An AI pricing agent should recommend and execute changes inside written limits. One bad vehicle-history match, weak comp set or missing cost field turns automated pricing into an automated gross loss.
An AI pricing agent repriced 80 used vehicles on Sunday night.
By Monday morning, a Silverado listed at $32,000 had been reduced to $24,000. The agent used weak MMR comparisons and misread the vehicle-history data.
The truck sold within two hours.
The expected selling price was $30,800. The $24,000 transaction removed $6,800 from expected gross before management opened the morning report.
$30,800 expected price − $24,000 sale price = $6,800 lost
The agent followed its goal: move inventory faster.
Nobody gave the system a second instruction strong enough to stop a bad price.
Before controls, the dealership recorded 12 material pricing errors a month. After six approval rules went live, the monthly count fell to one. That is 11 fewer errors. The title uses 10 as the conservative operating target.
The six dealership AI agent controls below define where automation stops and human approval begins.
The GM’s AI Pricing Approval Sheet
| # | AI Approval Rule | Error Prevented | Example Cost | How the Rule Works | Alert or Action |
| 1 | No price reduction above 5% without approval | Large overnight price cut | $6,800 | Agent stops when proposed reduction exceeds 5% | GM receives approval request |
| 2 | Low confidence or fewer than eight valid comps blocks execution | Price based on weak market data | $2,400 | Current price stays live until UCM review | Hold and flag |
| 3 | Structural, title or major history flag requires manual pricing | Clean-history pricing applied to damaged-history unit | $4,200 | History flag removes auto-pricing permission | Manual review only |
| 4 | Projected front loss above $500 requires Owner and GM approval | Agent prices below total cost | $500 or more | DMS cost, recon, pack and fees create a gross floor | Block and alert |
| 5 | More than two price changes in seven days freezes the unit | Repeated reductions without new evidence | $1,200 | Third proposed change enters review queue | Seven-day freeze |
| 6 | New listing receives no automatic reduction for seven days | Early markdown before enough market response | $1,800 | Agent monitors but does not lower price | UCM controls first week |
The percentages, comp counts and loss limits are internal policy thresholds from the pricing-control model used for this article. Each dealership should set limits around inventory mix, gross targets and risk tolerance.
#1: No Price Drop Above 5% Without Human Approval
The Silverado started at $32,000.
A 25% reduction produced the $24,000 price:
$32,000 × 25% = $8,000
$32,000 − $8,000 = $24,000
Under the approval rule, the largest reduction reaching the website without human review would have been 5%:
$32,000 × 5% = $1,600
$32,000 − $1,600 = $30,400
The proposed $24,000 price would have stopped in the approval queue. The GM or Used Car Manager would have seen an $8,000 overnight change before publication.
Five percent should serve as an outer approval limit, not permission for every vehicle to drop 5%. On a $70,000 truck, 5% equals $3,500.
A safer control uses both a percentage and a dollar cap:
“Block any reduction above 5% or $1,000, whichever limit is reached first.”
The alert should show:
- Current price
- Proposed price
- Dollar reduction
- Percentage reduction
- Vehicle cost
- Current projected gross
- Reason supplied by the agent
- Market comparisons used
“Approve” and “Reject” are not enough. Add “Edit Price” and “Send to UCM” so management does not accept a bad recommendation to clear the alert.
Action: Find the maximum automatic reduction in your pricing settings. If no limit exists, remove automatic publishing until one is available.
#2: Weak Comparison Data Blocks Automatic Pricing
An AI pricing model needs useful comparisons.
Three vehicles spread across different trims, drivetrains and markets do not support a precise price. The system might still produce a polished recommendation.
Confidence language makes the guess look stronger than the evidence.
Use two separate controls:
- Model confidence must reach the dealership’s approved threshold.
- The comp set must contain at least eight valid vehicles.
The 85% figure belongs to the AI model’s confidence output, if the tool provides a documented score. Do not call the number “MMR confidence.” Manheim Market Report and an AI vendor’s confidence calculation are separate data points.
Every comp should pass basic matching rules:
- Same model generation
- Comparable trim
- Same drivetrain
- Similar mileage
- Similar title and accident history
- Relevant geographic market
- Recent transaction or active-listing date
Example:
Proposed price: $27,400
Valid comps: 3
Comp range: $23,900 to $32,800
Model confidence: 62%
That recommendation has a $8,900 comparison range. The current price should stay live until a manager reviews the unit.
A tighter example looks different:
Valid comps: 12
Comp range: $28,200 to $29,900
Model confidence: 91%
The second set supports a smaller automatic adjustment inside the dealership’s gross and percentage limits.
Confidence scores also need testing. Compare each score with the eventual retail sale, wholesale exit and days in inventory. A 90% score has no value when those recommendations miss the market repeatedly.
Action: Require the dashboard to show confidence, comp count, comp range and excluded vehicles. No evidence means no automatic change.
#3: Frame, Salvage, Lemon and Structural Flags Require Manual Pricing
A history report shows prior structural damage.
The AI agent matches the vehicle against clean-title, no-accident units and recommends $28,000. A manager later determines the correct local retail position is closer to $21,000.
The dealership now faces a pricing problem, disclosure risk and possible deal unwind.
Do not let the system treat every negative-history term as the same issue. An accident report differs from a salvage title. Airbag deployment differs from confirmed structural damage.
Create a manual-review block for:
- Salvage or rebuilt title
- Manufacturer buyback or lemon history
- Flood history
- Confirmed structural or frame damage
- Odometer discrepancy
- Airbag deployment
- Branded-title record
- Unresolved history-report conflict
The block does not mean the vehicle should never be retailed. The rule means a manager must verify the history, legal disclosures, market comps and price before publication.
Feed the same status into the VDP, pricing tool and deal documents. A manager who corrects the price but leaves clean-history language on the website has solved only one part of the problem.
If two data sources conflict, hold the vehicle offline. Do not let the agent choose whichever source supports the easier price.
A $4,200 unwind estimate should include more than the refunded front gross:
- Transportation
- Detail and recon
- Registration work
- Employee time
- Replacement transportation
- Legal review, when required
Action: Connect vehicle-history flags to a hard pricing block. Only the GM or Used Car Manager should release the unit after review.
#4: A Projected Front Loss Above $500 Requires Owner and GM Approval
Market pricing alone does not protect dealership gross.
The agent sees a $26,000 market price. The DMS shows $27,000 total cost.
Without the cost feed, the recommendation appears competitive. From the dealer’s P&L, the proposed price creates a $1,000 front loss.
Use total cost:
| Cost Field | Amount |
| Acquisition cost | $25,100 |
| Auction and transport | $650 |
| Completed recon | $900 |
| Pack or policy adjustment | $350 |
| Total cost | $27,000 |
Proposed selling price:
$26,000
Projected front gross:
$26,000 − $27,000 = −$1,000
The pricing agent should stop.
A strong gross-floor rule reads:
“If proposed advertised price produces front gross below negative $500, block publication and require GM plus Owner approval.”
Use your accounting definition of total cost. Missing recon invoices create a false gross figure. If recon is incomplete, include an estimated reserve until the RO closes.
The alert should identify which cost created the loss. Management needs to see whether the problem came from acquisition, transport, recon, pack or a falling market.
Some aged units need a controlled loss. Automation should still present the loss for approval. The system should never hide a negative deal inside a “market adjustment” label.
Action: Test the cost feed on 20 units. Compare agent cost with the DMS and accounting report before activating the gross floor.
#5: More Than Two Price Changes in Seven Days Freezes the Unit
The agent drops a vehicle by $500 on Monday.
No sale arrives, so the agent drops another $500 on Wednesday. Friday brings a third $500 reduction.
Total weekly change:
$500 + $500 + $500 = $1,500
The third reduction might produce the sale, but the system has not proved whether the first two prices failed. Inventory feeds need time to update. Search filters, merchandising and buyer activity also affect response.
Use this control:
- First change: Allowed inside approval limits
- Second change: Allowed only with new market evidence
- Third change within seven days: Frozen for human review
The review screen should show what changed between recommendations:
- New competing listing
- Sold comp
- MMR movement
- VDP views
- Leads
- Appointments
- Price rank
- Days in inventory
A flat two-change rule also needs an emergency path. A data-entry mistake or OEM incentive change might require an immediate correction. Log the exception and the person approving it.
Do not force every unit to wait 14 days. A badly priced commodity vehicle deserves a prompt correction. The control targets repeated automated reductions without new evidence.
Action: Add a seven-day price-change counter. Place the third proposed reduction in the UCM queue.
#6: New Listings Get a Seven-Day Automatic-Reduction Freeze
A vehicle goes live Monday morning.
The agent reviews traffic Tuesday night and sees weak VDP activity. By Wednesday, the price drops $1,800.
The listing had barely entered shopping feeds. Photos were still processing. Some third-party sites had not updated. The dealership reduced the price before gathering a usable market response.
Place every new listing into a seven-day protected period.
During the first week, the agent should monitor:
- Price rank
- Comparable inventory
- VDP views
- Leads
- Photo completion
- Description status
- Feed errors
- Days to front-line readiness
The system should alert management when the opening price sits outside policy. It should not publish a reduction without approval.
A seven-day freeze is not an excuse for a bad launch price. The UCM needs to price the vehicle correctly before publication.
Do not start the seven-day clock at acquisition. Start when the vehicle becomes retail-ready with complete photos, correct equipment, accurate mileage and a live VDP.
Emergency exceptions include:
- Incorrect original price
- Wrong trim or equipment
- Duplicate stock record
- Material market change
- Manager-approved retail campaign
Action: Tag the retail-live date separately from acquisition date. Block automatic reductions through day seven.
Owner Policy: The One-Page AI Pricing Control
Your policy should name who has authority at each level.
| Pricing Action | AI Authority | UCM Authority | GM Authority | Owner Authority |
| Reduction inside approved percentage, dollar and gross limits | Recommend or execute | Review exceptions | Audit | No routine action |
| Reduction above 5% or dollar cap | Block | Recommend | Approve | Review large exceptions |
| Low-confidence recommendation | Block | Set price | Review | No routine action |
| Structural or branded-title vehicle | Block | Review and recommend | Approve | Review policy exception |
| Projected gross below negative $500 | Block | Recommend | Approve | Approve |
| Third change within seven days | Freeze | Review | Approve exception | No routine action |
| New listing inside seven days | Monitor only | Set price | Approve exception | No routine action |
Require an audit log for every action:
- Old price
- New price
- Date and time
- Data sources
- Comp set
- Confidence score
- Projected gross
- Rule triggered
- Approver
- Reason for approval or rejection
Do not let managers approve from a notification showing only the new price. The approval screen needs the cost, gross and market evidence.
Owner Math: What 12 Errors Cost
Assume 12 pricing errors each month with a $1,200 average gross loss:
12 × $1,200 = $14,400 per month
Annual loss:
$14,400 × 12 = $172,800 per year
After the controls, the store records one monthly error:
1 × $1,200 = $1,200 per month
Estimated monthly reduction:
$14,400 − $1,200 = $13,200
Estimated annual reduction:
$13,200 × 12 = $158,400
The controls reduced 11 errors in the operating example. A conservative policy target is 10 prevented errors each month.
Before activating automatic pricing, verify six settings:
- Maximum percentage and dollar reduction
- Minimum model confidence and valid-comp count
- Vehicle-history block
- Total-cost gross floor
- Seven-day frequency cap
- New-listing freeze
An AI agent should work inside those limits. Pricing authority outside the limits belongs to a named dealership manager.
Related Reading
7 AI Call-Scoring Rules That Save BDC Managers 8 Hours a Week
7 Used-EV Inventory Rules That Protect $1,500 in Dealer Gross
Sponsored by Gas.net — powering dealership growth through intelligent data.
Alt text: “Gas.net connects franchise dealers with integrated analytics and marketing tools.”