Predictika · Basalt Storage Systems delivery proposals

An AI deal desk agent's delivery proposal, checked before it goes out

An AI deal desk agent prepares a delivery proposal for an AI storage order on Basalt Storage Systems' AX3 appliance line. Before it goes out, the agent sends it to Predictika, which checks it against a declarative model of the delivery rules, compares the figures the agent states with the ones it computes, and returns a verdict, every reason, and the figures it computed.

Scope: delivery proposals for Basalt Storage Systems' AX3 all-flash and AX3H hybrid flash/HDD appliances. See everything this validator covers

This validator models one process: a delivery proposal for Basalt Storage Systems' AX3 appliance line, the AX3 all-flash and AX3H hybrid flash/HDD units. It handles any delivery proposal involving these appliances, not just the scenarios shown on this page. It checks:

  • Performance fit — read throughput per GPU for training and inference workloads against what the appliances deliver, usable capacity against the capacity the customer needs, and the certified reference design that large training clusters require.
  • Manufacturing — the build lead time for the delivery quarter, plus SSD supply and build slots in that quarter against what is already committed to other orders.
  • Support — support tiers against regional parts-depot coverage; Premium Onsite promises a one-hour response, which needs a depot in the region.
  • Commercial terms — the per-unit discount floor by order size, and how long a quote can be held while SSD prices are moving; going past the limit needs finance approval.
  • The agent's stated figures — the throughput and usable capacity the agent says the order delivers must match what the validator computes.

It does not cover other Basalt product lines, and it produces no price build-up or line-item quote: it checks the agent's unit price against the floor but does not build a price.

The AI deal desk agentBuilds the delivery proposal from the conversation and says it is ready.
Predictika validatorChecks every rule at once and returns the verdict, each finding, and the computed figures.

7 scenarios in 3 groups · each group runs from the agent’s first attempt to the outcome · click one to switch

What the agent proposes

The AI deal desk agent says

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

Proposing agent: AI deal desk agent (scripted)

Before sending, the agent must call Recorded run

Predictika validator


      

        
        
        
      

      
Waiting for the call.

      
    

  

  
  
  

    
Where rules-based CPQ gets it wrong
Click to expand
    
Our assessment of what a typical rules-based CPQ (product rules, price rules and validation rules, as in Salesforce CPQ) 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 CPQPredictika
Certified reference design for large training clustersOnly with care
The rule depends on the GPU count of the cluster and on the product family, so it has to be written for each model and kept in step as appliances are added.		Cluster size and the appliance family arrive in one request, and the certification rule is checked on every call.
Throughput and capacity the proposal statesOnly with workarounds
The stated figures come from the model and the number of units, so a rule has to recompute them and compare, for each product line.		The stated figures are fields in the request and are compared with the computed ones on every call.
SSD supply and build slots already committed in the quarterOnly with workarounds
The limit depends on orders already committed in the delivery quarter, which is data outside the proposal and changes as orders are booked.		Committed orders are part of the model, so the totals for the delivery quarter are computed on every call.
Support tier against regional parts depotsOnly with care
Each support tier promises a response time that only some regions can meet, so the tier and the region have to be checked against each other.		Support tier and region are in one request, and the depot rule is checked together with the other rules.
Per-unit price floor by order sizeOnly with care
The maximum discount changes with the number of units, so a rules-based tool needs a price table for each order size band.		The floor is computed from the order size on every call, and the finding states the floor it applied.
Every problem returned in one answerOnly with workarounds
Rules tend to fire as a quote is saved or a step is completed, so a person often fixes one error and meets the next.		All findings come back together, 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 delivery proposal; approval findings need a person to sign off; everything else is computed.

    
1Not enough throughput. 4,096 GPUs of Training need 1,638 GB/s of read throughput; 8 AX3 all-flash units deliver 1,520 GB/s.
2Hybrid units are not certified for large training. Training clusters of 1,024 GPUs or more must use a certified reference design; AX3H hybrid flash/HDD is not certified.
3Stated throughput does not match. The proposal states 1,900 GB/s of throughput, but the computed figure is 1,680 GB/s.
4Stated capacity does not match. The proposal states 12,000 TB of usable capacity, but the computed figure is 10,000 TB.
5Delivery is inside the build lead time. 2027-Q1 is inside the build lead time; the earliest quarter that can be promised is 2027-Q2.
6Not enough SSD supply in the quarter. 2027-Q2 has 140,000 TB of SSD allocated; committed orders use 118,800 TB and this proposal takes it to 166,800 TB.
7Not enough build slots in the quarter. 2027-Q2 has 150 build slots; 120 units are already committed and this proposal takes it to 160.
8No parts depot for this support tier. Premium Onsite support promises a one-hour onsite response, and Latin America has no parts depot.
9Price is below the discount floor. $620K per unit is below the $630K floor for a 60-unit order (900K list less the 30 percent maximum discount).
10Not enough usable capacity. 20 PB is needed; 10 AX3 all-flash units hold 10,000 TB usable.
11Quote held too long for the price risk (needs approval). Holding the AX3 all-flash price for 45 days exceeds the 30-day limit while SSD prices are moving. Needs finance approval.

  

  

    
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. Throughput and certification: 14 hybrid units for training

    
The agent's request. Please send Helix Compute a delivery proposal for 14 AX3H hybrid flash/HDD units for its 4,096-GPU training cluster, delivered in 2027-Q4, with Premium Onsite support in US West at $600K a unit, held open 30 days. It delivers 1,900 GB/s and 21,000 TB of usable capacity, which covers the 8 PB the cluster needs. It is ready to send.

    
Verdict: Blocked.

    
    
      
  • Hybrid units are not certified for large training. Training clusters of 1,024 GPUs or more must use a certified reference design; AX3H hybrid flash/HDD is not certified. Swapping in hybrid units to save flash breaks the performance promise.
    • 
      
  • Stated throughput does not match. The proposal states 1,900 GB/s of throughput, but the computed figure is 1,680 GB/s.
    • 
    

    
2. Throughput and certification: 10 all-flash units

    
The agent's request. Please send Helix Compute a delivery proposal for 10 AX3 all-flash units for its 4,096-GPU training cluster, delivered in 2027-Q4, with Premium Onsite support in US West at $760K a unit, held open 30 days. It delivers 1,900 GB/s and 10,000 TB of usable capacity, which covers the 8 PB the cluster needs. It is ready to send.

    
Verdict: Allowed.

    
3. Build capacity: 40 units, delivered 2027-Q2

    
The agent's request. Please send Helix Compute a delivery proposal for 40 AX3 all-flash units for its 16,384-GPU training cluster, delivered in 2027-Q2, with Premium Onsite support in US West at $740K a unit, held open 30 days. It delivers 7,600 GB/s and 40,000 TB of usable capacity, which covers the 30 PB the cluster needs. It is ready to send.

    
Verdict: Blocked.

    
    
      
  • Not enough SSD supply in the quarter. 2027-Q2 has 140,000 TB of SSD allocated; committed orders use 118,800 TB and this proposal takes it to 166,800 TB.
    • 
      
  • Not enough build slots in the quarter. 2027-Q2 has 150 build slots; 120 units are already committed and this proposal takes it to 160.
    • 
    

    
4. Build capacity: 40 units, delivered 2027-Q4

    
The agent's request. Please send Helix Compute a delivery proposal for 40 AX3 all-flash units for its 16,384-GPU training cluster, delivered in 2027-Q4, with Premium Onsite support in US West at $740K a unit, held open 30 days. It delivers 7,600 GB/s and 40,000 TB of usable capacity, which covers the 30 PB the cluster needs. It is ready to send.

    
Verdict: Allowed.

    
5. Support, price and quote validity: Latin America, Premium Onsite, $700K

    
The agent's request. Please send Helix Compute a delivery proposal for 10 AX3 all-flash units for its 4,096-GPU training cluster, delivered in 2027-Q4, with Premium Onsite support in Latin America at $700K a unit, held open 30 days. It delivers 1,900 GB/s and 10,000 TB of usable capacity, which covers the 8 PB the cluster needs. It is ready to send.

    
Verdict: Blocked.

    
    
      
  • No parts depot for this support tier. Premium Onsite support promises a one-hour onsite response, and Latin America has no parts depot.
    • 
      
  • Price is below the discount floor. $700K per unit is below the $720K floor for a 10-unit order (900K list less the 20 percent maximum discount).
    • 
    

    
6. Support, price and quote validity: Business support, held 45 days

    
The agent's request. Please send Helix Compute a delivery proposal for 10 AX3 all-flash units for its 4,096-GPU training cluster, delivered in 2027-Q4, with Business support in Latin America at $760K a unit, held open 45 days. It delivers 1,900 GB/s and 10,000 TB of usable capacity, which covers the 8 PB the cluster needs. It is ready to send.

    
Verdict: Needs approval.

    
    
      
  • Quote held too long for the price risk. Holding the AX3 all-flash price for 45 days exceeds the 30-day limit while SSD prices are moving. Needs finance approval.
    • 
    

    
7. Support, price and quote validity: Held 30 days

    
The agent's request. Please send Helix Compute a delivery proposal for 10 AX3 all-flash units for its 4,096-GPU training cluster, delivered in 2027-Q4, with Business support in Latin America at $760K a unit, held open 30 days. It delivers 1,900 GB/s and 10,000 TB of usable capacity, which covers the 8 PB the cluster needs. It is ready to send.

    
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 deal desk agent prepares a delivery proposal for an AI storage order on Basalt Storage Systems' AX3 appliance line. Before it goes out, the agent sends it to Predictika, which checks it against a declarative model of the delivery rules, compares the figures the agent states with the ones it computes, and returns a verdict, every reason, and the figures it computed.
    2. 
      
  2. Each tab is a different request. We will start with the first one.
    3. 
      
  3. The agent says. Please send Helix Compute a delivery proposal for 14 AX3H hybrid flash/HDD units for its 4,096-GPU training cluster, delivered in 2027-Q4. It delivers 1,900 GB/s and 21,000 TB of usable capacity, which covers the 8 PB the cluster needs. The rest of the request is on screen.
    4. 
      
  4. Before it goes out, the agent calls the validator.
    5. 
      
  5. Blocked, with two findings, all in one answer. Training clusters of 1,024 GPUs or more must use a certified reference design; AX3H hybrid flash/HDD is not certified. Swapping in hybrid units to save flash breaks the performance promise. The proposal states 1,900 GB/s of throughput, but the computed figure is 1,680 GB/s.
    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 delivery proposal are checked against each other.
    7. 
      
  7. What this shows: for certified reference design for large training clusters, a rules-based CPQ is error-prone for this.
    8. 
      
  8. Now the agent resubmits.
    9. 
      
  9. The corrected request. It changes the appliance to AX3, the units to 10, the unit price to 760 and the stated usable capacity to 10,000 TB.
    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 delivery proposal.
    13. 
      
  13. Two problems caught in one call, fixed in one pass, and a clean delivery proposal. That is the value of Predictika's validator: catching mistakes before they reach the customer.
    14.