Exercise: place one project's decisions on the landscape

Original fictional scenario. No cloud account, API calls or paid services are required. Difficulty: beginner · Estimated duration: 15 minutes

An insurer's generative AI programme has a backlog of open decisions. All names and details below are invented.

#	Open decision
1	Should the research team train its own claims model on TPUs or GPUs?
2	Which foundation model should power all five internal tools?
3	Who is allowed to deploy a new agent to production?
4	May the claims agent issue refunds under 200 without a person approving?
5	Should adjusters get meeting notes captured automatically in Meet?
6	Should staff search Jira and SharePoint from one place, seeing only what they may?
7	Should agent quality be measured continuously on live traffic?
8	Should the coding assistant be rebuilt on a different model?

Your decision

Place each decision on one layer: infrastructure, models, platforms, agents or applications.
For each, name who usually makes it: an engineering or procurement team, a builder team, or a business owner.
Decision 2 changes something under five tools at once. Explain what that means for testing.
Decision 4 is about autonomy. State what bounds what the agent may do on its own.

Rubric (10 house points)

4 points: all eight decisions on the right layer (half a point each).
2 points: a sensible owner for each layer.
2 points: decision 2 explained as a model change reaching every application built on it.
2 points: decision 4 answered with tools, permissions and a human review point.

Reference solution

1 — infrastructure (accelerators, networking and storage). 2 and 8 — models. 3 and 7 — platforms (governing who deploys; assessing and refining agent quality). 4 — agents (what an agent may do through its tools). 5 and 6 — applications (Gemini in Workspace for notes; Gemini Enterprise for permissions-aware search across connected apps).

Owners follow the layer: engineering and procurement for infrastructure; builder teams for models, platforms and agents; business owners for applications.

Decision 2: foundation models are the core many applications are built on, so a model change must be retested in all five tools. Rejected alternative: "test only the tool whose team asked".

Decision 4: an agent acts through the tools it is given, within permissions granted to it, with people supervising. Give it a refund tool only up to the limit, grant that permission explicitly, and keep a person reviewing above it. Rejected alternative: "trust the model's judgment because it is accurate on average".

Sources and scope

All names, decisions and rubric points above are house-authored. The layer definitions applied are grounded in:

Place one project's decisions on the landscape — decision exercise

Exercise: place one project's decisions on the landscape

Your decision

Rubric (10 house points)

Reference solution

Sources and scope