Three questions decide it: how sensitive the data is, how specific the workflow is, and how much engineering the team can carry.
Configure is the default. You still buy the platform; what you own is the workflow on top. Enterprise platforms let you wire custom workflows, retrieval, and guardrails onto a hosted model you do not own. For contract review and internal knowledge, this captures most of the value for far less maintenance.
Buy for commodity tasks. Transcription, translation, and first-pass summarization are mature, well-served categories. Building them in-house means a year rebuilding what you could license this week.
Build only for workflows that are both sensitive and differentiating. Building does not remove third parties: our own Bedrock tooling still runs on cloud infrastructure. It is the right call when off-the-shelf terms cannot meet your requirements on data use, retention, residency, and audit, and the workflow itself sets you apart. You still carry the full lifecycle burden.
Bloomberg Law reported in 2026 that for roughly 67 percent of large law firms, new-technology spending stayed under 2 percent of the 2025 budget. It counts firms, not in-house teams, but the signal carries.
Capacity decides the rest. ILTA's 2024 survey found generative AI use ranging from 20 percent at small firms to 74 percent at the largest. My read: that gap follows engineering capacity more than ambition.
Match the tool to the workflow: configure first, buy the commodity, build only what you cannot safely hand to anyone else.
Sources