Evidence note 03 · Build vs buy
The AI market is maturing. So why is build winning?
Custom engineering accounts for 71% of classified deployments. Its lead has grown as the market has filled with products.
The enterprise AI shelf has never been fuller. Copilots, agent platforms, workflow products and low-code builders now cover almost every business function. The deployment evidence keeps moving toward custom engineering. Build rose from 73% in Q2 2023 to Q1 2024 to 81% in Q3 2025 to Q2 2026.
My read is simple. The market is packaging interfaces faster than companies are standardizing work. A vendor can ship the assistant. The difficult part still sits inside process design, permissions, data, exceptions and accountability.
Custom-built
71%
1,707 of 2,399 classified deployments
Build share over time
+8 pp
73% in Q2 2023 to Q1 2024; 81% in Q3 2025 to Q2 2026
Highest agentic share
84%
Automotive operations automation; 74% custom-built
The long view
More products, more building
The direction matters more than any single quarter. Pooled across the first four completed quarters, custom build represented 73% of classified deployments. The latest four completed quarters reached 81%. That is a 8 point move toward custom engineering.
Buying had a real window. It peaked at 28% in Q3 2024. The orange line then fell as build recovered and mixed approaches started to appear more often. The catalog grew throughout this period, yet packaged adoption stayed concentrated in a few repeatable categories.
Build approach by quarter
Is the mix shifting?
Share of classified deployments by publication quarter
Quarters need at least 25 classified deployments.
What build means now
Build is an ownership choice
Build here means custom, pro-code engineering. It can include a bought model, managed search, a commercial agent framework and several SaaS APIs. The classification follows the main implementation path. Teams still build when their advantage comes from how those parts are connected.
This also explains why a bigger product market can produce more custom deployments. Better components lower the cost of assembling a tailored system. They make building accessible to more teams, especially when the workflow crosses several systems or carries company-specific rules.
Build vs buy x autonomy
Autonomy pulls the work in-house
The second chart shows where the ownership pressure comes from. Business process automation is 60% packaged and 20% agentic. It sits close to the shape vendors can productize: a recognizable process, a bounded job and familiar handoffs.
Automotive operations automation reaches 84% agentic. Its implementations remain 74% custom-built. As software gains authority to plan, call tools and move work forward, companies keep tighter control over the orchestration layer.
+3 unlabeled where space ran out — hover the dots: Predictive maintenance, Agriculture optimization, Cloud migration.
Where products win
Buying works when the process has settled
Legal practice management has the strongest off-the-shelf share in the current directory at 67%. Public sector automation leads low-code composition at 70%. These categories have common inputs, recognizable outputs and enough shared practice for a product to travel between customers.
The timing is the useful part. Teams buy once a workflow becomes legible to the market. They compose when the workflow is common and local variation still matters. They build when the operating model itself carries the value, or when autonomy raises the cost of handing control to a generic product.
Every use case runs on its own clock. Product coverage catches up as interfaces stabilize and exceptions become familiar. New capabilities then open another custom phase. The aggregate line can stay build-heavy even while individual categories gradually become easier to buy.
Conclusion and outlook
The build lead will move, not disappear
A clean decision starts with the workflow. How stable are the handoffs? How much company context shapes the answer? How expensive is an exception? Who carries accountability when the system acts? Those questions age more slowly than a vendor comparison.
Revisit the answer as the market moves. A custom system can migrate toward packaged components. A bought assistant can grow into a composed workflow. An agentic layer may bring a mature process back into custom engineering because the ownership boundary changed.
The current evidence gives custom build a clear lead. It also shows the path products take when they eventually catch up: bounded work, repeatable decisions and interfaces that many companies share.
I expect this cycle to keep repeating. Vendors will package yesterday's custom patterns, while new models and higher autonomy create fresh integration work. The build lead may narrow in mature categories, then reappear wherever the workflow changes faster than the product.
Method
How the charts were built
Each deployment receives one implementation approach: Build, Buy, Compose, Mixed or Unknown. Shares exclude Unknown cases. The trend groups deployments by original publication quarter and requires at least 25 classified cases per quarter. The comparison pools the first and last four completed quarters in the chart.
The autonomy map covers the highest-volume use-case types with at least 25 cases and 12 classified approaches. Packaged share combines Buy and Compose. Agentic share comes from the per-case classifier. Public case studies favor successful deployments, and publication choices change over time. The charts describe visible deployment evidence.
Exact build approach values
| Approach | Cases | Share |
|---|---|---|
| Build | 1,707 | 71% |
| Buy | 233 | 10% |
| Compose | 291 | 12% |
| Mixed | 168 | 7% |