Short answer: Buy commodity AI, such as copilots, meeting-notes tools and the AI features your software already ships. Build only where the process, the data and the approval rules are your own, and even then build on models you rent rather than training your own. Volume changes the answer more than almost anything else.
Most companies should buy commodity AI and build only where the workflow is their own. Buy the copilot, the meeting-notes tool and the AI features your existing software already ships. Build where the process, the data and the approval rules are specific to your business, and even then, build on top of models you rent rather than training your own. That middle option, configuring and composing on bought models and tools, is where most of the real decisions sit.
This guide sets out the three options, what the evidence says, the costs nobody puts in the proposal, a decision table and a short scoring sheet.
What are the real options when you build or buy AI?
“Build or buy” sounds like two choices. In practice there are three, and the middle one is the one most companies end up using.
1. Buy off-the-shelf. A SaaS product or copilot you subscribe to and configure: Microsoft 365 Copilot, ChatGPT Business, the AI features inside your CRM or accounting system, or a vertical tool built for one job. You get speed and someone else’s roadmap. You accept their workflow.
2. Build on rented models. You rent a model from a provider through its API and build the workflow around it: your documents, your integrations, your approval steps, your audit trail. The model is bought. The workflow is yours.
3. Build your own model. You train or heavily modify a model yourself. This is rarely sensible for a company of 20 to 500 people.
The analysts draw the same lines. McKinsey’s CIO and CTO guide calls them Taker, Shaper and Maker, and expects most companies to combine Taker for commodity services with Shaper for proprietary capability. It puts building a foundation model at tens or even hundreds of millions of dollars. Gartner’s deployment cost table splits the same range into five steps: consume commercial apps, embed model APIs in custom apps, extend models with your own data, customize them by fine-tuning, and build custom models from scratch.
What does the evidence say about build vs buy AI?
The direction is consistent: companies are buying more and building less from scratch.
Menlo Ventures surveyed 495 US enterprise AI decision-makers in November 2025. It found 76% of AI use cases were purchased rather than built internally, up from a 53/47 split the year before.
Andreessen Horowitz (a16z) surveyed 100 CIOs in its 2025 report and saw a marked shift toward buying third-party applications. In customer support, over 90% of respondents were testing third-party apps. It found companies discovering that internally developed tools were hard to maintain and often gave them no business advantage. Its 2026 survey of 100 Global 2000 companies found the shift continuing, even in knowledge management and workflow automation, where in-house builds had dominated.
Gartner expects that in 2026 AI “will most often be sold to enterprises by their incumbent software provider rather than bought as part of a new moonshot project,” as analyst John-David Lovelock put it in January. By September, it reported that lock-in, data-sovereignty and cost risks were not deterring buyers from those embedded features.
MIT NANDA’s GenAI Divide report found that external partnerships with learning-capable, customized tools reached deployment about 67% of the time, against about 33% for internally built tools. Employee usage was nearly double for externally built tools.
Read the last one carefully. NANDA’s “buy” is not shrink-wrapped software. It means procuring tools and co-developing them with vendors, which is closer to option 2 than to option 1. The sample is 52 organizations, outcomes are self-reported, and the authors warn that the correlation does not prove causation: companies that partner may simply have different risk tolerance or procurement skill. As an implementation company, we benefit from this finding, so weigh it accordingly.
Two more caveats apply to all of it. Most of this evidence comes from large enterprises: the a16z 2026 sample started at $500 million in revenue. And a16z noted one exception to the buy trend: regulated, high-risk industries such as healthcare, where privacy and compliance weigh more.
Build vs buy AI: the decision table
Run each AI use case through these questions. Most use cases land clearly on one side. If yours splits evenly, start by buying and revisit in six months.
Question | Lean buy if… · Lean build on rented models if… |
|---|---|
How specific is the workflow? | Buy: everyone does it the same way (notes, drafting, search) · Build: it runs on your rules, formats and exceptions |
How sensitive is the data, and where must it stay? | Buy: the vendor offers the residency, retention and contract terms you need · Build: you need to control exactly where data goes and what is logged |
How deep does the integration go? | Buy: it works inside one tool you already use · Build: it has to read from and write to several systems |
How much volume? | Buy: low or irregular volume · Build: high, steady volume where per-seat or per-case fees add up |
Does a person need to approve, with an audit trail? | Buy: no sign-off needed, or the tool supports yours · Build: you need a defined approval step and a record of who approved what |
How fast do you need it? | Buy: weeks · Build: you can wait one to three months for something that fits |
Who will own and maintain it? | Buy: nobody internal has time · Build: a named owner, plus a partner or team for upkeep |
How much lock-in can you accept? | Buy: switching later would be cheap · Build: the workflow is core and you want to swap models freely |
What is the 3-year cost? | Buy: seats plus setup stay below a build · Build: the build plus running costs beat fees at your volume |
Building your own model belongs in this table only if every answer points to build and you also have proprietary data at scale, a machine-learning team and a budget in the millions. For most of our readers, that is a no.
Data residency used to push companies toward building. It pushes less now. OpenAI, for example, has offered European data residency for ChatGPT Enterprise and its API since February 2025, with in-region processing and zero data retention for eligible API projects. Check the terms of whatever you buy, but do not assume residency rules out buying.
What does each option really cost over three years?
Gartner’s 2024 cost estimates show the spread. Consuming commercial apps cost roughly $100,000 to $200,000 upfront, with recurring costs of about $280 to $550 per user per year. Building custom models from scratch cost $8 million to $20 million upfront and $11,000 to $21,000 per user per year. These are enterprise-scale figures from mid-2024. Your numbers will be far smaller, but the ratio is the point.
Seat prices for the buy option are public. As of October 2026, Microsoft 365 Copilot Business lists at $21 per user per month on an annual plan for up to 300 users (with a first-year promotion at $18), on top of a qualifying Microsoft 365 plan. ChatGPT Business costs $20 per standard seat per month billed annually.
A worked example (hypothetical)
Take one workflow: matching supplier invoices against orders. Every number below is a hypothetical assumption for illustration, not a quote from any vendor or from us.
Item | Hypothetical 3-year cost |
|---|---|
Buy a vertical tool: €5,000 setup + €1.00 per invoice | 300 invoices a month: €15,800 · 2,000 a month: €77,000 |
Build on rented models: €25,000 build + €6,000 a year upkeep | Plus model and hosting costs of €150 to €400 a month: €48,400 at 300 · €57,400 at 2,000 |
At 300 invoices a month, buying wins by a wide margin. At 2,000, building wins, and it also gives you your own approval step and audit trail. Volume changes the answer more than almost anything else.
The hidden costs of buying
Seats nobody uses. MIT NANDA found only 40% of companies had bought an official LLM subscription, yet workers at over 90% of them used personal AI tools. In one case it describes, a corporate lawyer kept drafting in ChatGPT even though her firm had spent $50,000 on a specialized contract tool.
Prices that change. Promotional pricing often covers year one only. Usage-based fees grow with adoption.
Switching costs. CIOs told a16z that prompts and guardrails tuned for one provider make changing models an engineering project. In that survey, 37% of respondents used five or more models, partly to avoid lock-in.
Labels that oversell. Gartner estimates only about 130 of the thousands of vendors selling agentic AI are real, and expects over 40% of agentic AI projects to be canceled by the end of 2027.
The workflow still has to change. A tool does not redesign the process or train the team. You still pay for that in time.
The hidden costs of building on rented models
Maintenance. Google engineers showed more than a decade ago, in Hidden Technical Debt in Machine Learning Systems, that the model code is a small fraction of a real system. The rest is data pipelines, configuration and monitoring, and it needs upkeep.
Model changes. Providers retire and update models. Budget for re-testing against your past cases when they do.
Data readiness. Gartner predicts that through 2026, organizations will abandon 60% of AI projects not supported by AI-ready data. Its survey found 63% of organizations lack, or are unsure they have, the right data practices.
Dependence on whoever built it. Gartner predicts that by 2028, 70% of enterprises will abandon agentic AI built by vendor forward-deployed engineers, because of soaring costs and an inability to evolve it themselves. Analyst Mukul Saha: “FDE success starts with getting the engagement structure right, from scope and incentives to governance, ownership, and exit.”
The hidden costs of building your own model
Beyond the compute and the team, you own the model’s obligations. Under the EU AI Act, providers of general-purpose AI models have had obligations since 2 August 2025. The Commission’s guidelines say most fine-tuning does not make you a provider: that happens only when a modification uses more than a third of the original model’s training compute. Training your own model from scratch is another matter. This is general information, not legal advice.
A short scoring sheet
Score each question 0, 1 or 2. Then add them up.
Question | 0 points · 2 points |
|---|---|
Is the workflow specific to us? | Generic · Unique rules, formats or exceptions |
How many systems does it touch? | One · Three or more |
What is the volume? | A few cases a week · Hundreds a day |
Does a person need to approve, with a record? | No · Yes, always |
Is the data sensitive or regulated? | No · Yes, with strict residency or retention needs |
Does a good tool already cover 80% of it? | Yes · No |
Is there a named owner with time? | No · Yes |
Is this workflow part of how we compete? | No · Yes |
0 to 6: buy. Configure a tool you already have or a proven product. 7 to 11: start by buying, then compose. Use the bought tool and add integrations or an approval step around it. 12 to 16: build on rented models, inside the tools your team uses. If the owner question scores 0, fix that first: a build without an owner will stall.
Where does Postmodel sit, and when would we tell you to buy?
We sit in option 2. We build workflows on rented models inside the tools a team already uses, with a person approving where it matters, and then train the team and measure the result. Our projects start at €10,000 (from $12,000 in the US).
That means we have an interest in this answer, so here is when we would tell you not to hire us:
The workflow is generic. Meeting notes, email drafting and general research are solved problems. Buy a copilot.
Your existing vendor already ships it. If your CRM, ERP or document system has a feature that covers 80% of the need, turn it on first.
Volume is low. A few cases a week rarely justifies a build. A tool, or plain automation, will do.
You need it next week. Buy now. Revisit once you know what the tool cannot do.
If you hire someone to build, us included, take Gartner’s warning on dependence seriously. Agree before you sign who owns the workflow, the prompts and the test cases, make sure the model can be swapped, and plan the handover so your team can run it without the builder. As we argued in The model is no longer the advantage, the model is a commodity. The workflow is the asset, so it should belong to you.
Where to start
List your AI use cases and run each one through the table and the scoring sheet. Most will come out as buy. Keep the one or two that score high, and pick the first using our guide on how to choose your first AI workflow. For budgets, see what AI implementation costs in 2026.
Have a list of candidates to build or buy? Book a fit check. We’ll tell you which ones we’d buy, and whether any is worth a diagnostic.
Frequently asked questions
What are the options besides build or buy?
There are three. Buy off-the-shelf software or a copilot and accept its workflow. Build on rented models, where the model is bought and the workflow, integrations and approval steps are yours. Or build your own model, which is rarely sensible for a company of 20 to 500 people.
Are companies buying or building more AI?
Buying. Menlo Ventures found in November 2025 that 76% of AI use cases were purchased rather than built internally, up from a 53/47 split the year before. Most of this evidence comes from large enterprises, so read it as a direction for smaller companies, not a rule.
When should you not build a custom AI workflow?
When the workflow is generic, when your existing vendor already ships the feature, when volume is low, or when you need it next week. Meeting notes, email drafting and general research are solved problems. Buy a tool, then revisit once you know what it cannot do.
What are the hidden costs of building on rented models?
Maintenance of data pipelines, configuration and monitoring, retesting when providers retire or update models, getting the data ready, and dependence on whoever built it. Agree before you sign who owns the workflow, the prompts and the test cases. Plan the handover so your team can run it.
Does data residency rule out buying AI?
Usually not anymore. OpenAI, for example, has offered European data residency for ChatGPT Enterprise and its API since February 2025. Check the terms of whatever you buy, but do not assume residency rules out buying.
Related reading
Sources
McKinsey & Company. “Technology’s generational moment with generative AI: A CIO and CTO guide.” July 11, 2023. https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/technologys-generational-moment-with-generative-ai-a-cio-and-cto-guide
Gartner. “Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025.” Press release, July 29, 2024, including Figure 1, “Costs Incurred in Different GenAI Deployment Approaches.” https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025
Menlo Ventures. “2025: The State of Generative AI in the Enterprise.” December 2025. Survey of 495 US enterprise AI decision-makers, November 7 to 25, 2025. https://menlovc.com/wp-content/uploads/2025/12/menlo_ventures_enterprise_ai_report-2025-123125.pdf
Andreessen Horowitz. “How 100 Enterprise CIOs Are Building and Buying Gen AI in 2025.” Sarah Wang, Shangda Xu, Justin Kahl and Tugce Erten. June 10, 2025. https://a16z.com/ai-enterprise-2025/
Andreessen Horowitz. “Leaders, gainers and unexpected winners in the Enterprise AI arms race.” January 30, 2026. https://a16z.com/leaders-gainers-and-unexpected-winners-in-the-enterprise-ai-arms-race/
Gartner. “Gartner Says Worldwide AI Spending Will Total $2.5 Trillion in 2026.” Press release, January 15, 2026. https://www.gartner.com/en/newsroom/press-releases/2026-1-15-gartner-says-worldwide-ai-spending-will-total-2-point-5-trillion-dollars-in-2026
Gartner. “Gartner Forecasts Worldwide AI Spending to Grow 49.5% in 2026.” Press release, September 16, 2026. https://www.gartner.com/en/newsroom/press-releases/2026-09-16-gartner-forecasts-worldwide-ai-spending-to-grow-49-point-5-percent-in-2026
MIT NANDA. “The GenAI Divide: State of AI in Business 2025.” Aditya Challapally, Chris Pease, Ramesh Raskar and Pradyumna Chari. July 2025. Preliminary findings. https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf
OpenAI. “Introducing data residency in Europe.” February 5, 2025, updated January 16, 2026. https://openai.com/index/introducing-data-residency-in-europe/
Microsoft. Microsoft 365 Copilot for business, plans and pricing. Accessed October 4, 2026. https://www.microsoft.com/en-us/copilot/solutions/business
OpenAI. ChatGPT business pricing. Accessed October 4, 2026. https://openai.com/business/pricing/
Gartner. “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027.” Press release, June 25, 2025. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
D. Sculley et al. (Google). “Hidden Technical Debt in Machine Learning Systems.” NeurIPS 2015. https://papers.nips.cc/paper_files/paper/2015/file/86df7dcfd896fcaf2674f757a2463eba-Paper.pdf
Gartner. “Lack of AI-Ready Data Puts AI Projects at Risk.” Q&A with Roxane Edjlali, February 26, 2025. https://www.gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk
Gartner. “Gartner Predicts 70% of Enterprises Will Abandon Agentic AI Built by Vendor Forward-Deployed Engineering by 2028.” Press release, September 29, 2026. https://www.gartner.com/en/newsroom/press-releases/2026-09-29-gartner-predicts-70-percent-of-enterprises-will-abandon-agentic-ai-built-by-vendor-forward-deployed-engineering-by-2028
European Commission, Shaping Europe’s digital future. “Guidelines on obligations for General-Purpose AI providers” (FAQ). Last update November 11, 2025. https://digital-strategy.ec.europa.eu/en/faqs/guidelines-obligations-general-purpose-ai-providers