Build, Buy, or Partner for AI Capability

A sourcing decision is only coherent at a named boundary. 'Build AI' could mean training a foundation model, adapting a model, operating inference, creating an evaluation platform, integrating a managed API, or owning product workflow and policy. Draw the value stream and label what must be unique for the product, what creates regulated or safety-critical responsibility, what can be standardized, and what can remain behind an interface. Then compare modes for that bounded capability, not for an imagined all-or-nothing stack.

1. 01Freeze comparable scenariosUse the same demand distribution, data classes, regions, quality slices, reliability target, horizon, growth cases, and organizational constraints for every option.
2. 02Declare eligibility gatesSet minimum task and boundary quality, maximum latency and cost, required controls, data and rights constraints, launch horizon, concentration, and exit readiness before seeing favorable results.
3. 03Price the lifecycleInclude integration, evaluation, migration, platform, support, security, incident, vendor-management, staffing, retraining, reserved capacity, egress, overlap, and termination—not only API or salary lines.
4. 04Choose one primary objectiveFor example, minimize risk-adjusted lifecycle cost subject to all gates. State deterministic tie-breakers and keep the underlying dimensions visible.
5. 05Sequence reversible evidenceUse time-boxed discovery, a representative pilot, contract review, load and failure tests, and an exit exercise before a larger commitment.

1def decide_sourcing(contract: SourcingContract, evidence: SourcingEvidence):
2    contract = validate_record(contract, SourcingContract)
3    evidence = validate_record(evidence, SourcingEvidence)
4    if evidence.scope != contract.scope or evidence.contract_content_id != contract.content_id:
5        raise ValueError("evidence belongs to another sourcing contract")
6    # Disqualify each build, buy, and partner option independently.
7    # Select among eligible options by declared annual cost, then stable tie-breakers.

example=illustrative_only
decision=SELECT_ELIGIBLE_OPTION
selected=managed-service;mode=buy
eligible=internal-platform,managed-service
violations=delivery-partner/quality
claim=LOCAL_SOURCING_AUDIT_NOT_PROCUREMENT_APPROVAL

The fixture binds capability, risk, evaluation, security, exit, and fixture identities into one scope. Each option carries exact integer quality counts, latency, lifecycle cost units, launch time, supplier concentration, exit readiness, and four controls bound to its option identity, sourcing mode, and evidence digest. Cross-option control reuse is rejected. The partner option misses its own quality floor; that failure cannot be averaged with the other candidates. The eligible managed service wins the declared cost objective over the internal platform. All numbers are invented, and the result authorizes no purchase or production use.

• Contract for export formats, provenance, configuration, evaluation records, logs, deletion, assistance, notice, and post-termination access—not a generic portability promise.
• Retain internal competence to evaluate releases, operate incidents, interpret telemetry, enforce product policy, and execute the exit even when delivery is external.
• Keep product-owned interfaces and tests at the boundary so replacement does not require rewriting business workflows around a vendor-shaped payload.
• Treat dual-running and migration capacity as part of the exit cost; an export file alone is not a working replacement.

The decision record needs an owner, an evidence snapshot, explicit uncertainty, a review cadence, and event-driven triggers. Monitor realized task quality and boundary slices, latency, availability, spend, integration toil, incidents, supplier roadmap and control changes, internal staffing, concentration, and exit readiness. Reopen the record when an assumption moves beyond its declared envelope; do not wait for renewal pressure to discover that the exit path has decayed.

• Review control evidence and supplier changes on a schedule proportional to impact, not only during procurement.
• Reconcile invoices and metering with product outcomes; token or seat growth is not automatically delivered value.
• Exercise export, fallback, credential rotation, incident contact, and replacement interfaces before they are urgently needed.
• Use expiry dates for exceptions, exclusivity, pilots, and temporary partner authority.
• Publish why rejected options failed so future teams can update evidence instead of restarting a political debate.