The AI platform decision is the org's biggest AI bet — and under structural platform churn, it cannot be a one-time decision. This playbook builds the weighted decision record across Bedrock, Azure AI Foundry, Vertex AI, and self-hosted, with model exclusivity, ecosystem gravity, and exit costs scored — and a quarterly revisit built in.
Tier: Advanced · Hands-on time: ~40 minutes
The comparison literature of 2026 is consistent on the dimensions even where specific numbers drift: model exclusivity is real (GPT-5 first-party only via Foundry and OpenAI; Gemini only via Vertex; Claude available broadly including Bedrock, which markets low-latency claims), and catalog sizes differ by an order of magnitude — figures like ~1700 models on Foundry versus ~200 on Vertex versus ~100 on Bedrock circulate as directional, to be verified before any deck cites them, and catalog size is not catalog quality. The deeper lesson is churn itself: the Bedrock Agents classic-to-AgentCore transition is this decision's standing case study in how platform commitments age. Hence the artifact: a weighted decision record with a revisit date, not a marriage certificate.
Fix the dimensions and weights before touching platforms. The research-backed set: model access (which models, at what exclusivity), catalog breadth versus curation, ecosystem gravity (existing cloud, data locations, team skills), pricing model fit, compliance and residency capability, operational burden, and exit cost. Weight them for your org — a BFSI enterprise weights compliance and exit; a startup weights velocity — and the weights are yours, not the assistant's.
Scaffold the matrix with verification hooks. Vikram, setting the BFSI platform stance:
Scaffold a platform decision matrix: Bedrock vs Azure AI Foundry vs
Vertex AI vs self-hosted vLLM, for a BFSI enterprise with an existing
Azure estate, DPDP-scoped data, and three planned workloads (attached).
Dimensions and weights (mine, fixed): model access 20, ecosystem fit
20, compliance/residency 25, pricing fit 10, ops burden 10, exit cost 15.
For each cell: qualitative assessment, then a VERIFY line naming the
exact claim to confirm in current vendor docs (model availability in
our regions, residency guarantees, private-endpoint support). Mark
model-exclusivity claims explicitly — which models can we NOT get on
each platform — and mark every catalog-size or latency figure as
UNVERIFIED until I check it. Do not score; I score.
Verify the load-bearing cells hard. Model availability by region, exclusivity claims, private-connectivity and residency features, and pricing mechanics all come from vendor pages checked this week — platform capabilities are the fastest-decaying facts in this track, and a decision record citing a stale catalog number loses the room. Date-stamp every verified cell.
Score self-hosted honestly on both sides. The threshold argument is real — sufficient volume or hard compliance constraints can force it — but pb-42's utilization math and a named operational-burden line (serving stack, upgrades, security, on-call) go in the matrix, not a hand-wave. Self-hosted scores well on control and exit, and pays for it in ops; make both visible.
Price the exit for every arm. What leaves with you (prompts, evals, data, portable abstractions like MCP layers per pb-37) versus what is hostage (platform-specific agent configs, managed-index formats, fine-tunes). Cite the AgentCore transition as the worked example: teams on the deprecated service faced a migration they didn't schedule. Exit cost is the dimension most orgs under-weight and most regret under-weighting.
Score, decide, and make the AI argue the loser's case. You fill the scores; then have the assistant steelman the second-place option and attack the winner's weakest verified cells. Surviving arguments become consequences in the record; a decision that has never heard its best counterargument isn't decided, it's defaulted.
Write the weighted decision record with its revisit machinery. Decision, matrix with dated verifications, exit-cost sketches per arm, and — the part that makes it durable — the revisit triggers: quarterly review date, plus event triggers (model-exclusivity change, major platform deprecation, pricing restructure, new compliance rule). Platform churn is structural; the record's honesty about that is its authority.
Platform strategy work is mostly public-information analysis and safe for approved external assistants — the sensitive parts are your workload portfolio, volumes, and contract terms, which should be abstracted (workload classes and order-of-magnitude volumes) before leaving the boundary. For client engagements, the client's existing cloud agreements and MSA-approved vendor lists are constraints extracted before the matrix is built — often eliminating arms outright and shrinking the exercise usefully. The decision record itself is client work product in consulting contexts; build it in their environment, and keep your firm's reusable asset as the dimension framework and verification checklist, not the client's numbers.
A weighted record with dated verifications, exit sketches per arm, and a quarterly revisit — under structural churn, the platform decision's quality is measured by how gracefully it can change its mind.