In a measured test of 100 realistic freight broker emails, letting a lightweight classifier sort every email first, and sending only the harder ones on to Claude, kept accuracy at 99% and cut the cost per 1,000 emails from $2.30 to $1.00: 57% less. Two-thirds of the inbox never needed the larger model.
| Setup | Handled correctly | Cost per 1,000 emails | Reached Claude |
|---|---|---|---|
| Claude Sonnet on every email | 99% | $2.30 | 100 of 100 |
| Jev first, Claude only when needed | 99% | $1.00 | 34 of 100 |
66% of the inbox never needed the larger model. The 57% saving is the financial result; the 66% is the reason for it.
These figures come from one specific, measured run of 100 emails, each checked against a known right answer. The rest of this article explains how the test was built, why the approach works, and when it is not worth doing.
Why does AI email sorting cost more than it needs to?
Because the simplest design sends every email to the strongest model, and in this inbox, most emails turned out to be easy.
A freight broker’s shared inbox is a good example. Shippers ask for quotes. Carriers offer trucks. Drivers send check calls. Signed rate confirmations come back. Customers ask about invoices. In between arrive “thanks, got it” replies, out-of-office notices and newsletters. Someone has to read every one just to decide who handles it.
A capable model reads all of that well. But it charges the same rate to recognise an out-of-office reply as it does to pull a lane, equipment type, pickup date and weight out of a quote request buried in a long thread. Without measuring which emails actually need the stronger model, you end up paying its price for all of them.
How does sorting first actually work?
It is a two-step process with an escape hatch.
Step one: a lightweight classifier reads every email. We used Jev, a lightweight AI model from TypeSafe built for fast classification. Rather than writing text, it picks one answer from a fixed list (here, which of six categories an email belongs in) and reports how confident it is. It costs a small fraction of what a large model does and answers in a fraction of a second.
Step two: the larger model only when needed. Two kinds of email go on to Claude. Quote requests always do, because someone needs the load details pulled out ready to price, and a classifier cannot write those out. And any email Jev is not confident about goes on too.
Everything else is done. Emails Jev is confident about never reach Claude. That is where the saving comes from.
The confidence bar is a dial. Turn it down and the classifier handles more on its own: cheaper and faster, with more chance of a mistake. Turn it up and more goes to the larger model. The right setting depends on what a mistake costs you, which is a business decision, not a technical one.
Don’t use one model for everything. Route the work: a fast, inexpensive model for the easy majority, a stronger one for what genuinely needs it.
How do you know the cheaper version is as accurate?
You test it against emails where you already know the right answer, not against another AI’s opinion.
We wrote an inbox of 100 broker emails across six categories: quote requests, truck offers, status updates, rate confirmations, billing, and everything else. About a fifth are deliberately tricky: long threads where the real request is three replies down, signatures longer than the message, a quote request buried in small talk, a check call that mentions a rate. Because every email was written from known values, the correct category and every correct load detail are known in advance. “Correct” is a fact.
The scoring is strict. An email only counts as handled correctly if the category is right and, for a quote request, every extracted field is right: origin, destination, equipment, pickup date and weight. One wrong field and the whole email counts as wrong.
We then ran the same inbox several ways, Claude Opus, Sonnet and Haiku each on every email, and the sort-first process, recording every answer, its cost and its time per email.
Which emails still needed the larger model?
The quote requests and the emails the classifier was unsure about: 34 of the 100.
With the confidence bar at 90%, Jev settled the other 66 on its own, and the overall result still matched Claude Sonnet reading every email. The larger model was still doing the hard work; it simply was not being paid to do the easy work as well.
Because every model’s answer to every email is stored, the dial can be moved after the fact without running anything again. You can see exactly which emails would have gone wrong at a lower setting, or how much more a higher one would cost, before choosing.
The saving grows in step with volume, because in this test two-thirds of the emails never reached Claude. If your own mail has a similar mix, the share handled by the inexpensive model, and so the saving per email, carries over as volume grows.
When is this the wrong approach?
When your volume is low, when nearly every item is genuinely hard, or when you cannot say what a correct answer is.
Low volume. If your team handles a few hundred emails a month, the difference between $1.00 and $2.30 per thousand is not worth building a routing layer for. Use one good model and spend the effort elsewhere.
Uniformly hard work. Routing saves money only when a large share of the input is easy. If nearly every item needs careful reading and extraction, nearly everything escalates, and you pay for two models instead of one.
No answer key. Without a set of real examples labelled with the right answer, you cannot tune the dial, and you are back to trusting a demo. Building that set usually means a few hundred of your own emails, labelled once by the person who normally sorts them. It is the least glamorous part of the project and the part that makes the numbers defensible.
And some mistakes are expensive enough that the dial should sit high regardless of cost. The same principle from what to automate first in logistics applies: automate where a mistake is visible and cheap, keep people where it is quiet and costly.
What would this look like on your own inbox?
The same process with your categories, your answer key and your threshold.
A 3PL might add appointment scheduling or claims as categories. The answer key comes from your own emails. The threshold is set by looking at which mistakes you are willing to accept at which price, with the numbers in front of you. And the output feeds wherever the work already happens, a TMS, a ticketing system or a shared mailbox with labels, rather than a new screen someone has to remember to check.
The measurement is what carries over. Before anything goes live, you know the accuracy and the cost per thousand emails on your own mail, which answers the question every operations leader asks first: will this hold up in production, and what will it cost at our volume?
See it live
The demo is a working application, not a slide. Watch all 100 emails get sorted four ways, open the tricky ones to see why they are tricky, and move the confidence dial to watch cost and accuracy change.
See the live inbox-routing demo →
Want to see what this would look like against your own workflow? Talk to us.