The short answer
Build in-house when the workflow embodies your team's judgment, changes often, and someone is genuinely assigned to operate it. Choose managed operations when the workflow is stable and repetitive, when its failure would be silent and expensive, or when operating it well is not the work your team exists to do.
Either way, two assets must remain yours: the authority model (what the system may do, what needs approval, what it must never do) and the run record. A model you cannot state and a history you cannot read are the real risks; the build location is secondary.
Decision criteria
- How many workflows are there, and how often does each one change?
- If a workflow failed silently tonight, who would notice, and when?
- Where does the authority model live: a document, a system setting, or one person's memory?
- What record could you show a client, an auditor, or yourself in six months?
- Is operating automation a core skill of this team, or a means to an end?
- What does a day of this workflow not running actually cost?
Side by side
| Dimension | In-house automation | Managed agent operations |
|---|---|---|
| Build | Your team designs and builds | Built with you; operated by the provider |
| Monitoring | Your responsibility, and often nobody's assigned job | The operator's standing job, with agreed reporting |
| The 2 a.m. failure | Found when someone notices, often by its effects | Found by observability; handled under agreed rules |
| Authority and approvals | As good as your team's documentation discipline | Defined in the operating agreement and enforced in the system |
| Knowledge risk | Concentrated in whoever built it | Held by the operator, against records you can inspect |
| Cost shape | Staff time plus tooling; spiky around incidents | Predictable operating fee; incidents are the operator's problem |
Tradeoffs that decide it
In-house wins on control and learning: the team that operates its own automation understands its business more deeply, and can change a workflow the same day the need appears. Managed operations win on continuity: the workflow does not depend on one person's memory, calendar, or continued employment, and its record is kept by someone whose job is keeping it.
The deciding question is what the workflow is. If it differentiates the business, own it and staff it properly. If it is infrastructure the business relies on, treat reliability as the feature, and be honest about whether the team will still be watching it in month nine.
Workflow observability servicesOperating the agent after the build
The build is the easy half. An agent that works on the day it ships still needs someone to watch it, and that watching is the durable cost. The question is not whether the agent can do the task. The question is who notices when it does the task badly, who can change what it does, and who is responsible when it acts on bad input.
Observability is the first requirement, and it is a record, not a dashboard. For every run, keep the inputs, the actions taken, the outcomes, and the exceptions, written so a person can read them later. The run log is how a team learns whether the agent is improving or drifting, and it is the evidence that supports a decision to expand what the agent is allowed to do. An agent without a run log is an agent nobody can defend.
Human review is the second requirement. Some actions need a person before they happen — anything that spends money, contacts a customer, or changes a public record. Some actions need a person after they happen — a sample of routine runs, reviewed on a schedule, so the team sees the quality of the work rather than only the exceptions. The review points are set when the agent is designed, not added after the first mistake.
The operating cost is the sum of the runs, the review time, and the exceptions. A managed operation carries that cost as a service: the provider watches the runs, handles the exceptions, and reports the outcomes. An in-house operation carries it as staff time. Either way, the cost exists, and the decision is about who carries it — not whether it can be avoided.
The review points are also the off switch. When the run log shows a pattern of bad actions, the team that can see the log can also stop the agent, narrow what it is allowed to do, and restore the previous process while the cause is found. An agent that cannot be stopped is not an agent. It is a liability.
Common questions
Do we lose control with managed operations?
Control is the authority model: what may run, what needs approval, what never runs. If that is not explicit, you do not have control in either model; if it is, managed operation executes your model rather than replacing it.
What should stay in-house?
Workflows that embody your judgment and change weekly. Hand over the stable, repetitive ones first; they benefit most from being watched and least from being rebuilt.
What record should exist either way?
For every run: inputs, actions taken, outcomes, and exceptions, written so a person can read them later, and kept with the business rather than only inside a tool.