Step one
What we pull from the data
Picture a giant parking lot holding millions of used cars from dealers all over the country. We're hunting for one very specific situation — and it's easiest to understand with an example.
The idea: a car stuck in the wrong place
A franchise dealer is a store that sells one brand brand-new — a Toyota store, a Ford store. Sometimes a Toyota store ends up with a used Ford on its lot (often from a trade-in). That Ford is the wrong brand for them, so it's hard to sell and it just sits. Meanwhile, a Ford store nearby has a used Toyota stuck on its lot for the same reason. Each store is holding exactly the car the other store is built to sell.
Toyota store
Holding a used Ford F-150
Wrong brand for them. Sitting 210 days.
Ford store
Holding a used Toyota Camry
Wrong brand for them. Sitting 190 days.
Swap the two cars and they each land where they actually sell. That's the whole game — and our job is to find every pair like this, automatically.
What makes a pair worth pulling
Every pair we keep clears all of these, with no exceptions:
- Both are our partner dealers. We only match stores that already work with Autoturn (Redline accounts).
- Both are new-car stores. Each one sells at least one brand brand-new, so it can properly recondition and resell what it receives.
- Each car is the wrong brand for its store — but the right brand for the other store. That's the swap.
- Both cars have sat 90+ days. A car that won't sell is aging money — so both dealers have a reason to move it.
- The two cars are close in price. Within $20,000 of each other, and the pricier one never more than about double the cheaper — so the swap feels fair.
- Each car gets its single best partner. A car could pair with many others — we keep only its strongest match, then the best pairs overall.
From everything down to a short list
Those rules narrow an enormous haystack down to a tight, ranked shortlist. Here's roughly how the numbers shrink on a typical run:
Those 4,318 cars could be paired roughly 50,000 different ways. We keep each car's single best partner, then the 150 strongest pairs. (Representative numbers from a recent run.)
We check the real distance
For every pair we ask Google Maps how far apart the two dealers really are and how long the drive is. Moving cars costs money, so closer trades are easier to pull off.
We pull each car's history
We look up the history report on each car — accidents, number of owners, title problems, and whether it was ever a rental or lease. About 6 in 10 cars have a report; the rest are simply treated as "unknown."
Step two
We give each swap to an AI analyst
Now we hand every pair to an AI analyst — one trade at a time — and give it a clear job. We don't ask it to do anything technical. We just give it a briefing, the same way you'd brief a new hire on the trade desk.
In
One swap, in plain numbers
prices, days sitting, distance, history
AI weighs it against the rules
Out
A score, the reasons, two emails
0–100 · pros & cons · ready to send
The briefing we give the AI
Your job
"You're a trade-desk analyst for a dealer-to-dealer brokerage. For each swap: give it a score from 0 to 100 for how likely the trade is to actually happen, list the reasons for and against, and write a short, friendly email to each dealer. A human reviews and sends every email — you never contact anyone yourself."
The rules you judge by
The price gap (our CEO's rule). A gap under $7,500 doesn't hurt the odds at all. Bigger gaps slowly lower them — the wider the gap, the more it bites.
Fit is everything. Each store ends up with a car it's built to sell. A clean fit both ways is the heart of a good trade.
Both cars stuck a long time = both dealers eager. The longer the two cars have sat, the more motivated everyone is to make a deal.
Closer is better. The shorter the drive between the two stores, the cheaper and easier it is to move the cars.
Read the history. Two damaged cars matched together is fair to both sides. A branded (salvage) title is a big red flag. Former rental or taxi use is a minor caution. A missing report is treated as neutral — never held against the trade.
Watch for same-owner stores. If both stores secretly belong to the same company, the swap is pointless — they'd just move the cars internally. The AI flags this and scores it way down.
How a score turns into a rating
The AI is given fixed examples so a "75" means the same thing on every trade. The 0–100 score lands the trade in one of three buckets you'll see on the report:
under 50 · needs a push
50–69 · most signals good
70+ · everything lines up
A top score (85–100) looks like: both cars sitting 200+ days, a price gap under $7,500, the two stores within 150 miles, and a clean fit both ways. Two hard caps keep things honest — a likely same-owner pair can't score above 49, and stores at the same address score below 35.
And the two emails
Each email is written at a 5th-grade reading level — short, plain, and friendly — using the trade's real numbers (how long the car has sat, the price gap) and ending with one simple ask: "Reply yes and I'll set it up." The AI is told never to invent facts, like transport promises it can't keep.
See the exact instructions we give the AI
You are a trade-desk analyst for a dealer-to-dealer brokerage. Each trade you
receive is a two-way FRANCHISE SWAP: dealer A holds an aged unit that dealer B
is franchised for, and dealer B holds an aged unit that dealer A is franchised
for. Your job: score each trade 0-100 on how likely it is to produce a
positive trade, explain why with pros and cons, and draft a short persuasive
email to each dealer. A human broker reviews and sends — you never contact
anyone.
## Ranking rubric
- Price delta — the CEO's rule, apply it exactly:
"Trade deltas less than $7500.00 have no impact on the likelihood of producing a positive trade.
Trade deltas more than $7500.01 have a slightly lower likelihood of producing a positive trade.
The reduced likelihood of a positive trade reduces as the cost delta increases."
- Franchise fit is the demand signal: each dealer is franchised for the unit
it receives, so it can CPO/retail it. Clean fit both ways is the core of a
good trade.
- Combined days_on_lot = mutual motivation: both units stuck means both
dealers want out.
- distance_mi: lower = cheaper transport and easier logistics.
- Vehicle history (each unit's "history" field):
* "unavailable" = no report found. NEUTRAL: it must not move the score and
must NOT appear as a pros or cons bullet at all.
* Both units with damage/accidents reported: a PRO — damaged units are best
offered against other damaged units (matched condition, fair both ways).
* One damaged, one clean: slightly lower likelihood — the clean side takes
on the harder unit; name the damaged unit in a con.
* Both clean ("No accidents"): a small pro; one-owner adds a touch more.
* branded_title true on either unit: strong negative — major con, score it
down hard.
* Prior rental / taxi / commercial / government use: mild watch-out, worth
a con bullet naming that unit. "Previously leased" is common and NOT a
penalty — never a con.
- Same-group check: compare the two dealer names and addresses. If both
rooftops plausibly belong to one auto group — shared name root ("Cavender
Toyota" / "Cavender Buick"), a known group brand on both, or the same or
adjacent street address — a brokered swap adds little: the group would just
transfer the units internally. Flag it as a con naming the evidence, and
score it down; identical addresses make the trade near-worthless.
- Sanity-check the price ratio and mileage-for-year on both units.
## Score calibration anchors
Batches are scored in independent sessions; these anchors keep scores
comparable across batches. Calibrate to them, not to the batch in front of you.
- 85-100: both units 200+ days on lot, delta <= $7,500, < 150 mi apart, clean
franchise fit both ways.
- 60-84: solid — most signals good, one soft (e.g. delta slightly over $7,500
OR 150-400 mi apart).
- 35-59: stretch — multiple soft signals (big delta with the CEO curve biting,
far apart, marginal fit).
- <35: weak.
- Likely same ownership group caps the score at 49 regardless of other
signals; same street address caps it below 35.
tier = "strong" (score >= 70), "solid" (50-69), "stretch" (<50).
## Email guidance
Draft TWO emails per trade: email_a TO dealer A, email_b TO dealer B, each as
{"subject": ..., "body": ...}. VERY simple and clear: 5th-grade words, short
sentences, one idea per sentence, ~90-130 words per body.
Write the body as Markdown so it reads as a real, scannable email: open with a
short greeting line, then 2-3 short paragraphs separated by BLANK lines, and put
the closing CTA on its own line. Bold the one or two numbers that matter most
(days on lot, the price gap) with **...**. No headings, tables, or bullet lists
— keep it a normal short email.
Persuasion levers — use the trade's real numbers:
- Loss aversion: their unit has sat N days and is aging money.
- Fit: "you're the {make} store — this {model} sells on your lot."
- Reciprocity/fairness: the other dealer takes your aged unit; the price gap
is only $X.
- Clean history, only when the incoming unit's history SAYS "No accidents":
"no accidents reported" is a real selling point. Never mention history that
is unavailable, and never put a unit's damage in the email pitching it.
- Ease + a single CTA: "Reply yes and I'll set it up."
NEVER invent facts: no transport promises, no market claims beyond the data.
Subject <= 8 words, plain — "A {their aged unit} for your {incoming unit}?"
style.
## Output rules
- Respond with a JSON object {"rankings": [...]} containing EXACTLY one entry
per trade in the batch, keyed by its trade_id.
- pros and cons: 2-4 short bullets each, every bullet citing concrete numbers
from the trade (days, dollars, miles); a same-group flag instead cites the
matching names/addresses.
- The score 0-100 is your judgment against the rubric and anchors, not a
formula.
What comes back for our Toyota ↔ Ford swap
Why it works
- Both cars have sat 190–210 days — both dealers want out.
- Each store receives the exact brand it sells new.
- Price gap of $3,100 — well under the $7,500 line.
Watch outs
- The two stores are 280 miles apart — some transport cost.
Hi there,
You've got a Camry that's been on your lot 190 days. It's a great car — it just sells faster at a Toyota store.
I've got a Toyota store that will take it and send you a clean F-150 in return. The price gap is only $3,100, and it's exactly the truck your buyers ask for.
Reply yes and I'll set it up.
Step three
We rank them and hand them off
Once every swap has a score and a pair of emails, we put it all in order and present it for a person to act on.
Best trades to the top
We sort by score, highest first. When two trades tie, the one whose cars have sat longest wins — those dealers are the most motivated.
Grouped into tiers
Each trade is tagged Strong, Solid, or Stretch, so a broker can start with the surest wins and filter the rest.
A person makes the call
The list lands in the broker's report. They read the reasons, check the emails, tweak if needed, and hit send. The AI never emails anyone itself.
That ranked list is the report
Everything above ends up in the live report your team works from every day.