Credit unions agent orchestration Escalation Paths | Moring
Escalation paths are the defined routes by which an orchestrated AI agent hands work to a human when it reaches the edge of its authority, confidence, or competence. At a credit union they are a governance control, not a support feature: the path determines who reviews what, how fast, and with what evidence attached.
Short answer
Escalation paths are the defined routes by which an orchestrated AI agent hands work to a human when it reaches the edge of its authority, confidence, or competence. At a credit union they are a governance control, not a support feature: the path determines who reviews what, how fast, and with what evidence attached.
How it works
Every orchestrated agent at a credit union has a boundary, and the escalation path is what happens at that boundary. A member asks something the service agent is not authorized to answer, a fraud agent scores a case as genuinely ambiguous, a lending agent meets a condition it cannot clear on its own. The quality of the handoff at that moment determines whether the member experiences one continuous institution or a jarring transfer, and it determines whether examiners see designed supervision or improvised improvisation.
Escalation should be designed as tiers, not as a single panic button. The first tier moves a case to the human who can approve the specific action the agent identified, such as a fee waiver within a defined limit. The second tier moves to a specialist for judgment the agent cannot make, such as a complex dispute. The third tier reaches an officer when the case carries reputational, compliance, or member-harm weight. After-hours coverage deserves explicit design rather than assumption: a member who reaches an agent at night should land in a queue with a named on-call owner and a promised callback window, not in an escalation that silently waits for the next business day while the member waits too. Each tier should define what evidence travels with the handoff, what response time the member is promised, and who is accountable for the outcome.
Orchestration platforms are commonly evaluated on exactly this dimension, listed among context, routing, interoperability, evidence, and cost as authority and human intervention. That framing is useful because escalation is an authority question before it is a workflow question. The institution decides, in policy, which decisions an agent may make alone and which require a human signature; the orchestration layer enforces that policy at runtime, checking whether each delegated action is expected, permitted, and reversible before the agent changes anything, and escalating when the answer is no.
Escalation design is also where agent honesty is built. An agent that over-escalates burns staff time and members' patience; an agent that under-escalates takes actions it should not. The fix is calibrated triggers tied to the specific workflow: dollar thresholds on transfers and fee decisions, defined conditions in lending, confidence thresholds on member intent, and hard rules that always escalate regardless of confidence, such as anything touching a deceased member's account or a potential elder financial exploitation case. These triggers belong in the agent registry entry, where they can be reviewed and updated.
The handoff record matters as much as the handoff itself. When an agent escalates, the receiving human should inherit the full context: what the member asked, what the agent did, what it found, and why it escalated. When the human resolves the case, that resolution and its reasoning should flow back into the same audit trail. Examination readiness depends on exactly this continuity; a common supervisory question is what happens when an agent hits an exception, and the strongest answer is a documented path with records at every step rather than a policy statement alone.
Escalation performance should be measured like any other control. Escalation rate by workflow, time to human response, resolution rate after handoff, and the share of escalations that were appropriate versus premature or missed all belong on the management report. The lifecycle guidance in SR 26-2 and the monitoring discipline of SR 11-7 both expect ongoing supervision of AI systems, and escalation metrics are the most direct window into whether an orchestrated agent's authority boundaries still match reality as the workflows and the models change.
For a lean credit union the practical version is simple to start. Define two tiers for the first orchestrated workflow, name who staffs each tier, attach full context to every handoff, and review the trigger calibration quarterly. Most institutions discover that a small number of well-designed escalation triggers cover the large majority of cases, and that the designed path quickly becomes faster for members than the informal one it replaced.
Frequently asked questions
What is the difference between an escalation path and a fallback message?
A fallback message ends the agent's involvement with an apology; an escalation path transfers the case to an accountable human with full context attached. At a credit union the distinction is governance: the path defines which tier receives the case, what evidence travels with it, the promised response time, and who owns the outcome. Escalation is one of the core evaluation dimensions for orchestration platforms, grouped under authority and human intervention, because it is where delegated authority returns to named people with records.
Which cases should always escalate regardless of agent confidence?
Cases where the cost of a wrong autonomous action is unacceptable no matter how confident the model appears. Typical always-escalate triggers at a credit union include anything touching a deceased member's account, suspected elder financial exploitation, requests to override a fraud hold, large or unusual transfers, and any action outside the agent's defined permission scope. Writing these hard triggers into the orchestration policy, rather than trusting the agent's judgment, keeps the most consequential judgment calls with accountable staff.
How should escalation performance be reviewed over time?
Track escalation rate by workflow, time to human response, post-handoff resolution rate, and the proportion of escalations judged appropriate versus premature or missed, then review those numbers on a regular management cadence. Rising escalation rates can signal model drift or stale triggers; falling rates with rising complaints can signal under-escalation. This review aligns with the ongoing monitoring expectations in SR 26-2 and the supervision tradition of SR 11-7, and it produces the documented oversight examiners expect between examinations.
What context should travel with an escalated case?
Everything the human needs to resume without asking the member to repeat themselves: the authenticated request, the conversation history, the actions the agent already took, the data it retrieved, the reason the trigger fired, and the account details relevant to the decision. After resolution, the human's decision and reasoning should flow back into the same audit trail. That continuity is what lets a credit union answer the supervisory question of what happened when an agent hit an exception with records rather than recollection.