Every boardroom argument about AI ends up sounding like a broken record: build or buy. That binary choice feels satisfying because it looks decisive. But this framing hides the real decision. The real question is whether you can capture capability arbitrage. In other words, does your organization gain a durable advantage by owning and operating a capability that others cannot easily replicate, or is it better to access that capability from an external provider and focus your energy elsewhere?
Here is a practical way to think about the problem, illustrated with modern examples and a decision framework you can apply today.
Why capability arbitrage matters more than build or buy
Most build-versus-buy debates reduce to speed versus control. Buy is fast. Build gives control. That is true but incomplete. AI changes the calculus because value often sits in proprietary data, entrenched workflows, and regulatory guardrails. If those elements create a wedge that others cannot easily copy, you have capability arbitrage. If not, you probably should buy and move on.
Industry observers now recommend an integrated strategy that mixes building, buying, and partnering to optimize for impact rather than ideology. Harvard Business Review explicitly recommends investing intelligently across build, buy, blend, and partner strategies.
How to spot capability arbitrage: three lenses
- Proprietary data moat. Ask whether your internal data materially improves model performance in ways customers or competitors cannot replicate. Companies that monetize industry-specific data by fine-tuning models are turning that data into product. Bayer’s work with Microsoft to publish agri models trained on Bayer data shows how industry data can become a licensed product.
- Workflow control and integration. Some gains come from deep integration, not raw model skill. Legal teams at Hewlett Packard Enterprise built tools in-house because sharing sensitive contract data with vendors was unacceptable and internal tools were easier to iterate on for their workflows. When integration and iterative fixes matter more than raw model cleverness, build becomes attractive.
- Compliance and risk containment. Regulated industries can face compliance and privacy requirements that make buying risky unless the vendor offers strict contractual and technical guarantees. When control over data lineage, audit trails, and governance is a business requirement, that is a strong reason to build or to co-develop with a vendor who will commit to enterprise-grade controls.
A decision matrix you can use
Score each use case 1 to 5 on these dimensions: proprietary data value, workflow integration complexity, regulatory risk, speed-to-value need, and internal AI capability. Add the scores.
- High proprietary data value + high regulatory risk + strong internal talent = lean toward build or co-develop.
- Low proprietary data value + need for speed + limited talent = buy.
- Mixed scores = buy-and-extend or partner model.
This is not just theory. The product community now argues that the future will be some blend of built components and bought components, with customers and vendors both building parts of the final system. The Silicon Valley Product Group notes that yes to both is likely the future.
Hidden costs that shift the answer
Don’t be fooled by sticker price only. The economics of building a production-grade AI agent are non-trivial. Recent analyses of building agentic solutions show very large one-time and recurring costs for fully custom agents and warn that a DIY AI stack can become brittle and expensive. In some modern comparisons, the economics favor platform-based approaches because building all foundational layers results in repeated rebuilds and higher long-term costs.
That does not mean building is always worse. Where capability arbitrage exists, the long-term return can justify the investment. But you must include operational costs, governance, monitoring, and retraining budgets when you compare total cost of ownership.
Cases that map to the framework
- HPE legal team. A legal department built tailored tools because sensitivity of contract data and need for rapid iteration made vendor options unattractive. The internal build proved cheaper and more controllable for their needs.
- Bayer and Microsoft. Bayer used proprietary agronomy data to train industry models that Microsoft makes available to customers, turning specialized data into a monetizable capability. That is capability arbitrage converted into a product.
- Platform economics. Vendors and platform providers push the unified platform argument where economies of scale, integrated security, and predictable operating costs make buying or partnering more attractive for many use cases. Analysts recommend evaluating unified platforms carefully to avoid hidden DIY stack costs.
Practical playbook for leaders
- Inventory use cases and data. Map where you have unique datasets and where data quality will materially change outcomes.
- Score each use case with the matrix above. Be brutal on talent and operational readiness.
- Pilot fast. For borderline cases, run a buy-and-extend pilot where you integrate a vendor solution with your data in a secure sandbox. If performance and control are insufficient, move to build.
- Negotiate vendor contracts with lock-in and compliance clauses. If you buy, ensure exportability of models and agreed SLAs on governance.
- Consider co-development or licensing. If you have unique data but lack scale, partnering with a large platform can monetize your data while sharing costs. Bayer’s model catalog approach is a good example.
How to write the internal memo that ends the debate
Frame the memo around capability arbitrage, not ideology. Ask three questions for each use case:
- Does our data create measurable advantage?
- Will workflow integration or compliance require ongoing tight control?
- Can we support the operational overhead for a production AI system?
If the answer is yes to the first two and yes or maybe to the third, prioritize building or a co-development model. If not, buy and move the freed capacity to the next highest-arbitrage use case.
Closing note
The binary build-versus-buy debate is an old mental model for a new world. The smarter leaders treat AI decisions as capability allocation problems. Invest where your data, control, and regulatory posture create asymmetric advantages. For everything else, buy, integrate, and ship.
Click here to read this article on Dave’s Demystify Data and AI LinkedIn newsletter.