From 12 Reviewers to 6: The Economics of AI Transaction Processing
How a leading automotive marketplace cut review time from 20 minutes to under 2, dropped error rates from 7% to 1%, and halved its review team.
A leading automotive marketplace was running roughly 1,000 vehicle-purchase evaluations per week. Each evaluation required reviewing about five documents (title, lien status, ownership, compliance records), and a team of roughly 12 full-time reviewers worked through that queue every day. A typical review took around 20 minutes, and the error rate hovered near 7%.
None of that was unusual. It’s what document-heavy transaction processing looks like almost everywhere: a trained team, a checklist, a queue, and a throughput ceiling set by headcount.
What triggered the change
The company hadn’t gone looking for AI. What changed was ownership. Under private equity, profitability pressure landed directly on operations: the mandate was to grow transaction volume without growing the cost base. The review team was the constraint: every additional 1,000 evaluations per week meant hiring, training, and managing another cohort of reviewers, and the 7% error rate was generating rework and financial exposure on top of the labor cost.
That’s the pattern we see repeatedly. The trigger is almost never “we want AI.” It’s “we can’t keep hiring into this” or “our error rate is eating margin,” and someone with a P&L is being asked to fix it.
What the process actually was
Before anything could be automated, it had to be made explicit. The review process was nominally documented, but the real version lived largely in the heads of the Deal Support team: which documents count as valid, which discrepancies are deal-breakers versus formalities, how rules shift by state. Rules, experience, and procedures, accumulated over years, mostly unwritten.
The first phase of the work was discovery: sitting with the reviewers, walking through real transactions, and capturing that institutional knowledge as executable logic. Document types, extraction fields, verification checks, state-specific rules, and, most critically, the exception patterns that separate a routine approval from a case that genuinely needs human judgment.
What changed after deployment
The system that went into production follows the same shape the human process did: documents come in, structured data is extracted and verified, business rules are applied, and a decision comes out: approve, reject, or escalate to a human with the context pre-assembled.
The numbers moved in every dimension that mattered. Around 50% of deals are now auto-approved in under 2 minutes, against the previous 20-minute manual review. About 70% of total volume is AI-managed end to end. For the deals humans still touch, review time dropped from roughly 20 minutes to 1–2 minutes, because reviewers receive extracted, verified, rule-checked cases instead of raw document stacks. The error rate fell from about 7% to about 1%.
And the team went from roughly 12 reviewers to 6, with the other half redeployed, not laid off. Freed capacity went into growth work, including expanding the company’s mobile purchasing units, which had been capped by how fast the back office could clear evaluations.
The economics, in plain terms
The way we frame the value of a project like this is three buckets: labor, incremental capacity, and error reduction. Halving a 12-person review function is a direct labor saving. Removing the review bottleneck unlocked transaction capacity that had nothing to do with demand: the deals were there; the queue was the constraint. And cutting errors from 7% to 1% eliminated most of the rework and leakage that never shows up as a line item but compounds every month.
Our rule of thumb: a process is worth converting to an AI-operated workflow when those three buckets add up to $1.2M or more per year, and the automation should cost no more than about 20% of the value it captures. This process cleared both thresholds comfortably, which is why it went from pilot to production instead of stalling as another AI experiment.
What this looks like from the COO seat
The before/after is easy to summarize: 20 minutes to under 2, 7% errors to 1%, 12 reviewers to 6, and a throughput ceiling removed. But the structural change matters more than any single metric. The company’s transaction capacity is no longer a function of how many reviewers it can hire and train. The process itself (the rules, the exceptions, the accumulated judgment) is now an asset the company owns, rather than knowledge that walks out the door with attrition.
If you’re running a document-heavy operation with a similar shape (high volume, multiple documents per transaction, definable rules, a team of five or more working the queue), the full anonymized breakdown is in our AI Deal Engine case study.