The thesis. Build, buy or partner is not a technology choice. It is a decision about where the organization intends to be different. Build artificial intelligence (AI) capability only where your data and process knowledge create an advantage a competitor cannot purchase; buy where the capability is becoming a commodity; partner where speed and scarce skills matter more than ownership.
Why the old rules need updating
Classic build-or-buy logic assumed that software, once built, kept its value and needed modest upkeep. AI breaks both assumptions. The underlying models improve and get cheaper on a cycle measured in months, and the systems around them need constant evaluation, monitoring and retraining.
The pace is visible in the data. Stanford’s 2025 AI Index reports that the performance gap between leading open-weight and closed models narrowed from 8 percent to 1.7 percent on some benchmarks in a single year. It also reports that the inference cost of a system performing at the level of OpenAI’s GPT-3.5 fell more than 280-fold between November 2022 and October 2024 [1]. When a capability becomes that much cheaper and more widely available that quickly, owning it stops being an advantage on its own.
Enterprise buyers have noticed. In a survey of 100 chief information officers published in June 2025, Andreessen Horowitz reported a marked shift over the previous twelve months towards buying third-party AI applications. Respondents said internally developed tools were difficult to maintain and frequently did not give them a business advantage [2]. The same survey found 37 percent of respondents using five or more models, up from 29 percent a year earlier [2].
Four tests that decide the answer
I use four tests. Each pulls the decision in one direction, and the answer comes from weighing them together, not from any single one.
1. Differentiation
Ask whether customers would notice, and pay, if this capability were better than a competitor’s. Pricing logic, underwriting judgment, a proprietary design method or a distinctive service experience can be differentiating. Document summarization, transcription and general drafting rarely are. Spending scarce engineering capacity on undifferentiated capability is a strategic cost, even when the build itself succeeds.
2. Data advantage
Models are increasingly available to everyone. Proprietary data, and the process knowledge to use it, are not. Gartner predicts that by 2027, organizations will use small, task-specific AI models at least three times more than general-purpose large language models, because the accuracy of general models declines on tasks that need specific business context [3]. Where your data and context are the source of accuracy, owning the resulting capability has real value. Where they are not, you are building what a vendor will sell more cheaply next year.
3. Time to value
Buying is usually fastest to a first result. Building is usually slowest. Partnering sits between them. But time to a demo is the wrong measure. The relevant measure is time to a redesigned workflow that moves a business number. McKinsey found that, of 25 attributes tested, the redesign of workflows had the biggest effect on an organization’s ability to see earnings before interest and taxes (EBIT) impact from generative AI. Yet only 21 percent of respondents reporting generative AI use said their organizations had fundamentally redesigned even some workflows [4]. A purchased tool dropped into an unchanged process delivers little, whichever route was chosen.
4. Lock-in and exit cost
Every option creates dependency. Building ties the organization to its own team and technical choices. Buying ties it to a vendor’s roadmap and pricing. Partnering ties it to a relationship. The question is whether the exit cost is known and acceptable. The same survey of chief information officers noted that the rise of agentic workflows is making it more difficult to switch between models [2]. Regulators are also acting on switching costs. The European Union’s Data Act will apply from 12 September 2025. It requires providers of cloud and other data processing services to support customers who switch, and it prohibits switching charges from 12 January 2027 [5].
The decision table
| Test | Build when | Buy when | Partner when |
|---|---|---|---|
| Differentiation | It is core to how you win | It is table stakes | It is core, but skills are scarce |
| Data advantage | Your data drives accuracy | Generic data is enough | The data is yours, the methods are not |
| Time to value | You can wait for depth | You need results this year | You need speed and depth |
| Exit cost | You can staff it for years | Exit terms are clear | Roles and ownership are contracted |
Most organizations will end up using all three routes. That is healthy, provided each choice is deliberate. The risk is accidental sprawl. In IBM’s 2025 study of 2,000 chief executives, half said the pace of recent investment had left their organization with disconnected, piecemeal technology [6].
Making partnership work
Partnering is the least understood route and the one most often structured badly. Three design choices decide whether it creates value or friction.
- Intellectual property. Agree up front who owns the models, prompts, evaluation sets and fine-tuned model weights built on your data. Ambiguity here becomes a dispute at renewal.
- Capability transfer. Decide whether the partner will run the capability indefinitely or build it and hand it over. Price and staff the engagement accordingly.
- Shared measures. Tie a meaningful share of the partner’s economics to the business outcome, not to hours or volume consumed. Aligned incentives do more than detailed statements of work.
From the perspective of strategic partnership development, the partnerships that last are the ones where both sides can state, in one sentence, what each gains that it could not get alone. If that sentence is hard to write, the structure is probably wrong.
Vendor choice follows strategy
None of this argues for or against any particular vendor or model. Most enterprises will rely on several, and they should. The point is that vendor choice follows from strategy. When the organization knows where it intends to be different, the rest of the portfolio becomes a procurement discipline: clear evaluation criteria, clear exit terms and regular review as prices and capabilities move.
Questions for the leadership team
- Which three AI capabilities, if they were better than a competitor’s, would customers notice and pay for?
- For each capability we plan to build, what proprietary data or process knowledge makes our version better?
- What is our exit cost, in time and money, for each major AI vendor or partner today?
- Who owns the models, evaluation data and fine-tuned weights created in our partnerships?
- Are we measuring each option by time to a redesigned workflow, or by time to a demo?
References
- N. Maslej et al., “The AI Index 2025 Annual Report,” AI Index Steering Committee, Institute for Human-Centered AI, Stanford University, Stanford, CA, Apr. 2025.
- S. Wang, S. Xu, J. Kahl and T. Erten, “How 100 enterprise CIOs are building and buying gen AI in 2025,” Andreessen Horowitz, Jun. 2025.
- Gartner, “Gartner predicts by 2027, organizations will use small, task-specific AI models three times more than general-purpose large language models,” press release, Apr. 9, 2025.
- A. Singla, A. Sukharevsky, L. Yee, M. Chui and B. Hall, “The state of AI: How organizations are rewiring to capture value,” McKinsey & Company, Mar. 2025.
- European Parliament and Council of the European Union, “Regulation (EU) 2023/2854 on harmonised rules on fair access to and use of data (Data Act),” Official Journal of the European Union, Dec. 2023.
- IBM Institute for Business Value, “5 mindshifts to supercharge business growth: 2025 CEO Study,” May 2025.