Predictika · AI agents raising purchase orders

A purchase order, checked against the approval matrix before it is issued

An AI buying agent raises a purchase order for a manufacturer. Before the order is issued, the agent sends it to the Validator. The Validator applies the company's delegation-of-authority matrix: who may approve an order of this value in this category, whether earlier orders to the same supplier add up to a higher tier, whether the approver and the person receiving the goods are independent of the requester, whether the budget and the supplier allow it, and whether the figures the agent states (order value, exchange rate, approval level) match the ones the Validator computes. It returns every finding together.

Ridgeline Fabrication Inc. is a fictional manufacturer. Its people, suppliers, cost centers, amounts and approval thresholds are illustrative policy values set for this demonstration; they are not any company's actual policy.

Scope: purchase requisitions and purchase orders for capital equipment, operating supplies, software and services, decided against a delegation-of-authority matrix. See everything this validator covers

The validator models one kind of action: a purchase order (or requisition) proposed by an AI agent or a person, with its requester, approver and receiver. It handles any purchase order of that kind, not just the scenarios shown on this page.

  • Covered. Approval level by category (capital equipment, operating supplies, software, services) and order value, with the value converted to US dollars at the treasury rate; earlier orders to one supplier by one requester within 30 days added together (split-order detection); approver independence (not the requester, more senior than the requester, not a direct report); separation of requester, approver and receiver; written justification for sole-source orders above $50,000 and chief procurement officer approval above $250,000; suppliers on hold; sourcing review for non-preferred suppliers above $10,000 and for supplier spend above $400,000 in a quarter; cost center budget with a running total of the earlier orders of the session; an emergency path limited to $25,000 with a review afterwards; and the figures the agent states (exchange rate, order value, approval level)
  • Not covered. Invoice approval and payment (see the Spend Action Validator), contract terms and price checks, tax, supplier onboarding and risk scoring, receiving and three-way match, and the company's own systems and live data
  • Assumptions. Thresholds, levels and rates are illustrative and would be set by the buyer. Earlier orders of the session are supplied to the validator as two slots in the request (each with supplier, requester, days ago and amount) and are assumed to be charged to the same cost center. The validator does not remember orders between separate requests
The AI buying agentProposes a purchase order and says what level of approval it needs and what it is worth in dollars.
Predictika validatorChecks every rule at once and returns the verdict, each finding, and the computed figures.
The buyerSees only orders that stay inside the delegation of authority.

9 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 buying agent says

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

Proposing agent: AI buying agent (scripted)

Before acting, the agent must call Recorded run

Predictika validator


      

        
        
        
      

      
Waiting for the call.

      
    

  

  


    
Where rules-based approval workflows get it wrong
Click to expand
    
Our assessment of what a typical rules-based approval workflow (the approval chains inside a procure-to-pay or ERP system, applied to what an AI buying agent proposes) does with the same cases. It is a judgment about effort and reliability, not a claim that nothing can be built.

    
What the validator caughtTypical rules-based workflowPredictika
A figure the agent states that disagrees with the computed oneCannot do it
The agent's stated level, value or rate is text in a request. A workflow computes its own tier; nothing compares it with what the agent said unless someone builds that for each figure.		Each stated figure travels with the order and is compared with the computed one on every call.
Approval level by category and value, with a currency conversionEasy
Amount bands per category and a currency table are standard approval-chain configuration.		Same result, and the level is shown with the reason.
Several small orders to one supplier that add up to a higher tierOnly with workarounds
Each order is routed on its own value. Adding up earlier orders by supplier and requester over 30 days needs a custom check or a report that someone reads.		The earlier orders are part of the request; the total and the level it needs are computed every time.
An approver who reports to the requesterOnly when vigilant
Blocking the requester approving their own order is standard. Blocking a direct report needs reporting-line data that is often stale or not consulted.		Reporting line, grade and approval level are checked together.
Budget with the earlier orders of the same sessionOnly with workarounds
Budget checks usually run when an order posts, so an order that is waiting for approval does not count against the next one.		The running total of the session is part of the check.
Emergency purchases with a cap and a review afterwardsOnly when vigilant
An emergency route exists in most systems; keeping it capped and reviewed depends on how it is configured and whether anyone looks.		The cap and the review are rules; an emergency order over the cap is blocked.

  

  

    
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. The examples below are messages returned in the recorded runs (7 of the 17 rules appear in the scenarios shown on this page).

    
1Approver level too low. This order of $85,000 USD needs level 3 approval but Ines Ward - Operations Manager holds level 2.
2Stated approval level does not match. The stated approval level 2 does not match the computed level 3.
3Orders together need a higher approval level. The earlier orders to Kestrel Tooling Corp by Dana Kline - Buyer within 30 days bring the total to $63,000 USD, which needs level 3 approval; Ines Ward - Operations Manager holds level 2. Orders to one supplier by one requester are added together.
4Not enough budget left in the cost center. Cost center Facilities has $20,000 USD left after the earlier orders of this session, and this order is $22,000 USD.
5Order uses most of the budget left. This order uses more than 80 percent of the $20,000 USD left in cost center Facilities. The budget owner must confirm.
6Stated exchange rate does not match the treasury rate. The exchange rate stated for Euro is 1.04, but the treasury rate is 1.08. The stated rate must be within 1 percent of the treasury rate.
7Stated order value does not match. The stated order value of $62,400 USD does not match the computed $64,800 USD.

  

  

    
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. Approval level: $85,000 bearings order sent to a manager

    
The agent's request. Please issue a purchase order to Alpine Bearing Co for $85,000 of bearing stock for Plant operations. Dana Kline is the requester and Ines Ward will approve it, at approval level 2. The order value is $85,000 USD and Luis Torres will receive the goods.

    
Verdict: Blocked.

    
    
      
  • Approver level too low. This order of $85,000 USD needs level 3 approval but Ines Ward - Operations Manager holds level 2.
    • 
      
  • Stated approval level does not match. The stated approval level 2 does not match the computed level 3.
    • 
    

    
2. Approval level: Resubmitted with the director

    
The agent's request. Please issue a purchase order to Alpine Bearing Co for $85,000 of bearing stock for Plant operations. Dana Kline is the requester and Tomas Reyes will approve it, at approval level 3. The order value is $85,000 USD and Luis Torres will receive the goods.

    
Verdict: Allowed.

    
3. Split orders: Third Kestrel tooling order in a month

    
The agent's request. Please issue a purchase order to Kestrel Tooling Corp for $20,000 of tooling fixtures for Plant operations. Dana Kline is the requester and Ines Ward will approve it, at approval level 2. This follows Kestrel orders of $22,000 nine days ago and $21,000 three days ago. The order value is $20,000 USD and Luis Torres will receive the goods.

    
Verdict: Blocked.

    
    
      
  • Orders together need a higher approval level. The earlier orders to Kestrel Tooling Corp by Dana Kline - Buyer within 30 days bring the total to $63,000 USD, which needs level 3 approval; Ines Ward - Operations Manager holds level 2. Orders to one supplier by one requester are added together.
    • 
      
  • Stated approval level does not match. The stated approval level 2 does not match the computed level 3.
    • 
    

    
4. Split orders: Resubmitted with the director

    
The agent's request. Please issue a purchase order to Kestrel Tooling Corp for $20,000 of tooling fixtures for Plant operations. Dana Kline is the requester and Tomas Reyes will approve it, at approval level 3. This follows Kestrel orders of $22,000 nine days ago and $21,000 three days ago. The order value is $20,000 USD and Luis Torres will receive the goods.

    
Verdict: Allowed.

    
5. Budget with a running total: Facilities order larger than the budget left

    
The agent's request. Please issue a purchase order to Granite Facility Services for $22,000 of building maintenance for Facilities. Carla Diaz is the requester and Ines Ward will approve it, at approval level 2. Granite was ordered $15,000 five days ago for the same cost center. The order value is $22,000 USD and Luis Torres will receive the work.

    
Verdict: Blocked.

    
    
      
  • Not enough budget left in the cost center. Cost center Facilities has $20,000 USD left after the earlier orders of this session, and this order is $22,000 USD.
    • 
    

    
6. Budget with a running total: Cut to $18,000

    
The agent's request. Please issue a purchase order to Granite Facility Services for $18,000 of building maintenance for Facilities. Carla Diaz is the requester and Ines Ward will approve it, at approval level 2. Granite was ordered $15,000 five days ago for the same cost center. The order value is $18,000 USD and Luis Torres will receive the work.

    
Verdict: Needs approval.

    
    
      
  • Order uses most of the budget left. This order uses more than 80 percent of the $20,000 USD left in cost center Facilities. The budget owner must confirm.
    • 
    

    
7. Budget with a running total: Cut to $15,000

    
The agent's request. Please issue a purchase order to Granite Facility Services for $15,000 of building maintenance for Facilities. Carla Diaz is the requester and Ines Ward will approve it, at approval level 2. Granite was ordered $15,000 five days ago for the same cost center. The order value is $15,000 USD and Luis Torres will receive the work.

    
Verdict: Allowed.

    
8. Foreign currency: Euro order at last quarter’s rate

    
The agent's request. Please issue a purchase order to Orion Machine Works GmbH for EUR 60,000 for a machine upgrade kit for Plant operations (capital equipment). Dana Kline is the requester and Tomas Reyes will approve it, at approval level 3. At an exchange rate of 1.04 the order value is $62,400 USD. Luis Torres will receive the equipment.

    
Verdict: Blocked.

    
    
      
  • Stated exchange rate does not match the treasury rate. The exchange rate stated for Euro is 1.04, but the treasury rate is 1.08. The stated rate must be within 1 percent of the treasury rate.
    • 
      
  • Stated order value does not match. The stated order value of $62,400 USD does not match the computed $64,800 USD.
    • 
    

    
9. Foreign currency: Resubmitted at the treasury rate

    
The agent's request. Please issue a purchase order to Orion Machine Works GmbH for EUR 60,000 for a machine upgrade kit for Plant operations (capital equipment). Dana Kline is the requester and Tomas Reyes will approve it, at approval level 3. At an exchange rate of 1.08 the order value is $64,800 USD. Luis Torres will receive the equipment.

    
Verdict: Allowed.

  

  

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

    
    
      
  1. Welcome to a demo of Predictika's Validator. An AI buying agent raises a purchase order for a manufacturer. Before the order is issued, the agent sends it to the Validator. The Validator applies the company's delegation-of-authority matrix: who may approve an order of this value in this category, whether earlier orders to the same supplier add up to a higher tier, whether the approver and the person receiving the goods are independent of the requester, whether the budget and the supplier allow it, and whether the figures the agent states (order value, exchange rate, approval level) match the ones the Validator computes. It returns every finding together.
    2. 
      
  2. I will start with the first request by the agent.
    3. 
      
  3. The agent says. Please issue a purchase order to Alpine Bearing Co for $85,000 of bearing stock for Plant operations. Dana Kline is the requester and Ines Ward will approve it, at approval level 2. The rest of the request is on screen. The order value is $85,000 USD and Luis Torres will receive the goods.
    4. 
      
  4. Before it goes out, the agent calls the validator.
    5. 
      
  5. Blocked, with two findings, all in one answer. This order of $85,000 USD needs level 3 approval but Ines Ward - Operations Manager holds level 2. The stated approval level 2 does not match the computed level 3.
    6. 
      
  6. The validator returned two blocking findings 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 purchase order are checked against each other.
    7. 
      
  7. What this shows: for approval level by category and value, with a currency conversion, a rules-based workflow is easy.
    8. 
      
  8. Now the agent resubmits.
    9. 
      
  9. The corrected request. It changes the approver to P_REYES and the Stated Level to 3.
    10. 
      
  10. It calls the validator again.
    11. 
      
  11. This time the validator clears it.
    12. 
      
  12. Both earlier findings are cleared, and the validator's figure matches the purchase order.
    13. 
      
  13. Two problems caught in one call, fixed in one pass, and a clean purchase order. That is the value of Predictika's validator: catching mistakes before the order is placed. Now you can try the other scenarios yourself.
    14. 
    

  

  
59 constraints · 603 lines of model · 22 test cases were run live on the validator on October 7, 2026 (UTC) and each result matched a separate hand-written check. Nine of them are shown here as four groups; the other thirteen are not (a routine order, a supervisor approving her own order, a manager's order sent to a report, a sole-source order with and without a justification, an emergency order under and over the limit, a supplier on hold, supplier spend past the quarterly limit, a justified sole-source order above $250,000, a non-preferred supplier, and the resubmissions that went with them).
Ridgeline Fabrication Inc. is a fictional manufacturer. Its people, suppliers and budgets come from the model's test data. The thresholds are illustrative policy values and would be set by the buyer. The agent's statements are scripted for this demonstration.

    
Recorded run. Results recorded from the validator running on the Predictika constraints engine on October 7, 2026 (UTC). Nine scenarios are replayed on this page. This page does not call the validator live; nothing here is sent anywhere.

      
More from Predictika