The short answer
Buy an off-the-shelf copilot when the workflow is horizontal, the data lives in the vendor's tools, and your differentiation is not in the workflow. Extend an existing platform (Salesforce Einstein, ServiceNow Now Assist, M365 Copilot) when 80% of the data is already there and the gap is configuration, not capability. Build custom when the workflow is core differentiation, spans systems no vendor consolidates, or requires governance that vendor SLAs do not cover. Most enterprises will run all three patterns — the trap is forcing one model on every problem.
The three patterns
Buy (off-the-shelf copilot — ChatGPT Enterprise, Microsoft 365 Copilot, Glean, Notion AI), Extend (platform-native AI — Salesforce Einstein, ServiceNow Now Assist, HubSpot Breeze, NetSuite AI), and Build (custom agentic system on your stack with MCP, LangGraph or equivalent, your own evals and governance).
Five axes for the decision
Score each candidate workflow on five axes; the pattern of high and low scores tells you the answer.
- ›Workflow specificity: horizontal (buy) vs vertical (extend) vs your specific competitive logic (build).
- ›Data locus: in one vendor's tools (buy/extend) vs spanning many systems (build).
- ›Governance ceiling: standard SaaS controls (buy/extend) vs custom audit, retention, or jurisdiction needs (build).
- ›Integration depth: read-only (buy) vs write-back (extend) vs multi-system orchestration and tool use (build).
- ›Differentiation potential: parity with peers (buy/extend) vs durable competitive advantage (build).
When buy is the right answer
Knowledge search, meeting summarization, draft assistance on documents that already live in the vendor's stack, code completion. The workflow is horizontal, the data is local to the tool, governance needs are standard, and nobody wins by building their own ChatGPT.
When extend is the right answer
CRM hygiene where 80% of the data already lives in Salesforce. ITSM ticket triage where ServiceNow already holds the source of truth. Email assistance inside M365. Extending platform-native AI is faster than building and avoids the integration tax — as long as you are not bumping against the platform's governance or extensibility ceiling.
When build is the only answer
Workflows that span 3+ systems no vendor consolidates. Industry-specific reasoning that no horizontal copilot encodes. Governance requirements (data residency, custom audit, regulator-grade explainability) that exceed vendor SLAs. And — critically — workflows where the way you do the work is differentiation. Outsourcing that to a vendor commoditizes your edge.
The hybrid reality
Real enterprises run all three patterns. Knowledge search via Glean, CRM hygiene via Einstein, and the bespoke proposal-to-SOW agent built custom on top of both. The mistake is mandating one pattern across the portfolio — it always produces either over-investment in commodity work or under-investment in differentiation.
Key takeaways
- Score on workflow specificity, data locus, governance ceiling, integration depth, and differentiation potential.
- Buy commodity work, extend platform-native workflows, build differentiation.
- Hybrid portfolios beat single-pattern mandates every time.
- If you are tempted to build a horizontal copilot, buy one instead and free the engineers for differentiation.
Frequently asked questions
Should we wait for vendor platforms to catch up before building?+
For commodity workflows, yes. For differentiating workflows, no — by the time vendors ship parity, your competitors who built early have a 2-year head start on the data and evaluation harness.
How do we evaluate a platform-native AI feature?+
Pilot it on the exact workflow you are considering, not on the demo workflow. Vendor demos are tuned for sales; your workflow tests integration depth, governance, and edge-case handling.
What about open-source agent frameworks?+
Open-source frameworks (LangGraph, CrewAI, AutoGen) are the substrate for the 'build' pattern, not a fourth pattern. Choose the framework that fits your team's language and runtime, not the one with the loudest GitHub stars.