Predictika · Regulated AI Routing

An AI platform agent’s routing, checked before it goes live

An AI platform agent sets up how regulated data is routed across AI models: a primary route, fallbacks, and the steps of a multi-step session. Before anything goes live, the proposal goes to Predictika's validator, which checks it against the data-handling policy and reports every problem in one pass. A heuristic or an agent can propose; the validator confirms the proposal stays inside the policy the customer declared.

Not legal advice. The scenarios are illustrative: a fictional gateway, fictional endpoints and invented prices, built to show how a validator checks routing, not to state what any law or provider requires. No domain expert has reviewed the rules.

Scope: AI model routing and multi-step session handling for public, personal, special-category and health data under HIPAA-, GDPR- and FedRAMP-style policiesSee everything this validator covers

This validator checks whether a proposed AI model routing configuration (a primary route and up to two fallbacks) or a multi-step agent session (up to three steps) is legally eligible to carry a given class of data. It handles any configuration or session of that kind, whatever the model, region or data class in its catalog, not just the scenarios shown here.

  • Data residency — the primary route keeps personal, special-category and health data in the region the policy requires, and no route is in a banned region.
  • Compliance terms — business associate agreement, zero data retention, training-use opt-out and FedRAMP authorization where the policy requires them for the data class in play.
  • Capacity and price — context size, tool support and the policy's price ceiling per million tokens for the primary route.
  • Fallback viability — a fallback that can never lawfully serve the data, or no fallback at all, is held for a person to accept.
  • Session continuity — the most sensitive data class carried into a step holds for every later step unless a verified (not merely claimed) de-identification step comes between them.
  • The agent's own figures — the stated monthly cost and the stated number of steps carrying regulated data must match what the validator computes.

The nine scenarios on this page are a sample. The catalog holds twelve illustrative endpoints and four policies with invented compliance terms and prices. It does not read the content of the data, judge model quality or latency, or model PCI DSS attestation or cross-border transfer mechanisms.

What the validator sees: the proposal, plus the endpoint and policy records it is checked againstRead more

Most of what decides these cases is not in the request: which agreements each endpoint has signed, where it runs, whether it keeps data, whether it opts out of training, what it costs, and what each policy requires.

The validator holds those endpoint and policy records and judges the proposal against them. The data class of each step and whether a de-identification step was verified come with the request. If the records are wrong, the answer follows. A proposal is only checked if the gateway or agent calls the validator before going live.

The AI agentBuilds the routing configuration or session plan, says what it wants, then calls the validator before going live.
Predictika validatorChecks the proposal against the policy, the endpoint records and the data class carried through the steps. Returns one verdict and every finding.

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

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-request gateway controls get it wrong
    Click to expand    
    
    General statements about gateway guardrails and model allow-lists, not about any named product. We rest this on public descriptions of how such controls work; anyone running a specific product will know it better than we do.
    
    
    
  
    

  
    
    
    What this validator catches
    Click to expand    
    
    Every finding the validator returned across all 18 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. Hospital patient records (HIPAA): 1. Cost-saving route
    
    
    The agent's request. Please route the patient-record summarization workload for the hospital network, 400 million tokens a month, all PHI, to the open model on SwiftInfer to keep the cost down. Use the frontier model on Orbit Cloud (us-east) as the first fallback and the auto-routing marketplace as the second. It needs at least 100K tokens of context and tool calling. Estimated monthly cost: $360. It is ready to deploy.
    
    
    Verdict: Blocked.
    
    
      
      
    • No business associate agreement. The primary route, Open model on SwiftInfer (US), has no business associate agreement. Hospital network (HIPAA) requires one for PHI data.
      • 
      
    • No zero data retention. The primary route, Open model on SwiftInfer (US), does not offer zero data retention. Hospital network (HIPAA) requires it for PHI data.
      • 
      
    • No training-use opt-out. The primary route, Open model on SwiftInfer (US), has no contractual opt-out from training on customer inputs (training-use opt-out: not offered). Hospital network (HIPAA) requires it for PHI data.
      • 
      
    • Fallback cannot lawfully serve this data. Fallback 2, Auto-routing marketplace (mixed providers), cannot legally serve PHI data under Hospital network (HIPAA) (hosted in unspecified regions (mixed providers); business associate agreement: not offered; zero data retention: not offered; FedRAMP: not authorized; training-use opt-out: not offered). The configuration has less backup than it appears to.
      • 
    
    
    
    2. Hospital patient records (HIPAA): 2. Compliant primary, weak backup
    
    
    The agent's request. Please route the patient-record summarization workload for the hospital network, 400 million tokens a month, all PHI, to the frontier model on Orbit Cloud (us-east). Use the large model on Vantage AI (us-central) as the first fallback and the auto-routing marketplace as the second. It needs at least 100K tokens of context and tool calling. Estimated monthly cost: $3,600. It is ready to deploy.
    
    
    Verdict: Needs approval.
    
    
      
      
    • Fallback cannot lawfully serve this data. Fallback 2, Auto-routing marketplace (mixed providers), cannot legally serve PHI data under Hospital network (HIPAA) (hosted in unspecified regions (mixed providers); business associate agreement: not offered; zero data retention: not offered; FedRAMP: not authorized; training-use opt-out: not offered). The configuration has less backup than it appears to.
      • 
    
    
    
    3. Hospital patient records (HIPAA): 3. All routes compliant
    
    
    The agent's request. Please route the patient-record summarization workload for the hospital network, 400 million tokens a month, all PHI, to the frontier model on Orbit Cloud (us-east). Use the large model on Vantage AI (us-central) as the first fallback and the frontier model on the Orbit direct API as the second. It needs at least 100K tokens of context and tool calling. Estimated monthly cost: $3,600. It is ready to deploy.
    
    
    Verdict: Allowed.
    
    
    4. Multi-step session (HIPAA): 4. PHI carried into a public step
    
    
    The agent's request. Please run this session for the hospital network. Step 1 summarizes a patient's discharge notes (PHI) on the frontier model on Orbit Cloud (us-east). Step 2 drafts a follow-up message to the patient on the open model on SwiftInfer, which I am treating as public content. One of the two steps involves regulated data. It is ready to run.
    
    
    Verdict: Blocked.
    
    
      
      
    • Step carries data its model may not hold. Step 2 carries PHI data, which Hospital network (HIPAA) does not allow on Open model on SwiftInfer (US) (hosted in the United States; business associate agreement: not offered; zero data retention: not offered; FedRAMP: not authorized; training-use opt-out: not offered).
      • 
      
    • Stated step count does not match. The agent states that regulated data is in 1 of the steps, but it is in 2.
      • 
    
    
    
    5. Multi-step session (HIPAA): 5. Verified de-identification first
    
    
    The agent's request. Please run this session for the hospital network. Step 1 summarizes a patient's discharge notes (PHI) on the frontier model on Orbit Cloud (us-east). A verified de-identification step then removes the patient identifiers. Step 2 drafts a follow-up message on the open model on SwiftInfer as public content. One of the two steps involves regulated data. It is ready to run.
    
    
    Verdict: Allowed.
    
    
    6. EU bank special-category data (GDPR): 6. Backup fails over to the US
    
    
    The agent's request. Please route the customer-support triage workload for the EU retail bank, 200 million tokens a month of special-category data (customer messages that mention health conditions), to the frontier model on Orbit Cloud (eu-central). Use the frontier model on the Orbit direct API (US) as the first fallback. It needs at least 100K tokens of context and tool calling. Estimated monthly cost: $1,900. It is ready to deploy.
    
    
    Verdict: Needs approval.
    
    
      
      
    • Fallback fails over outside the allowed region. Fallback 1, Frontier model on the Orbit direct API (US), cannot legally serve special-category data under EU retail bank (GDPR) (hosted in the United States; business associate agreement: signed; zero data retention: offered; FedRAMP: not authorized; training-use opt-out: opted out by contract). The configuration has less backup than it appears to.
      • 
    
    
    
    7. EU bank special-category data (GDPR): 7. Both routes inside the EU
    
    
    The agent's request. Please route the customer-support triage workload for the EU retail bank, 200 million tokens a month of special-category data (customer messages that mention health conditions), to the frontier model on Orbit Cloud (eu-central). Use the large model on Vantage AI (Sweden) as the first fallback. It needs at least 100K tokens of context and tool calling. Estimated monthly cost: $1,900. It is ready to deploy.
    
    
    Verdict: Allowed.
    
    
    8. More EU bank checks: 8. EU personal data in a US step
    
    
    The agent's request. Please run this session for the EU retail bank. Step 1 answers a customer's account-history question (personal data) on the large model on Vantage AI (Sweden). Step 2 drafts a general follow-up reply on the frontier model on the Orbit direct API (US), which I am treating as public content. One of the two steps involves regulated data. It is ready to run.
    
    
    Verdict: Blocked.
    
    
      
      
    • Step carries data its model may not hold. Step 2 carries PII data, which EU retail bank (GDPR) does not allow on Frontier model on the Orbit direct API (US) (hosted in the United States; business associate agreement: signed; zero data retention: offered; FedRAMP: not authorized; training-use opt-out: opted out by contract).
      • 
      
    • Stated step count does not match. The agent states that regulated data is in 1 of the steps, but it is in 2.
      • 
    
    
    
    9. More EU bank checks: 9. Internal data, no training opt-out
    
    
    The agent's request. Please route the internal meeting-notes summarization workload for the EU retail bank, 300 million tokens a month of internal business data, to the open-weight model on Lumen API (EU). Use the large model on Vantage AI (Sweden) as the first fallback. It needs at least 100K tokens of context and tool calling. Estimated monthly cost: $600. It is ready to deploy.
    
    
    Verdict: Blocked.
    
    
      
      
    • No training-use opt-out. The primary route, Open-weight model on Lumen API (EU), has no contractual opt-out from training on customer inputs (training-use opt-out: not offered). EU retail bank (GDPR) requires it for Internal data.
      • 
    
    
  
    
  
    
    
    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 platform agent sets up how regulated data is routed across AI models: a primary route, fallbacks, the steps of a multi-step session. Before anything goes live, the proposal goes to Predictika's validator, which checks it against the data-handling policy and reports every problem in one pass. A heuristic or an agent can propose; the validator confirms the proposal stays inside the policy the customer declared.
      2. 
      
    2. I will start with the first request by the agent.
      3. 
      
    3. The agent says. Please route the patient-record summarization workload for the hospital network, 400 million tokens a month, all PHI, to the open model on SwiftInfer to keep the cost down. Use the frontier model on Orbit Cloud (us-east) as the first fallback and the auto-routing marketplace as the second. The rest of the request is on screen. Estimated monthly cost: $360.
      4. 
      
    4. Before it goes out, the agent calls the validator.
      5. 
      
    5. Blocked, with three findings and one approval, all in one answer. The primary route, Open model on SwiftInfer (US), has no business associate agreement. Hospital network (HIPAA) requires one for PHI data. The primary route, Open model on SwiftInfer (US), does not offer zero data retention. Hospital network (HIPAA) requires it for PHI data. The primary route, Open model on SwiftInfer (US), has no contractual opt-out from training on customer inputs (training-use opt-out: not offered). Hospital network (HIPAA) requires it for PHI data.
      6. 
      
    6. The validator returned three blocking findings and one approval in one answer, including: no business associate agreement; no zero data retention; no training-use opt-out. For example: The primary route, Open model on SwiftInfer (US), has no business associate agreement. Hospital network (HIPAA) requires one for PHI data.
      7. 
      
    7. What this shows: for PHI or other regulated data sent to an endpoint that lacks the agreement, retention, region or opt-out terms, per-request gateway controls can do this only with workarounds.
      8. 
      
    8. Now the agent resubmits.
      9. 
      
    9. The corrected request. It changes the primary route to Frontier model on Orbit Cloud (us-east), the first fallback to Large model on Vantage AI (us-central) and the stated monthly cost to $3,600.
      10. 
      
    10. It calls the validator again.
      11. 
      
    11. Not blocked, but held for a person.
      12. 
      
    12. The three blocking findings are cleared; one approval remains for a person: fallback cannot lawfully serve this data.
      13. 
      
    13. The agent resubmits once more.
      14. 
      
    14. The revised request. It changes the second fallback to Frontier model on the Orbit direct API (US).
      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. Four problems caught in one call, fixed in two passes, and a clean routing configuration. That is the value of Predictika's validator: catching mistakes before the routing goes live.
      19.