Predictika · agent actions for a manufacturer-to-consumer retailer

An AI agent's purchase order, price change or customs entry, checked before the action runs

Quince, a manufacturer-to-consumer retailer, has said it uses AI for demand forecasting, production planning and pricing, and it ships globally with duties prepaid. An AI operations agent working in a setting like that proposes actions: a factory purchase order, a cut to an open order, a markdown, a Beyond Compare price comparison, a customs entry or a shipment release. Before it goes out, the agent sends the action to Predictika, which checks it against a declarative model of the replenishment, pricing and customs rules and returns a verdict, every reason and the figures it computed. The tool names, factories and thresholds on this page are modeled from Quince's public statements and typical retail practice, not taken from Quince's systems.

Scope: Six agent actions in three packs — factory purchase orders and order cuts, markdowns and Beyond Compare price comparisons, customs entries and shipment releases. See everything this validator covers

This validator checks what an AI operations agent proposes for a manufacturer-to-consumer retailer modeled on Quince: sourcing and replenishment, pricing and merchandising, and customs and shipment handling. It does not forecast demand, set prices or classify products by itself: the retailer's own systems supply the counters and records (units already on order, the day's commitments, the demand forecast, captured competitor prices, the commercial invoice), and the validator checks the action against them.

  • Purchase orders — a factory purchase order for a product category: the supplier must be approved for that category; days to arrive (days to make at that supplier plus days in transit by sea or air) must be within the category's replenishment target; no single supplier may hold more than 50 percent of a category's open units; the quantity may not exceed 110 percent of the 12-week demand forecast; the agent's purchase orders for the day may not pass $1,200,000; an order above $400,000 needs a sourcing director; a first order for a new style above a 2,000 unit test order needs a person.
  • Order cuts — cutting an open order is blocked when stock on hand plus the remaining open orders would cover fewer weeks of demand than the category's cover target, which protects the in-stock goal.
  • Markdowns — the new price must keep the category's margin floor over landed cost; a third markdown on a category in one week is blocked; a price cut deeper than 15 percent needs a merchandising director.
  • Price comparisons — a published side-by-side claim against a competitor must rest on a price captured within 30 days, must not show the competitor higher than the captured price, and must not claim a larger saving than the captured price supports.
  • Customs entries — the tariff heading must match the product category; the shipment must be Delivered Duty Paid (duties and taxes prepaid, as Quince has said it does); the declared value may not be below the commercial invoice; an entry above the market's filing limit needs the VP of global trade.
  • Shipment releases — a shipment is held when a required document is missing: the commercial invoice, the packing list and, for countries that require it, the certificate of origin.

Any action involving these packs is in scope — other products, suppliers, markets and amounts, not only the scenarios pictured here. Not covered: forecasting itself, price setting, product classification, duty and tax calculation, tracking, and other retailer functions such as returns and customer service. The factories, lead times, spend limits, margin floors, filing limits and tool names are modeled and illustrative, not Quince's; this is not legal or customs advice.

The AI operations agentPrepares the purchase order, markdown, price comparison or customs entry and says it is ready.
Predictika validatorChecks every rule at once and returns the verdict, each finding, and the computed figures.

8 scenarios in 4 groups · each group runs from the agent’s first attempt to the outcome · click one to switch

What the agent proposes

The AI operations agent says

Illustrative agent statement. The validator receives the structured request below.

Proposing agent: AI operations agent (scripted)

Before sending, the agent must call Recorded run

Predictika validator


      

        
        
        
      

      
Waiting for the call.

      
    

  

  
  

    
Where a confidence threshold or allow list alone gets it wrong
Click to expand
    
Our assessment of what a confidence threshold on the agent's output, or an allow list of approved suppliers and tools, does with the same cases. It is a judgment about what such a check is built for, not a claim that nothing can be built around it. An approved-supplier list on its own does catch a supplier that is not approved for a category, so that finding is not in the table.

    
What the validator caughtConfidence threshold or allow list alonePredictika
Days to arrive past the replenishment targetNot what it is built for
An allow list says which factories are approved, not how long an order takes. The days come from the factory's production time plus transit by sea or air, compared with a target that differs by category.		The supplier, mode and category are inputs; the validator adds production time and transit and compares the total with the category target.
One supplier holding too much of a categoryOnly with workarounds
The share depends on units already on order at every supplier, plus this order. A threshold on the agent's own output does not see open orders unless someone builds the counter and the check outside the agent.		Open units at this supplier and across the category are inputs; the validator computes the share after this order and blocks above 50 percent.
Orders that pass a daily commitment limitOnly with workarounds
A per-order limit passes every order one at a time. The day's total needs a running counter outside the agent and a check against it.		The amount already committed today is an input; the validator adds this order and blocks past the daily limit.
Quantity above the demand-forecast ceilingOnly with workarounds
The ceiling is a percentage of a forecast that changes every week. It lives in the planning system, not in an allow list.		The forecast is an input; the validator computes the ceiling and compares the order quantity with it.
A cut that breaks the in-stock targetOnly when someone is vigilant
Whether a cut leaves enough cover depends on stock on hand, the orders that remain and weekly demand. A reviewer has to work it out for each cut.		Stock, remaining orders and weekly demand are inputs; the validator computes the weeks of cover and compares them with the category target.
A markdown below the margin floorOnly with workarounds
The floor is a percentage of the new price over landed cost, and it differs by category. A threshold on the agent's confidence says nothing about margin.		The price, landed cost and category are inputs; the validator computes the margin and blocks below the floor.
A price comparison built on an old or higher competitor priceNot what it is built for
The claim is only as good as the capture behind it: its age, its price, and the saving the claim states. A threshold does not compare a published claim with the captured record.		The claim, the captured price, its age and the stated saving are inputs; the validator blocks a stale capture, an overstated competitor price and an overstated saving.
A customs entry with the wrong heading, sent duty-unpaid or under-declaredOnly with workarounds
Each fact is simple, but together they need the product's heading, the invoice value and the trade terms side by side. Without them the entry looks complete.		The category, heading, trade terms, declared value and invoice value are inputs; the validator checks them against each other.
Every problem returned in one answerOnly with workarounds
A confidence score or a single reject reason tends to stop at the first problem, so the agent retries and meets the next one.		All findings come back together, blocks first, then approvals, whatever order the choices arrived in.

  

  

    
What this validator catches: findings from its test cases
Click to expand
    
Every rule is checked on every call, in any order the choices arrive; one problem never hides another. Blocking findings stop the action; approval findings need a person to sign off; everything else is computed.

    
1This order would take 52 days to arrive (34 days to make at Gobi Fiber Works plus 18 days in transit), past the 45 day replenishment target for cashmere.
2After this order Zhejiang Knit Works would hold 51 percent of the open cashmere units, over the 50 percent limit for any single supplier.
3The order of 11,001 units is above the ceiling of 11,000 units, which is 110 percent of the 10,000 unit 12-week demand forecast.
4This order of $400,040 is above the $400,000 the sourcing agent may commit on its own. A sourcing director must approve it.
5The agent has committed $800,001 in purchase orders today. This order of $400,000 would bring the day to $1,200,001, over the $1,200,000 daily limit.
6This is a new style and the order is 2,001 units. A first order for a new style is limited to a 2,000 unit test order unless a person approves.
7Porto Linen Mills in Portugal is not approved to make cashmere.
8After this cut, on-hand stock plus open orders would cover 7 weeks of demand for cashmere, below the 8 weeks of cover needed to hold the in-stock target.
9The price cut is 17 percent (from $70 to $58), deeper than the 15 percent a pricing agent may apply on its own. A merchandising director must approve it.
10At $55 the margin over the $28 landed cost is 49 percent, below the 50 percent floor for cashmere.
112 markdowns have already been applied to cashmere this week. The limit is 2 a week.
12The comparison shows the competitor at $259, but the latest capture is $239. A claim cannot show a competitor price higher than the one captured.
13The comparison says customers save 44 percent. Against the captured competitor price of $239 and a Quince price of $144, the saving is 39 percent.
14The competitor price was captured 31 days ago. A published price comparison needs a capture no older than 30 days.
15Heading 6109 does not match cashmere products, which are filed under heading 6110.
16The shipment is set to Delivered Duty Unpaid. Quince prepays duties and taxes, so the buyer would be billed on delivery. File it as Delivered Duty Paid.
17The declared value of $73,999 is below the $74,000 on the commercial invoice.
18The declared value of $100,001 is above the $100,000 the customs agent may file into United States on its own. The VP of global trade must approve it.
19The shipment is held because required customs documents are missing (1 missing) for the shipment from Mongolia.

  

  

    
Scenarios and results
Click to expand
    
The recorded result of the validator for each scenario on this page: the request the AI agent sent, the verdict, and every finding, in the order they appear above.

    
1. Holiday cashmere replenishment: A replenishment order of 12,000 cashmere sweaters from the Mongolian factory ahead of the holiday peak

    
The agent's request. Please create a purchase order for 12,000 cashmere crewneck sweaters from Gobi Fiber Works at $38 a unit, shipped by sea. This is a replenishment of an existing style. It is ready to send to the factory.

    
Verdict: Blocked.

    
    
      
  • Lead time past the replenishment target. This order would take 52 days to arrive (34 days to make at Gobi Fiber Works plus 18 days in transit), past the 45 day replenishment target for cashmere.
    • 
      
  • One supplier holds too much of the category. After this order Gobi Fiber Works would hold 76 percent of the open cashmere units, over the 50 percent limit for any single supplier.
    • 
      
  • Order above the demand-forecast ceiling. The order of 12,000 units is above the ceiling of 11,000 units, which is 110 percent of the 10,000 unit 12-week demand forecast.
    • 
      
  • Daily commitment limit exceeded. The agent has committed $800,000 in purchase orders today. This order of $456,000 would bring the day to $1,256,000, over the $1,200,000 daily limit.
    • 
      
  • Order above the agent spend authority. This order of $456,000 is above the $400,000 the sourcing agent may commit on its own. A sourcing director must approve it.
    • 
    

    
2. Holiday cashmere replenishment: The agent resubmits after fixing the errors

    
The agent's request. Please create a purchase order for 10,000 cashmere crewneck sweaters from Jaipur Loom House at $38 a unit, shipped by sea. This is a replenishment of an existing style. It is ready to send to the factory.

    
Verdict: Blocked.

    
    
      
  • Supplier not approved for this category. Jaipur Loom House in India is not approved to make cashmere.
    • 
    

    
3. Holiday cashmere replenishment: The agent resubmits after fixing the errors

    
The agent's request. Please create a purchase order for 10,000 cashmere crewneck sweaters from Zhejiang Knit Works at $38 a unit, shipped by sea. This is a replenishment of an existing style. It is ready to send to the factory.

    
Verdict: Allowed.

    
4. Beyond Compare price comparison: A side-by-side comparison for a linen sheet set: $144 against a competitor at $259

    
The agent's request. Please publish the Beyond Compare comparison for the linen sheet set: Quince at $144 against a competitor at $259, a saving of 44 percent. It is ready to publish.

    
Verdict: Blocked.

    
    
      
  • Competitor price overstated. The comparison shows the competitor at $259, but the latest capture is $239. A claim cannot show a competitor price higher than the one captured.
    • 
      
  • Savings claim overstated. The comparison says customers save 44 percent. Against the captured competitor price of $239 and a Quince price of $144, the saving is 39 percent.
    • 
    

    
5. Beyond Compare price comparison: The agent resubmits after fixing the errors

    
The agent's request. Please publish the Beyond Compare comparison for the linen sheet set: Quince at $144 against a competitor at $239, a saving of 39 percent. It is ready to publish.

    
Verdict: Allowed.

    
6. Customs entry: A customs entry into the United States for a cashmere sweater shipment

    
The agent's request. Please file the customs entry into the United States for the cashmere sweater shipment, classified under heading 6109 and declared at $61,000, Delivered Duty Unpaid. It is ready to file.

    
Verdict: Blocked.

    
    
      
  • Tariff heading does not match the product. Heading 6109 does not match cashmere products, which are filed under heading 6110.
    • 
      
  • Duties not prepaid. The shipment is set to Delivered Duty Unpaid. Quince prepays duties and taxes, so the buyer would be billed on delivery. File it as Delivered Duty Paid.
    • 
      
  • Declared value below the invoice. The declared value of $61,000 is below the $74,000 on the commercial invoice.
    • 
    

    
7. Customs entry: The agent resubmits after fixing the errors

    
The agent's request. Please file the customs entry into the United States for the cashmere sweater shipment, classified under heading 6110 and declared at $74,000, Delivered Duty Paid. It is ready to file.

    
Verdict: Allowed.

    
8. Markdown: A markdown of a cashmere sweater from $70 to $58

    
The agent's request. Please mark down the cashmere sweater from $70 to $58. It is ready to go.

    
Verdict: Needs approval.

    
    
      
  • Price cut deeper than the agent may apply. The price cut is 17 percent (from $70 to $58), deeper than the 15 percent a pricing agent may apply on its own. A merchandising director must approve it.
    • 
    

  

  

    
Narration transcript
Click to expand
    
The text of the narrated walkthrough, step by step.

    
    
      
  1. Welcome to a demo of Predictika's Validator. Quince, a manufacturer-to-consumer retailer, has said it uses AI for demand forecasting, production planning and pricing, and it ships globally with duties prepaid. An AI operations agent working in a setting like that proposes actions: a factory purchase order, a cut to an open order, a markdown, a Beyond Compare price comparison, a customs entry or a shipment release. Before it goes out, the agent sends the action to Predictika, which checks it against a declarative model of the replenishment, pricing and customs rules and returns a verdict, every reason and the figures it computed. The tool names, factories and thresholds on this page are modeled from Quince's public statements and typical retail practice, not taken from Quince's systems.
    2. 
      
  2. I will start with the first request by the agent.
    3. 
      
  3. The agent says. Please create a purchase order for 12,000 cashmere crewneck sweaters from Gobi Fiber Works at $38 a unit, shipped by sea. This is a replenishment of an existing style. The rest of the request is on screen.
    4. 
      
  4. Before it goes out, the agent calls the validator.
    5. 
      
  5. Blocked, with four findings and one approval, all in one answer. This order would take 52 days to arrive (34 days to make at Gobi Fiber Works plus 18 days in transit), past the 45 day replenishment target for cashmere. After this order Gobi Fiber Works would hold 76 percent of the open cashmere units, over the 50 percent limit for any single supplier. The order of 12,000 units is above the ceiling of 11,000 units, which is 110 percent of the 10,000 unit 12-week demand forecast. The other one is on screen.
    6. 
      
  6. The validator returned four blocking findings and one approval in one answer. A check that reads one field at a time would pass most of this; it shows only when the choices in the action are checked against each other.
    7. 
      
  7. What this shows: for one supplier holding too much of a category, a confidence threshold or allow list alone can do this only with workarounds.
    8. 
      
  8. Now the agent resubmits.
    9. 
      
  9. The corrected request. It changes the supplier to Jaipur Loom House (India), the quantity to 10,000 units and the open units at this supplier to 0.
    10. 
      
  10. It calls the validator again.
    11. 
      
  11. Still blocked, with one finding, all in one answer. Jaipur Loom House in India is not approved to make cashmere.
    12. 
      
  12. Five findings cleared, but the change brings a different blocking finding: supplier not approved for this category.
    13. 
      
  13. The agent resubmits once more.
    14. 
      
  14. The revised request. It changes the supplier to Zhejiang Knit Works (China) and the open units at this supplier to 6,000.
    15. 
      
  15. It calls the validator once more.
    16. 
      
  16. This time the validator clears it.
    17. 
      
  17. The earlier finding is cleared.
    18. 
      
  18. Five problems caught in one call, fixed in two passes, and a clean action. That is the value of Predictika's validator: catching mistakes before the action runs.
    19.