Exercise: pick the path, the processor and the control

Original fictional scenario. No cloud account, API calls or paid services are required. Difficulty: intermediate · Estimated duration: 20–25 minutes

Harbor Mutual, a fictional insurer, has five AI requests on its desk. It runs Google Workspace, has three developers, and its claims staff do not code.

#	Request	Who would build it	Data involved
1	Translate policy pages into Spanish and Portuguese	Developers	Public web text
2	Pull policy number and amount from scanned claim forms	Developers	Customer personal data
3	Automate a weekly claims-status summary in Workspace	Claims staff	Internal email and docs
4	An agent that checks a claim, queries two internal systems and drafts a decision	Developers	Customer personal data
5	Train a pricing model whose maths needs high-precision arithmetic	Data science team	Historical claims

Your decision

Assign each request a path: pre-trained API, no-code, low-code or code. Justify each.
For request 5, say whether TPUs are the right processor, and why.
For requests 3 and 4, name the data control that applies at that layer and who must set it up.

Rubric (10 house points)

4 points: a defensible path for all five requests, choosing the highest-level path that works.
2 points: request 5 steered away from TPUs, with the reason.
2 points: request 3 tied to the Workspace commitment that interactions stay within the organization.
2 points: request 4 tied to a per-agent identity and to the training restriction, with the customer's part named.

Reference solution

Request 1 → pre-trained API. Cloud Translation's Basic API gives quick, plug-and-play access to a standard translation model; building a translator would spend months on a solved problem. Request 2 → pre-trained API. Document processing services do exactly this workflow: get the raw text, then extract the fields needed. Rejected: a custom agent written with a development kit. Request 3 → no-code. Workspace Studio automates everyday work with no coding required, so claims staff can own it. Request 4 → code. The multi-system logic justifies developers; low-code is reasonable only if the team's canvas can express both system calls.

Request 5: TPUs are optimized for specific workloads, and Google lists workloads that require high-precision arithmetic among those TPUs do not suit — plan for other processors. Rejected: "they are Google's own chips, so they must be best."

Controls. For request 3, Google's Workspace privacy hub says interactions with Gemini stay within the organization; the administrator decides who gets the feature. For request 4, give the agent its own managed identity, which enables access control and auditing, and rely on the commitment that Google won't train on the data without permission. Harbor Mutual still sets the agent's permissions; zero data retention, if required, needs its own configuration steps.

Sources and scope

All company names, requests and the rubric are house-authored. Product facts are grounded in:

Pick the path, the processor and the control — decision exercise

Exercise: pick the path, the processor and the control

Your decision

Rubric (10 house points)

Reference solution

Sources and scope