Build, buy or wrap: three ways to add an AI capability
Build on a model API, buy a vertical product, or turn on a feature in software you already pay for. The three paths compared on cost curve, EU data residency and the switching cost nobody quotes.
Contents
An AI capability arrives by one of three routes. You build on a model API, you buy a vertical product that already does the job, or you turn on a feature inside software you already pay for. Capability rarely decides it. Cost curve, data residency and exit cost do.
Key takeaways
- A system matching GPT-3.5 on MMLU cost $20.00 USD per million tokens in November 2022 and $0.07 USD by October 2024.
- Building has the lowest unit cost and the highest fixed cost, because retrieval, evaluation and fallbacks become your responsibility.
- European data residency is now documented configuration: OpenAI opened it on 5 February 2025, Microsoft on 6 November 2024.
- Wrapping a feature you already own is cheapest to start and dearest to leave, because the workflow stays inside the vendor.
- Most Spanish companies decide this without precedent: 12.4% of firms with 10 or more employees used AI in the first quarter of 2024.
What the three paths actually are
Build means calling a model API and owning everything around it: prompts, retrieval, evaluation, cost control, logging, fallbacks. The system is yours.
Buy means a vertical product where AI is the product — contract review, ticket deflection, demand forecasting. The vendor owns the model choice and the upgrade path.
Wrap means switching on a feature inside software you already run. No new supplier, usually a per-seat line on an invoice you already approve. These are not three levels of ambition but three ways of distributing the ongoing work, and the questions worth asking before an AI deployment apply to all three.
How the cost curve behaves
Building looks expensive at the start and cheap at volume. In April 2025 GPT-4.1 listed at $2.00 USD per million input tokens and $8.00 USD per million output. Claude 3.7 Sonnet, announced on 24 February 2025, lists at $3.00 USD and $15.00 USD.
Those figures are small next to a single engineer-month, the real cost of building. The trap runs the other way too: per-seat pricing scales with usage, not with value.
Running a model scoring 64.8 on MMLU, the level of GPT-3.5, cost $20.00 USD per million tokens in November 2022. By October 2024 it cost $0.07 USD, a reduction of more than 280 times in roughly 18 months (Stanford HAI, April 2025).
What you build gets cheaper to run on its own. What you rent gets repriced when the vendor decides.
Where the data sits, and who can prove it
For a company operating in Spain this is usually the first hard constraint, and it is where the answer changed most recently.
OpenAI launched European data residency on 5 February 2025 for the API platform, ChatGPT Enterprise and ChatGPT Edu. Microsoft made Azure OpenAI Data Zones generally available for the EU on 6 November 2024. Anthropic distributes Claude through Amazon Bedrock and Google Cloud Vertex AI, so residency follows the region you deploy in.
OpenAI announced European data residency on 5 February 2025. Eligible API projects are processed in region with zero data retention, and ChatGPT Enterprise and Edu content is stored at rest in the region (OpenAI, 5 February 2025).
The difference between paths is who demonstrates it. Build and you hold the evidence. Buy and you inherit a sub-processor list. Wrap and you inherit whatever the suite decided, alongside the compliance dates the EU AI Act puts in your calendar.
Switching cost is the number nobody quotes
Switching cost is not the migration invoice. It is the work accumulated inside the option you chose: prompts tuned to one model's behavior, evaluation sets built for one output format, workflows that assume a record structure.
Building has the lowest switching cost at the model layer and the highest at the team layer. Swapping APIs is a week of evaluation if you kept your test set. Losing the people who understand the retrieval pipeline is not.
Wrapping inverts it. Nothing is easier to turn on and harder to unpick, because the feature lives inside the system of record. The EU Data Act, in force since 11 January 2024 and applicable from 12 September 2025, requires providers of data processing services to remove obstacles to switching. It covers infrastructure, not the tuning you did on top.
When each path stops making sense
Wrapping stops when the feature becomes the workflow rather than a convenience inside it. From then on you depend on a roadmap you do not influence.
Buying stops when the product does something your data makes unusual, or when per-seat cost crosses the engineering it replaces. Both thresholds are calculable before signing, and public co-funding rarely changes them. The eligible categories in what Spain's Kit Digital actually funds are written per solution, not per technology.
Building stops when nobody owns it. A system without an evaluation set, a cost dashboard and a named engineer is a subscription with worse uptime. The work in what LLMs in production actually require does not disappear because a budget was approved once.
FAQ
Which path is cheapest for a mid-sized company in Spain?
Wrapping, for the first year, and usually not after that. Per-seat features add a predictable line to an existing invoice and need no integration. The cross-over arrives once usage becomes routine: seat pricing grows with headcount, token pricing has been falling.
Does European data residency mean the model was trained in Europe?
No. Residency covers where requests are processed and content is stored, not where the model was built. OpenAI's February 2025 announcement describes in-region handling and zero retention for eligible API projects. It states separately that business and API data is not used for training unless a customer opts in.
How do we test a vertical AI product before committing?
Hold out a labeled sample of your own records, run it through the product, and score the output yourself. Ask in writing for the sub-processor list and the deployment region. A vendor that cannot answer both inside a week has answered the second one.
Decide in this order. Residency first, because it removes options rather than ranking them. Then switching cost, because it compounds. Unit price last, since it is the only one that improves on its own. Most mid-sized companies should wrap what is already paid for, buy where a vertical product does something specific, and build only where the capability is the business. Within a year the useful question becomes which path left you able to change your mind, and AI features inside a CRM are a good place to watch it.
