Build vs buy: a CTO's framework for AI products.
In 2026 the question isn't whether you'll add AI. It's whether you'll build it, buy it, or wait six months for the answer to change. Here's the framework we use with our buyers — including the third option most decision matrices skip.
Every product CTO is being asked the same question right now: should we build custom AI for this, or use something off-the-shelf? The honest answer is "it depends" — but it depends on a small set of variables, and most teams ask them in the wrong order.
Start with the wrong question, fail twice.
The default question is "how powerful does the model need to be?" That's the wrong start. The model is the cheapest part of any AI product in 2026; you'll swap it twice before the product matures. The right question is "who owns the data, and where does it live?"
If your data is in your customer's tenancy and they're regulated, your build-vs-buy answer is constrained before you start. If your data is yours and you're trying to differentiate on outputs, the answer space is much wider.
The four-axis matrix.
We score every AI feature against four axes:
1. Data sensitivity.
Public data, internal data, customer data, regulated data. Each step up adds requirements that off-the-shelf vendors can't always meet. Customer data + regulated industry ≈ build, or buy from a vendor with deep compliance posture.
2. Differentiation surface.
Is the AI feature the product, or a feature inside the product? An AI-native product (where the AI behavior is the value) needs control of the inference pipeline. A product where AI is one of twelve features can usually rent that pipeline.
3. Frequency of model upgrade.
If your differentiation is "we get the latest model fastest," buy. The major labs ship new models every 4–8 months and you cannot keep up with infrastructure work alone. If your differentiation is "we never need a new model because our retrieval is right," build.
4. Cost-of-failure asymmetry.
If the AI being wrong costs the customer real money or risk (legal, medical, financial), you need observable, auditable inference — usually a build, occasionally a buy from a vendor that exposes the audit trail. If wrong-but-recoverable is fine (autocomplete, summaries, sorting), buy.
The decision tree.
- Buy when: low data sensitivity, AI is one feature among many, no audit requirements, model upgrades matter.
- Build when: customer or regulated data, AI is the product, audit trail required, your retrieval/data is the moat.
- Hybrid when: you need control of orchestration but the model is interchangeable. Build the orchestration; rent the model. This is the answer for most B2B products in 2026.
- Wait when: the use case is real but the model isn't quite there. Sometimes the right answer is to ship a non-AI version, instrument it, and revisit in two quarters. Especially true for agentic features in 2026 — the gap between demo and production is still wide.
The cost math nobody runs.
Build estimates are usually high. Buy estimates are usually low. The numbers we see in 2026:
- Buy: $0.5–5 per 1k requests at scale, plus integration. Total Year-1 spend for a typical B2B feature: $50–250k.
- Hybrid (BYO model + custom orchestration): $80–180k engineering build + $0.3–2 per 1k requests. Year-1: $150–400k.
- Full build (custom retrieval, custom inference, vector store, eval pipeline): $300–800k engineering + $0.2–1 per 1k requests. Year-1: $400–1M+.
The number that flips most decisions: vendor lock-in cost two years out. A buy at $50k/year that can't be migrated is a $500k commitment by year three, because you can't leave. Always price the exit.
The "wait" option is underused.
Sometimes the right move in 2026 is a non-decision. The model isn't there yet (true for most "fully autonomous agent" pitches), the regulation is moving (true for healthcare and financial AI), or your team isn't ready (true honestly more often than people admit). Shipping a non-AI version, instrumenting the workflow, and revisiting in 90 days is a real strategy. It's also the one consultants will never recommend, because they can't bill for it.
The shortest version.
Score the feature on data sensitivity, differentiation, upgrade cadence, and cost-of-failure. If three or more axes lean "build," build. If three or more lean "buy," buy. If the matrix splits, the answer is hybrid — almost always.
And if your in-house team has never shipped a production AI feature, hybrid is probably your only realistic answer this year. The next post explains what that hybrid actually looks like in production.
Oviompt builds custom AI surfaces for regulated B2B buyers — legal AI, voice agents, retrieval pipelines that pass compliance. File an intent if you have a build-vs-buy decision in progress; references on request.