an inbox, measure what it reads before you tune how it reads. Most of the saving is in not paying for the same thing twice, and in spending less on mail that is probably noise without ever ignoring it.

We build the same controls into every AI document processing project.