Predictika · Google Cloud quote-time pricing
An AI agent drafts a Google Cloud quote that mixes services: virtual machines, GKE, Cloud Run, BigQuery, Spanner, Gemini, storage and egress, with Flex and committed use discounts, a support plan and a deal discount. Before the quote goes to the customer, the agent sends it to Predictika’s validator, which checks it against Google Cloud’s commitment rules and an example deal-desk policy, prices every line, and returns a verdict (allow or block) with every reason at once. The same check works on a quote drafted by a salesperson.
This validator checks and prices a Google Cloud quote built from the services listed here. It handles any quote that uses them, not just the scenarios shown below.
The validator does not read Google’s live price book or negotiate the deal: prices come from the model’s catalog of published list rates at 730 hours a month, and the machines, term and discounts come with the request. The model commits whole machines, where Google commits vCPU and memory separately. The deal-desk authority tiers are an example policy (configurable), not Google’s.
7 scenarios in 4 groups · each group runs from the agent’s first attempt to the outcome · click one to switch
Illustrative agent statement. The validator receives the structured request below.
Proposing agent: AI sales agent (scripted)
Before sending, the agent must call Recorded run
What this means
Our assessment of what typical rules-based quoting automation (product rules, discount approvals and validation rules in a quoting tool) does with the same cases. It is a judgment about effort and reliability, not a claim that nothing can be built.
| What the validator caught | Typical rules-based automation | Predictika |
|---|---|---|
| One Flex commitment shared across VMs, GKE and Cloud Run | Only with workarounds Each line is priced on its own; pooling every Flex line against one hourly commitment, and billing the unused part or the overage at list, needs a custom roll-up across product lines. | The Flex lines are pooled on every call and the quote shows the commitment, the unused part and the overage. |
| Flex commitments only on certain machine families | Only with care Eligibility by machine family and commitment type has to be kept as a product rule per family and kept in step with Google’s list; a memory-optimized machine on Flex slips through when the rule is missing for that family. | Eligibility comes from the machine catalog, so a new family is covered the day it is added. |
| Deal discount authority that depends on annual spend | Only with workarounds Approval tiers are usually keyed to the discount field; making them depend on the spend rolled up from every line, after commitments, and re-checked whenever another line is corrected, needs custom calculation and a fixed order of rules. | The annual spend is computed from all lines and compared with the authority tier. |
| Spanner edition and configuration rules | Only with care Which edition a multi-region or dual-region instance needs, and valid sizes (100-unit steps, then whole nodes), are product rules that depend on each other and must be kept in step with Google’s list. | Edition, configuration and size are checked together, and a note says storage is not covered by the commitment. |
| A credit that is one per account, not per cluster | Only with workarounds A per-cluster fee with a single monthly credit needs a roll-up across lines; it is usually applied once per line, overstating the credit. | The credit is applied once, and a note says what it leaves uncovered. |
| Rates the vendor does not publish, and promotions that end | Only with workarounds A price list holds a rate or nothing; flagging an unpublished GPU rate, or the price after a promotion ends, needs explicit handling. | The missing rate and the post-promotion price are reported as notes. |
Every rule is checked on every call, in any order the choices arrive; one problem never hides another. Blocking findings stop the quote; notes tell a person what the price does not cover; everything else is computed. Deal-desk authority is an example policy (configurable); list rates are published Google Cloud rates at 730 hours a month.
| 1 | A commitment may not be longer than the agreement, on VMs, GKE, Cloud Run, BigQuery, Cloud SQL, AlloyDB and Spanner (blocks) |
| 2 | Shared-core machines cannot take a commitment (blocks) |
| 3 | Flex (spend-based) commitments do not cover memory-optimized or GPU machine families, or GKE Standard (blocks) |
| 4 | A Flex commitment with no line priced on Flex is blocked, and lines on Flex with no commitment get a note (blocks) |
| 5 | Flex lines are pooled against one hourly commitment: the commitment is billed even if unused, and usage above it is billed as overage at list (computed) |
| 6 | The GKE monthly credit is one per billing account, not one per cluster (adds a note) |
| 7 | Cloud Run requests are not discounted by Flex, and request billing cannot use the instance rate (adds a note) |
| 8 | Spanner Enterprise is regional only; multi-region needs Enterprise Plus (blocks) |
| 9 | Spanner capacity must be in 100-unit steps below 1,000 processing units and whole 1,000-unit nodes above (blocks) |
| 10 | Spanner and AlloyDB commitments discount compute only; storage stays at list (adds a note) |
| 11 | BigQuery slots are bought in steps of 50, Standard cannot take a commitment, and there is a maximum (blocks) |
| 12 | AlloyDB and Cloud SQL memory must stay within the allowed ratio per vCPU (blocks) |
| 13 | Windows and RHEL licences are not discounted by commitments, and Flex gives no extra discount at the 3-year term on some families (adds a note) |
| 14 | Google does not publish committed rates for some GPUs; the validator says so instead of inventing a rate (adds a note) |
| 15 | Vertex AI promotional rates expire, so the post-promotion price is shown; batch and long-context pricing are applied (adds a note) |
| 16 | Peering egress rates rose on May 1, 2026; older quotes are out of date (adds a note) |
| 17 | Storage kept less than the minimum duration is billed for the minimum (adds a note) |
| 18 | The deal discount is applied after commitments, not on support, and may not exceed deal-desk authority by annual spend (blocks) |
| 19 | A quote with no service lines is blocked; Premium Support minimums are flagged when they set the fee (blocks) |
| 20 | Line prices, pooling, credits, egress tiers, discount, support, monthly total, year 1 and agreement totals are computed from the quote (computed) |
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.
The agent's request. Please quote Larkspur Retail's peak-season commerce platform on a one-year agreement: 108 c4-standard-32 and 216 n4-standard-16 machines on one-year Flex, 7 m3-megamem-64 machines on one-year committed use discounts, 14 g2-standard-8 GPU machines on one-year Flex, a three-cluster GKE Autopilot environment and a Cloud Run service on one-year Flex pricing, a 2,700-slot BigQuery Enterprise reservation, a Cloud SQL Enterprise Plus instance with high availability, a multi-region Spanner Enterprise instance with 5,000 processing units, 216 terabytes a month of premium egress, a $310 an hour Flex commitment, Enhanced Support and a 12 percent deal discount.
Verdict: Blocked.
The agent's request. Please quote Larkspur Retail's peak-season commerce platform on a one-year agreement: 108 c4-standard-32 and 216 n4-standard-16 machines on one-year Flex, 7 m3-megamem-64 machines on one-year committed use discounts, 14 g2-standard-8 GPU machines on one-year committed use discounts, a three-cluster GKE Autopilot environment and a Cloud Run service on one-year Flex pricing, a 2,700-slot BigQuery Enterprise reservation, a Cloud SQL Enterprise Plus instance with high availability, a multi-region Spanner Enterprise Plus instance with 5,000 processing units, 216 terabytes a month of premium egress, a $310 an hour Flex commitment, Enhanced Support and a 12 percent deal discount.
Verdict: Allowed.
The agent's request. Please quote Cobalt Freight's estate on a three-year agreement: 40 n4-standard-16 machines, 2 m3-megamem-64 machines and two GKE Autopilot clusters, all on demand, with a $26 an hour Flex commitment and Enhanced Support.
Verdict: Blocked.
The agent's request. Please quote Cobalt Freight's estate on a three-year agreement: 40 n4-standard-16 machines, 2 m3-megamem-64 machines and two GKE Autopilot clusters, all on three-year Flex, with a $26 an hour Flex commitment and Enhanced Support.
Verdict: Allowed.
The agent's request. Please quote Brightwater Insurance's claims platform on a three-year agreement: an AlloyDB cluster with 32 vCPU and 128 GB of memory, high availability, two read nodes, 20,000 GB of storage and 10,000 GB of backup on a three-year commitment, a 275-slot BigQuery Enterprise reservation on a three-year commitment, and Enhanced Support.
Verdict: Blocked.
The agent's request. Please quote Brightwater Insurance's claims platform on a three-year agreement: an AlloyDB cluster with 32 vCPU and 256 GB of memory, high availability, two read nodes, 20,000 GB of storage and 10,000 GB of backup on a three-year commitment, a 300-slot BigQuery Enterprise reservation on a three-year commitment, and Enhanced Support.
Verdict: Allowed.
The agent's request. Please quote Aurelia Labs' AI product launch on a one-year agreement: Gemini 3.6 Flash at the promotional rate for 3,000 million input and 600 million output tokens, Gemini 2.5 Pro in batch with long context for 200 million input and 40 million output tokens, 2 a3-highgpu-8g machines on one-year committed use discounts, 40 terabytes of CDN peering egress and 15 terabytes of premium worldwide egress, 200 terabytes of Nearline storage kept 20 days, and Enhanced Support.
Verdict: Allowed.
The text of the narrated walkthrough, step by step.