Every number here is reproducible.
No fabricated logos. No borrowed metrics. The figures below are computed live from LouieAuto's own production tables — proven across our own rooftops — with the methodology and the caveats stated next to each claim, the way a diligence team would want to read them.
lender_outcomes. Every routing claim on this page is aggregated from this table.Routing to the best-fit lender approves 8–17 more deals per hundred Proven
Sending each deal to its best-fit lender, instead of an average lender, approves 8 to 17 percentage points more deals — and the gap is widest in the subprime tiers, where an approval is hardest to get and worth the most.
Methodology. Source: lender_outcomes (4,715,030 rows). Approval rate computed per (fico_tier, lender), filtered to cells with ≥500 decisions. "Lift" = best-fit-lender approval minus average-lender approval — the deliberately conservative comparison. Best-minus-worst is far larger (e.g. near-prime best 78.6% vs. worst 0.2%). This is the value at stake in routing — the approval the brain protects by sending each deal to its best-fit lender. We do not claim the brain captures 100% of it.
Stores that lean on the routing brain run materially higher PVR Observational
Across our rooftops, periods with AI-routing compliance at or above 90% show +$628 PVR and +0.67 F&I products per deal versus periods below 90%.
| AI-routing compliance | Avg PVR | F&I products / deal | Periods |
|---|---|---|---|
| ≥ 90% | $3,139 | 2.45 | 6 |
| < 90% | $2,511 | 1.78 | 24 |
| Difference | +$628 | +0.67 | n = 30 |
store_pvr_stats across 5 rooftops — not a controlled trial. Higher-performing stores may both adopt the AI more and run higher PVR, so we frame it as "stores that lean on the routing brain run materially higher PVR," not "AI causes +$628." The controlled, single-dealer version of this number is exactly what the external pilot below is designed to produce.
What the lift is built on Proven
The routing isn't a static rules table. It is a learning substrate that recalibrates every night against real outcomes.
brain_patterns — encoded desk knowledge plus learned signal.lender_weight_cache, the live weight map that matchLenders() reads.sim_runs) plus 4.7M real-shaped lender outcomes.Nightly closed-loop reweight runs at 1:30am (lenderOutcomes.js → lender_weight_cache). The longer an instance runs, the more its routing calibrates to that store's actual deal outcomes and lender mix.
What we do not claim — yet
Integrity is the product. Here is exactly what we are not putting a number on until it is backed, and why.
- No headline AI-accuracy %. Our
ai_outcomestable mixes real-signal rows with seeded demo rows (some seed modules sit at an implausible 99.6%). Until live and seed rows are separated, we will not publish a single calibration number. - "Stip time 47→9 minutes" is retired here. No production table currently carries a stip resolution-time column to substantiate it, so it does not appear as a proven claim on this page.
- Inventory-score → days-to-sell is not yet usable.
louie_score_outcomeshas a single row. We are accumulating real outcomes before we feature it.
Every "Proven" figure on this page is reproducible via aggregate queries on the named production tables. A diligence team can re-run them against data/louieauto.db.
External reference pilot — in progress In progress
Everything above is proven across our own rooftops. The one thing it does not yet include is a third-party dealer who will put their real name behind the numbers. That pilot is being set up now: a controlled before/after at a single independent store, the dealer's metrics captured directly, results published under the dealer's own name with their written approval. This placeholder will be replaced by the named case study when the study completes — not before, and never with a fabricated stand-in.
See the brain make these calls live.
Open the demo and route a deal yourself — the same engine that produced the numbers above is on the floor, right now.