Self-hosting is often presented as the cautious choice, but it is a trade, not an upgrade: you exchange a subscription and someone else's uptime for a server you are now responsible for at 3am. Both options are legitimate — the mistake is choosing without pricing the side you will actually pay.
Data residency is a requirement, not a preference
If a legal, sector or client requirement says workflow data must stay in a given jurisdiction, that requirement decides the platform and the rest of the comparison is commentary. Check it before evaluating features: a processor list and international-transfer answer is far cheaper to produce before a workflow exists than after it is running in production.
Self-hosted still means operated
A self-hosted platform that nobody patches, backs up or monitors is a worse outcome than the managed cloud it replaced — the control was theoretical and the risk is real. Budget the operating work explicitly, or take the managed option honestly. If neither is comfortable, a managed deployment inside your own cloud account is the middle ground worth pricing.
01Real scenario
A document workflow handles regulated client data and a high execution volume. Managed cloud is faster to launch; self-hosting may satisfy residency and cost requirements but transfers uptime, patching and recovery to the team.
02Inputs and evidence
List data categories, processor restrictions, network access, execution volume, retention, recovery objectives and the people available for operations. Obtain the managed vendor's processor list, regions, limits and export terms.
03Actions to test
Run the same representative workflow in a managed environment and a hardened self-hosted test. Exercise credential rotation, upgrade, restore and queue recovery rather than comparing only a successful happy-path execution.
04Security and controls
Use isolated service accounts, encrypted backups, monitored queues, restricted administration and tested restore procedures. A managed option still needs scope control and audit logs; self-hosting still needs documented processors for external AI calls.
05Pilot KPI
Compare total monthly cost, operator hours, recovery time, failed executions, latency and evidence required for compliance review. Model the next volume tier and at least one platform upgrade.
06Errors and human hand-off
Reject self-hosting if nobody owns patching and restore tests; reject managed cloud if residency, network or retention requirements cannot be met contractually. Escalate outages through a named operator, not an unowned shared inbox.