It reads the request — typos and all. It pulls your lane history from the TMS and a live market rate from DAT or Greenscreens. It prices under your margin rules and drafts the reply. Then it stops, and waits for a human. Every quote, every time, until you say otherwise.
READS: EMAIL · TMS · DAT/GREENSCREENS · DRAFTS: ~90 SEC (DESIGN TARGET) · AUTO-SENDS: NEVER ON DAY ONE
Draft · held for rep approval
THE AGENT CANNOT SKIP THIS STEP. ONLY YOU CAN WIDEN IT — SEE THE TRUST LADDER BELOW. SAMPLE DATA.
Click any station to see what actually happens there. Station 08 is the one that matters: a human, with a button.
STATION TIMES ARE DESIGN TARGETS FROM OUR TEST CORPUS, NOT GUARANTEES. THE GATE HAS NO TIMER — IT WAITS FOR YOUR REP.
Real quote requests come in misspelled, half-specified, and sent from a phone at a loading dock. Here's how the agent reads one — and what it does about the parts it can't read. Click any highlighted phrase or table row to trace it.
Inbound · Tue 2:14 PM · quotes@yourbrokerage.com
GREEN = READ CLEAN · ORANGE = FLAGGED — THE REPLY WILL ASK
Extraction · per-field confidence
| Field | Read as | Conf |
|---|---|---|
| ORIGIN | KANSAS CITY, MO ("kanas city" — typo resolved) | 0.94 |
| DEST | MEMPHIS, TN ("memhpis" — typo resolved) | 0.95 |
| EQUIP | ⚑153' DRY VAN — ASSUMED FROM "van" | 0.82 |
| WEIGHT | ⚑2NOT GIVEN — 26 PALLETS GROCERY | — |
| PICKUP | WED 06/11, AM ("tomorow" resolved vs. received-date) | 0.91 |
Two fields didn't clear the bar. The agent doesn't silently guess on them — it prices what it can defend and puts both flags in the reply as questions.
Draft reply · held for rep approval · both flags confirmed in writing
Why this matters
A wrong assumption that's stated costs you an email. A wrong assumption that's hidden costs you a claim. Every flag the extractor raises becomes a sentence in the draft — so the shipper confirms it, in the thread, before anyone books a truck.
And if too much is unreadable — no lane, no date, gibberish — the agent doesn't draft at all. It escalates with whatever it managed to parse. See the escalation list below.
RATE SHOWN IS SAMPLE DATA. CONFIDENCE THRESHOLDS ARE TUNED PER CLIENT DURING THE PILOT.
Nothing auto-sends on day one — and nothing ever does unless you switch it on. The agent earns rungs with a scorecard; only you grant them. Drag the slider.
SPLITS SHOWN ARE ILLUSTRATIVE — YOUR WHITELIST AND CAPS DECIDE THE REAL NUMBER. ONE TOGGLE DROPS THE WHOLE DESK BACK TO RUNG 0, INSTANTLY.
The agent's most important skill is knowing what it doesn't know. This list never auto-quotes — at any rung of the ladder, forever. That's not a limitation we apologize for; it's most of why the desk is safe to run.
| What came in | Why it never auto-quotes | What your rep gets |
|---|---|---|
| Multi-stop loads | Stop-off fees and routing order are judgment calls, not lookups. | FULL EXTRACTION + LANE DATA, PACKAGED |
| Reefer partials | Temp liability plus consolidation math — humans only. | TEMP, PALLET COUNT, COMPATIBLE-FREIGHT NOTES |
| Hazmat | Placards, endorsements, insurance. Not guessing territory. | FLAGGED CLASS + HANDLING REQUIREMENTS |
| Brand-new shippers | No lane history, no credit signal, no pattern to price against. | LOOKUP RESULTS + NEW-SHIPPER CHECKLIST |
| Low-confidence extraction | If the agent isn't sure what it read, it doesn't price it. | RAW EMAIL + WHAT IT COULD AND COULDN'T PARSE |
| Anything that smells like fraud | Mismatched reply-to, look-alike domains, rates too good to be real. | RED-FLAG REPORT — BEFORE ANYONE REPLIES |
ESCALATIONS COUNT AS PROCESSED ITEMS — THE AGENT DID THE READING AND THE PACKAGING. WHAT IT DIDN'T DO IS GUESS.
Lane history comes out of your TMS through a read-only, credentialed connection scoped during setup week. Everything the agent does goes back in as a record your auditors — and your gut — can check.
| Lane history | PAST LOADS · RATE BAND · CARRIERS |
| Customer record | TERMS · CONTACTS · NOTES |
| Margin ruleset | YOUR FLOORS/TARGETS · VERSIONED |
| Market rates | DAT · GREENSCREENS |
DIFFERENT TMS? CHECK FIT — WE SAY NO WHEN IT WON'T WORK.
When a shipper disputes a number in October, you pull this record and see exactly what the market said, which ruleset priced it, and which human approved it — in one place.
Every DeadheadDesk agent runs inside an eval and monitoring harness. The failure mode is designed: a degraded agent escalates more. It never guesses more.
Nightly · regression
The agent re-runs a corpus of real freight quote emails — clean ones, messy ones, the ones that broke it before — scored per field against known answers. A score dip blocks changes from shipping.
Always on · drift watch
Confidence and rep-edit rates are tracked per sender. When a shipper's new ERP changes how their requests read, confidence sags — and the harness notices before a human would.
The failsafe
A low-confidence email doesn't get a worse quote. It gets no quote — it routes to your reps as if the agent were never there, and the alarm pages us. Not you.
The patch
We fix the parser, add the new format to the regression corpus so it can't quietly break again, and you read about it in the weekly report — in dollars and loads, not log files.
SILENT-FAILURE ALERTING IS A CONTRACT TERM, NOT A FEATURE FLAG. IF THE DESK DEGRADES AND YOU FIND OUT FROM A SHIPPER BEFORE YOU HEAR IT FROM US, THAT'S OUR FAILURE — IN WRITING.
0 SEC
Median request → draft
Design target from our test corpus.
0%
After-hours capture
Overnight requests with a draft waiting before your reps sit down.
+0%
Quote-to-cover
Illustrative model: faster replies win more of the loads you quote.
THESE ARE DESIGN TARGETS AND MODEL OUTPUTS — NOT CUSTOMER RESULTS. THE 30-DAY PILOT REPLACES THIS STRIP WITH YOUR OWN BEFORE/AFTER NUMBERS, MEASURED ON YOUR INBOX.
30-day paid pilot, $2,500, credited against setup. Shadow mode first, approval gates always, and a go/no-go decision made on hard before/after numbers — not a sales deck.
$2,500 PILOT (CREDITED) · $7,500 SETUP · $2,000/MO INCL. 500 ITEMS · $1.50/ITEM AFTER