Implementation
Build vs Buy vs Compose
Across 2,839 classified AI deployments, 72% are built custom (pro-code), 9% buy off-the-shelf, and 10% compose with low-code. Compare the build approach for every use-case type.
Dataset details
- Revision
- dsr-fc53bf0883afdfe4
- Canonical records
- 3,963
Show metrics
- Cases classified
- 2,839
- Build
- 72%
- Buy
- 9%
73% of 3,887
custom / pro-code
off-the-shelf
The whole corpus
Most AI gets built, not bought.
72% of 2,839 documented deployments are engineered custom — only 9% adopt an off-the-shelf assistant, 10% compose with low-code, and 8% blend approaches.
230 use-case types · 2,839 of 3,887 cases classified (73%)
The trend
Is the mix shifting?
Share of classified deployments by publish quarter · Build + Buy + Compose + Mixed = 100%
Building is pulling further ahead. Build rose from 75% of classified deployments in Q1 2021–Q4 2021 to 84% in Q4 2025–Q3 2026.
Quarters with under 25 classified deployments aren't charted.
Where buying & composing win
Off-the-shelf buying and low-code composition concentrate in a handful of use-case types — Legal practice management is the most bought at 100%. Everything else still leans heavily custom.
Use-case types with at least 10 cases, ranked by each approach's share of classified cases.
Build vs buy × autonomy
Are agentic workloads built or bought?
Packaged (buy + low-code) share of each type's implementations vs. share that run agentic · Select a dot to inspect it.
The most-bought category, business process automation (62.7% packaged), runs only 20% agentic — while the most-agentic, automotive operations automation (83.8% agentic), stays mostly custom-built. Autonomy skews toward building.
+3 labels hidden for clarity. Hover or focus a dot for details.
Packaged share is over cases with a classified approach (~2 in 3); "agentic" reflects our per-case classifier, not vendor self-description. Showing the 20 highest-volume types with at least 25 cases and a dozen classified approaches. Packaged split at the corpus average (19.3%).
Who builds it
Which use cases still need an implementation partner?
Packaged (buy + low-code) share of each type's implementations vs. share of deployments naming a delivery partner · Select a dot to inspect it.
Clinical documentation leans on integrators the most: 13.8% of its deployments name a delivery partner, against 8.7% across the corpus. At the other end, risk assessment is mostly packaged (21.2%) and only 6.2% partner-led, so buyers largely roll it out themselves.
+2 labels hidden for clarity. Hover or focus a dot for details.
Partner-led means the public evidence names at least one delivery partner other than a cloud provider. First-party vendor posts under-report partners, so treat it as a floor. Packaged share is over cases with a classified approach. Showing the 20 highest-volume types with at least 25 cases. Splits at the corpus average (19.3% packaged).
By industry
Who buys vs builds
Each industry's Build / Buy / Compose / Mixed mix, ranked by buy + compose share
Legal leans hardest on ready-made AI — 35% of its 66 classified deployments buy off-the-shelf or compose with low-code — while Agriculture engineers 86% custom.
% = buy + compose share of classified deployments · industries with at least 25 classified.
ⓘ How this is measured
Each documented deployment is classified by how it was built: Build (custom / pro-code engineering — bespoke pipelines, SDKs such as Azure AI Foundry, Vertex AI or Amazon Bedrock), Buy (an off-the-shelf assistant or copilot adopted as the main solution), Compose (low-code / no-code tooling such as Copilot Studio, Power Platform or n8n), or Mixed when more than one applies.
Per-type percentages are shown over classified cases. Sorting by cases stacks each row's four approaches into one 100% mix bar; sorting by an approach (Build, Buy or Compose) tints that share column and switches the bar to a single left-anchored meter for it, so the list reads as a ranked staircase. Deployments whose write-up doesn’t describe the build approach are left unclassified and reported as the “% classified” coverage figure, not forced into a bucket.
The trend chart buckets classified deployments by publish quarter and plots each approach’s share of that quarter. Quarters with fewer than 25 classified deployments aren’t charted (the sparse early tail of old publish dates), and the in-progress quarter is marked with an asterisk while it fills. The “shifting?” comparison pools the first and last four completed quarters, so a single thin quarter can’t swing it.
“Who buys vs builds” groups classified deployments by each case’s industry and ranks industries by their combined buy + compose share — the fraction adopted off-the-shelf or assembled with low-code rather than engineered custom. Industries with fewer than 25 classified deployments are left out rather than charted on a thin sample.