ai-build-buy-partner

AI Build, Buy, or Partner

Systematically evaluate AI capability sourcing options to find the right balance of speed, cost, control, and strategic differentiation.

Context

You are evaluating AI capability sourcing options for $ARGUMENTS.

Instructions

1. Define the capability and requirements:

   • What AI capability is needed? (language understanding, generation, vision, recommendations, predictions)
   • What are the quality, latency, cost, and compliance requirements?
   • How central is this capability to the product's competitive differentiation?
   • What data does the team currently own relevant to this capability?

2. Evaluate the four sourcing archetypes:

   Archetype 1 — Pure API (e.g., OpenAI, Anthropic, Google):

   • Fastest to market; no ML infra required; pay-per-use
   • Best for: commodity tasks, prototyping, teams without ML capability
   • Risks: vendor dependency, data privacy exposure, cost at scale, model deprecation

   Archetype 2 — Fine-tuned base model (e.g., fine-tuned Llama, Mistral):

   • Moderate customisation; requires labelled data and ML engineering
   • Best for: domain-specific language, consistent output format, moderate data advantage
   • Risks: training cost and time, ongoing maintenance, model update cycle

   Archetype 3 — Hybrid (API + RAG / retrieval augmentation):

   • Use a foundation model API with a proprietary knowledge base for grounding
   • Best for: knowledge-intensive applications with dynamic data (docs, policies, product info)
   • Risks: retrieval quality bottleneck, context window limits, latency from retrieval step

   Archetype 4 — Fully custom model:

   • Full control of architecture, training data, deployment, and updates
   • Best for: core differentiator with large proprietary dataset, regulatory data residency, extreme cost at scale
   • Risks: high up-front investment, long timeline, requires significant ML talent

3. Score each archetype on the decision matrix:

   Dimension	Weight	Pure API	Fine-tuned	Hybrid	Custom
   Strategic differentiation	20%	/5	/5	/5	/5
   Data advantage leverage	15%	/5	/5	/5	/5
   Cost at scale (10× volume)	15%	/5	/5	/5	/5
   Time-to-market	15%	/5	/5	/5	/5
   Compliance / data residency	15%	/5	/5	/5	/5
   Customisation / control	10%	/5	/5	/5	/5
   Team capability match	10%	/5	/5	/5	/5
   Weighted total	100%

4. Cost modelling — API vs. infrastructure:

   • API cost at scale: monthly_requests × avg_tokens × price_per_1k_tokens
   • Model the break-even point: at what volume does the infrastructure cost of a self-hosted model become cheaper than API cost?
   • Include: GPU instance cost, serving infra, MLOps tooling, ML engineer time
   • Rule of thumb: API is usually cheaper below ~$50K/month at current pricing

5. Make-or-buy signals:

   • Build custom when: it is a core differentiator, you have 10× more proprietary data than competitors, regulatory requirements prohibit third-party APIs, or API cost exceeds $500K/year
   • Buy/use API when: commodity task, fast time-to-market needed, team lacks ML capability, or volume is low
   • Fine-tune when: output style/format consistency is critical, domain vocabulary is specialised, or hallucination rate on off-the-shelf models is too high

6. Vendor evaluation checklist:

   • Model quality on your specific task (run offline eval, do not rely on leaderboards alone)
   • API reliability: uptime SLA, rate limits, support tier
   • Data privacy: zero retention policy, BAA availability, data residency options
   • Pricing transparency and volume discount availability
   • Model deprecation policy: advance notice, migration support
   • OpenAI-compatible API surface (reduces lock-in risk)
   • Roadmap alignment: is the vendor investing in capabilities you need?

7. Partnership vs. licensing:

   • Technology licensing: pay for access to a model weight or API; no co-development
   • Strategic partnership: joint development, data sharing, co-marketing; higher commitment, higher leverage
   • Evaluate: does the partner provide a distribution channel? Proprietary data? Talent access?
   • Negotiate: data usage rights, exclusivity window, model improvement sharing, SLA commitments

8. Produce recommendation report:

   • Recommended archetype with weighted score justification
   • Cost model at current and projected volume
   • Build trigger: the conditions under which a custom-build becomes justified
   • Risks and mitigation plan
   • Short-term (proof-of-concept) and long-term sourcing roadmap

Think step by step. Save as markdown.