Build, buy or fine-tune an AI capability?

When it comes up

When a team proposes a new AI feature or internal tool, when a vendor pitches an AI product, or when someone suggests training or fine-tuning a model.

Your options

Buy a SaaS product

Fits when: The capability is not a differentiator, a mature product exists, and data can be shared under acceptable terms.

Watch out: Lock-in, pricing changes, limited customization, data residency.

API model + your own application

Fits when: The workflow or data is a differentiator, and you want control over UX and data handling while renting the model.

Watch out: Provider dependency (mitigate with a gateway and evals); token costs at scale.

Self-host an open-weights model

Fits when: Data must stay in-house, volumes are high and steady, or you need offline or edge deployment, and you have ML ops skills.

Watch out: GPU cost and utilization, operations burden, keeping up with model releases.

Fine-tune a model

Fits when: Evals show a behaviour, format or narrow-task gap that prompting and retrieval can't close, and you have quality training data.

Watch out: Data preparation, retraining when base models improve, evaluation effort; retrieval often beats fine-tuning for knowledge.

How to decide

Define the use case and success metric, and build a small eval set from real examples.
Score each option 1–5 on cost, latency, quality on your evals, security and data control, lock-in and team skill.
Agree weights with product and finance before scoring.
Run the sensitivity check: if the winner flips when one weight moves ±10 points, run a time-boxed proof of concept before deciding.
Record the decision as an ADR with review triggers (price changes, new models, volume changes).

Use the build-vs-buy matrix to do steps 2–5 in one place.

Trade-offs

Buying is fastest but least controllable; building on APIs balances control and speed but creates provider dependency; self-hosting maximizes control at the cost of operations; fine-tuning improves narrow fit but adds lasting maintenance.

Real examples

Bloomberg built BloombergGPT (2023), a 50-billion-parameter model trained on a 363-billion-token dataset from Bloomberg's own data sources plus 345 billion tokens of general data — an illustration of the data and compute scale behind building a domain model from scratch. Source: Wu et al. — BloombergGPT: A Large Language Model for Finance (arXiv 2303.17564, 2023) ↗
Researchers comparing knowledge-injection methods (Ovadia et al., 2023) found that retrieval-augmented generation consistently outperformed unsupervised fine-tuning, for both existing and entirely new knowledge. Source: Ovadia, Brief, Mishaeli & Elisha — Fine-Tuning or Retrieval? Comparing Knowledge Injection in LLMs (arXiv 2312.05934, 2023) ↗

What to check first

Is this capability a differentiator for your product or business?
Do you have an eval set to compare options on your own data?
What are the data residency and confidentiality requirements?
Who will own and maintain the capability for the next five years?

Learn and use

Sources

Last reviewed 2026-10-22.