Ask five things: what production ready means to them, how a silent failure would be caught, what happens to the workflow at ten thousand runs a month, what a change costs six months in, and whether this should be built at all rather than bought. Each page below answers one of those, with dates and sources on every claim.
Changing an AI workflow six months later lands in one of four buckets: configuration, a contained code change, a large code change, or a rebuild. A system designed for change puts most requests in the first two. A glued workflow puts them in the last two, because the logic lives in the connections rather than in code anybody can edit.
updated 2026-09-14·4 dated receipts
Production ready means the system keeps working when nobody is watching it. Concretely: tests that gate every change, monitoring that reaches a human, a documented failure path for every dependency, retries and rate-limit handling, secrets held outside the app, and a rollback somebody has actually used. A demo that works on a call is none of these.
updated 2026-09-14·5 dated receipts
AI automation fails silently because nothing was watching the part that failed. Most automations report success when a step returns anything at all, so an empty result, a rate-limited call and a model that answered wrong all look like a green run. The failure reaches you as a customer complaint weeks later, by which point the data is already wrong.
updated 2026-09-14·5 dated receipts
Glued workflows stop being cheap and start being fragile at the same time. 10,000 runs a month is about 330 a day, which crosses the per-task pricing tier on most automation platforms, exhausts free API quotas, and turns every retry into a billed event. The glue usually holds. What breaks is what the glue calls, and nothing is catching it.
updated 2026-09-14·4 dated receipts
Buy the tool when your workflow is the same as everyone else's: payroll, accounting, ticketing, email. Build when the workflow is the part of the business you would not change, and bending it to fit a product would cost you the thing you are good at. Most operations are a mix, and the honest answer is usually both.
updated 2026-09-14·3 dated receipts
AI agentDry runHallucinationHuman in the loopObservabilityOrchestrationp95 latencyRate limitingRetrieval augmented generationRollbackTotal cost of ownershipWebhook
Want the arithmetic on your own numbers instead of an argument?
Run the cost model →