Weigh a platform pitch — decision exercise
Exercise: weigh a platform pitch
Original fictional scenario. No cloud account, API calls or paid services are required. Difficulty: intermediate · Estimated duration: 20–25 minutes
Northwind Outfitters, a fictional retailer, must decide whether to build its customer-service assistant on Google Cloud. Its leadership has a vendor pitch containing the four statements below, and its own requirement list. Its staff use Microsoft SharePoint and Jira; nobody uses Google Workspace. Its two developers have no infrastructure experience.
| # | Statement in the pitch | Evidence attached |
|---|---|---|
| 1 | "The models are built from the ground up for multimodality." | None |
| 2 | "Our custom chips make every AI workload faster." | None |
| 3 | "This is the most accurate assistant platform for retail." | A public leaderboard screenshot |
| 4 | "The platform is enterprise-ready: secure, private and reliable." | None |
Northwind's requirements: zero data retention for customer chats; the assistant stays up if one region fails; each agent can reach only the order system; answers draw on SharePoint and Jira content; the option to switch to a cheaper open model later.
Your decision
- Classify each statement as a description, a comparison or positioning, and say what test would make it useful to Northwind.
- For statement 4, list what Northwind itself must configure to make each requirement true.
- Say where Google's ecosystem integration helps Northwind and where it does not.
- Explain what switching to an open model later would require, and how Northwind could avoid running servers.
Rubric (10 house points)
- 3 points: all four statements classified, each with a concrete test on Northwind's own work.
- 3 points: a configuration for each of zero retention, regional resilience and per-agent access.
- 2 points: integration value tied to Northwind's real tools (SharePoint, Jira), not Workspace.
- 2 points: names re-evaluation after a model switch and a managed path for open models.
Reference solution
Statement 1 is Google's description of how it built Gemini; test it by running Northwind's own mixed customer messages, photos included. Statement 2 overreaches Google's own documentation, which says Cloud TPUs are optimized for specific workloads — treat it as positioning and ignore it until a Northwind workload is measured. Statement 3 is a comparison; a public leaderboard cannot settle it, because Google says evaluation on your own tasks gives insights that leaderboards and general benchmarks cannot. Run an evaluation on a few hundred past Northwind chats instead. Statement 4 is positioning that becomes true only with configuration.
For the requirements: zero data retention needs the specific customer actions Google lists — it is not automatic. Regional resilience means deploying to endpoints in several regions or using the global endpoint. Per-agent access means granting each agent narrow permissions with Agent Identity.
Integration helps through Gemini Enterprise's prebuilt connectors to SharePoint and Jira; the Workspace benefits do not apply, because Northwind does not use Workspace. Rejected: choosing Google Cloud because Gemini is embedded in Workspace.
A later switch to an open model means re-running the same evaluation before going live. Managed APIs serve open models without managing infrastructure, which suits Northwind's two developers. Rejected: "the platform is the same, so the new model needs no testing."
Sources and scope
All company names, statements, requirements and the rubric are house-authored. Product facts are grounded in:
- https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/evaluation-overview
- https://docs.cloud.google.com/tpu/docs/intro-to-tpu
- https://docs.cloud.google.com/gemini-enterprise-agent-platform/resources/zero-data-retention
- https://docs.cloud.google.com/architecture/framework/perspectives/ai-ml/reliability
- https://docs.cloud.google.com/gemini-enterprise-agent-platform/overview
- https://docs.cloud.google.com/gemini/enterprise/docs
- https://docs.cloud.google.com/gemini-enterprise-agent-platform/models