Specify the capability before comparing options
A vendor product and an internal proposal cannot be compared until both are asked to satisfy the same workflow, rights, deployment, quality, and operating constraints.
One workflow
Name the trigger, inputs, accepted output, user or downstream system, failure consequence, volume, latency, and person accountable for the result.
Hard requirements
Separate non-negotiable rights, runtime, deployment, security, integration, and quality requirements from preferences. A cheap partial component is not a valid option for a complete service.
Acceptance contract
Define representative test cases, baseline, review burden, release gate, and stop condition before the team becomes invested in a particular path.
Map build, buy, compose, and stop
Give each path a fair version. A weak vendor shortlist and a heroic internal estimate merely formalize the preference the team already had.
Build
Estimate the smallest maintainable capability that meets the contract, including evaluation, tooling, CI, review, deployment, monitoring, and ownership—not just generation time.
Buy or compose
Check functional coverage, integration work, rights, data terms, portability, vendor viability, change control, and the cost of operating the acquired capability in your environment.
Stop
Include the option to preserve the current workflow or redesign it without AI when the expected outcome cannot earn the total cost or operating risk.
Model full ownership, not token price
The decision changes when integration, evaluation, review, failure, and change costs sit beside engineering and inference.
Creation and acquisition
Model engineering or agent effort, tools, compute, CI, purchase or subscription cost, procurement, and the work required to prove that the capability meets the acceptance contract.
Operation
Include context, inference, retries, orchestration, observability, human review, incident handling, failure remediation, and vendor or model changes.
Sensitivity
Show which assumptions can reverse the recommendation: volume, review rate, provider price, integration depth, model quality, internal reuse, or time-to-value.
Fund a bounded next move
A good comparison ends with an owned action and evidence gate—not a scorecard that leaves the same debate alive.
Recommendation
State build, buy, compose, stop, or evaluate-further; include the assumptions, confidence, disqualifiers, and evidence gaps capable of changing it.
Pilot charter
If uncertainty remains, fund the smallest pilot that resolves it. Name the cases, owner, timebox, cost ceiling, success threshold, and stop rule.
Decision boundary
Keep planning separate from execution. A comparison should not spend funds, create a purchase, or silently authorize production access.