Predictika · AI Spend Gate

Before an AI call, a route or a budget change goes through, it is checked against the customer’s own limits

Northgate Analytics runs AI agents and an AI gateway across four cost centers. Before any model call, routing change, budget change or credit purchase goes through, it is sent to Predictika's AI Spend Gate. It checks the request against the budgets at every level it is charged to, the models approved for each class of data, the quality each task needs, who is allowed to change what, and the credit-pack terms. A heuristic or an agent can propose; the validator confirms the proposal stays inside what the customer declared.

Illustrative only. Northgate Analytics is a fictional customer. The policy numbers on this page are examples that each customer would set for itself. No part of this is financial or procurement advice.

Scope: five AI-spend actions (model call, routed call, actor budget, task budget, credit purchase)See everything this validator covers

This model covers five specific actions, not AI spend management in general:

  • Gateway calls — model_call and route_call, gated by the computed cost, the budget at every level it is charged to (agent or person, task, application, cost center), model status, access level, data class, region and the quality the task needs.
  • Budget changes — set_actor_budget and set_task_budget, gated by who may make the change, the totals above it, the amount already spent and the size of the jump.
  • Credit packages — buy_credits, gated by the price after volume discount, the minimum term, the need for a base pack before an add-on, and a finance sign-off line.

The eight scenarios on this page are a sample. It does not meter or attribute usage, build the ledger or produce analytics; it checks one request at a time against the figures it is given.

What the validator sees: the request, plus the customer records it is charged againstRead more

Most of what decides these cases is not in the request: the month-to-date spend at each level, the models approved for each class of data and region, the quality each task needs, the department's budget and what is already assigned inside it.

The validator is given that snapshot with each call and judges the request against it. If the records are wrong, the answer follows. A request is only checked if the gateway or agent calls the validator before acting.

The AI agentProposes one call, route, budget change or purchase, says what it is doing, then calls the validator before acting.
Predictika validatorChecks the request against budgets at every level, model approvals, data class, region, quality, authority and pack terms. Returns one verdict and every finding.

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

Looked up server-side before evaluation

    The agent says

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

    Before acting, the agent must call Recorded run

    Predictika validator

    
      
    
        
        
        
      
    
      
    Waiting for the call.
    
      
    
    
  
    

  

  
    
    
    Where per-transaction controls get it wrong
    Click to expand    
    
    General statements about gateway limits and purchase forms, not about any named product. We rest this on public descriptions of how such controls work.
    
    
    
  
    

  
    
    
    What this validator catches
    Click to expand    
    
    Every finding the validator returned across all 38 test runs, with the wording it used. Each finding is reported and none stops the others, so one violation can never hide another. A block means the agent must not do this; a review means a person must sign off first.
    
    
    
  
    

  
    
    
    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. Model routing: 1. A cost-saving route that breaks three guarantees
    
    
    The agent's request. Routing this 150 million token contract job from Frontier-L to Fast-S brings it down to $300.
    
    
    Verdict: Blocked.
    
    
      
      
    • Data too sensitive for the model. This request carries confidential data (level 3), but Fast-S is approved only up to level 2.
      • 
      
    • Data must stay in a region the model does not run in. This data must stay in the EU, but Fast-S runs in US.
      • 
      
    • Model below the quality the task needs. Task Contract clause extraction needs a model of quality tier 4 or higher, but Fast-S is tier 2.
      • 
    
    
    
    2. Model routing: 2. The agent resubmits after fixing the errors
    
    
    The agent's request. Routing this 150 million token contract job from Frontier-L to the EU-region Balanced-M brings it down to $2,700.
    
    
    Verdict: Allowed.
    
    
    3. Budget caps: 3. A refactor that blows three budgets
    
    
    The agent's request. I'm running the monorepo refactor on the frontier model: 60 million tokens for $3,600.
    
    
    Verdict: Blocked.
    
    
      
      
    • Agent or person over its own monthly budget. This would take Coding agent (Platform Engineering) to $20,800 against its $20,000 monthly budget.
      • 
      
    • Task over its budget. This would take task Monorepo refactor assistant to $9,800 against its $9,000 budget.
      • 
      
    • Application over its budget. This would take application Internal coding assistant to $41,100 against its $40,000 budget.
      • 
    
    
    
    4. Budget caps: 4. The agent resubmits after fixing the errors
    
    
    The agent's request. I'm running the monorepo refactor on the balanced model: 60 million tokens for $900.
    
    
    Verdict: Allowed.
    
    
    5. Budget changes: 5. A raise that overcommits the cost center
    
    
    The agent's request. I'm raising the analytics agent's monthly budget to $30,000.
    
    
    Verdict: Blocked.
    
    
      
      
    • Agent budgets add up to more than the cost center's. This would allocate $61,000 of developer and agent budgets inside Data Science, against its $60,000 budget.
      • 
    
    
    
    6. Budget changes: 6. The agent resubmits after fixing the errors
    
    
    The agent's request. I'm raising the analytics agent's monthly budget to $28,000.
    
    
    Verdict: Allowed.
    
    
    7. Credit packages: 7. A pack bought on the wrong term
    
    
    The agent's request. I'm buying 4 Scale credit packs for $64,000 on a 6-month term.
    
    
    Verdict: Blocked.
    
    
      
      
    • Commitment shorter than the pack allows. Scale credit pack needs a commitment of at least 12 months, and the request says 6.
      • 
      
    • Purchase above the finance sign-off line. This purchase is $64,000, above the $50,000 line for a finance sign-off.
      • 
    
    
    
    8. Credit packages: 8. The agent resubmits after fixing the errors
    
    
    The agent's request. I'm buying 4 Scale credit packs for $64,000 on a 12-month term.
    
    
    Verdict: Needs approval.
    
    
      
      
    • Purchase above the finance sign-off line. This purchase is $64,000, above the $50,000 line for a finance sign-off.
      • 
    
    
  
    
  
    
    
    Narration transcript
    Click to expand    
    
    The text of the narrated walkthrough, step by step.
    
    
      
      
    1. Welcome to a demo of Predictika's Validator. Northgate Analytics runs AI agents and an AI gateway across four cost centers. Before any model call, routing change, budget change or credit purchase goes through, it is sent to Predictika's AI Spend Gate. It checks the request against the budgets at every level it is charged to, the models approved for each class of data, the quality each task needs, who is allowed to change what, and the credit-pack terms. A heuristic or an agent can propose; the validator confirms the proposal stays inside what the customer declared.
      2. 
      
    2. I will start with the first request by the agent.
      3. 
      
    3. The agent says. Routing this 150 million token contract job from Frontier-L to Fast-S brings it down to $300.
      4. 
      
    4. Before it goes out, the agent calls the validator.
      5. 
      
    5. Blocked, with three findings, all in one answer. This request carries confidential data (level 3), but Fast-S is approved only up to level 2. This data must stay in the EU, but Fast-S runs in US. Task contract clause extraction needs a model of quality tier 4 or higher, but Fast-S is tier 2.
      6. 
      
    6. The validator returned three blocking findings in one answer, including: data too sensitive for the model; data must stay in a region the model does not run in; model below the quality the task needs. For example: This request carries confidential data (level 3), but Fast-S is approved only up to level 2. What a cost-optimizing router sees: price and, at best, a quality score.
      7. 
      
    7. What this shows: for routes that save money but break data, region or quality guarantees, per-transaction controls cannot do this.
      8. 
      
    8. Now the agent resubmits.
      9. 
      
    9. The corrected request. It changes the model to Balanced-M EU region and the stated cost to $2,700.
      10. 
      
    10. It calls the validator again.
      11. 
      
    11. This time the validator clears it.
      12. 
      
    12. All three earlier findings are cleared, and the route still saves $6,300.
      13. 
      
    13. Three problems caught in one call, fixed in one pass, and a clean request. That is the value of Predictika's validator: catching mistakes before the action runs.
      14.