MA
← Selected systemsENGINEERING CASE STUDY

Support operations · governed actions

ResolveOps

A support-operations system that separates what should happen from how it should be explained, keeps business rules deterministic, requires approval when risk demands it, and treats external execution as a reconciled operational state rather than a chat side effect.

Separationresolution reasoning ≠ business policy ≠ action execution
Action modeltyped proposal · risk classification · human approval when required
Execution statesSUCCEEDED · FAILED · UNKNOWN with reconciliation
CORE QUESTION
What governed business resolution should occur?
Owns · Governed business-resolution workflows

Authority split

What should happen is separate from how it should be explained.

MODEL / LANGUAGEInterpret caseRetrieve evidenceDraft explanationprobabilistic
BUSINESS POLICYEligibilityCeilingsRisk + approvaldeterministic authority
EXECUTIONTyped actionIdempotencyReconciliationoperational state

System model

Request to outcome without collapsing policy into chat.

External action truth

Timeout is not the same as failure.

SUCCEEDEDFAILEDUNKNOWN

UNKNOWN triggers reconciliation before retry; otherwise a payment/account action can be duplicated.

Cross-product boundary

High-impact execution can cross into AgentGuard.

ResolveOpsapproved business action→AgentGuardcredential-bound execution

Engineering decisions

Operational correctness lives in the failure states.

DECISION 01

Separate resolution reasoning from policy

The model may reason about a case, but refund eligibility, ceilings, and irreversible actions remain deterministic business rules.

DECISION 02

UNKNOWN is a first-class execution state

An external timeout may have succeeded remotely; blind retry can duplicate financial/account actions.

DECISION 03

Approval is risk-based, not universal

Low-risk reversible work can remain efficient while high-impact actions retain human control.

Capability surface

Governed participation in support operations.

01Customer-context retrieval
02Approved knowledge retrieval
03Intent classification
04Resolution planning
05Deterministic business policy
06Typed action proposals
07Risk classification
08Human-in-the-loop
09Replay-resistant approvals
10Execution adapters
11Idempotency
12Reconciliation
13Outcome tracking
14Policy simulation
15Case replay
16Feedback loop
17AgentGuard integration

Benchmark surface

What the system is designed to measure.

MEASUREMENT CONTRACT

The useful unit is a correctly governed resolution, not a fluent response.

Policy-labeled case fixtures, action proposal evaluation, approval/replay tests, external-failure simulation, and outcome replay.
01Resolution ratecases completed without escalation
02Evidence coverageresolutions grounded in current sources
03Action accuracytyped proposals match policy
04Review ratehuman intervention by risk tier
05Execution successincluding unknown/reconcile states
06Cost / resolutionmodel + operational cost

Evidence

Publicly defensible operational boundaries.

SignalEvidenceInterpretation boundary
Business-policy boundaryRefund/cancel/credit/plan rules remain deterministicSupportedLanguage-model output can propose or explain; it does not become the policy authority.
Typed action proposalOperational intent becomes an explicit, reviewable action objectSupportedProposal is distinct from execution and can be blocked or require approval.
Execution safetyIdempotency + replay-resistant approvals + adapter boundarySupportedExternal timeouts are reconciled; UNKNOWN is not treated as safe-to-retry by default.
Policy simulationHistorical cases can be replayed before changing business rulesSupportedSimulation informs policy change; model feedback does not autonomously rewrite policy.
AgentGuard integrationApproved high-impact actions can cross into credential-bound executionSupportedResolveOps owns business-resolution governance; AgentGuard owns runtime execution integrity.

Explicit boundaries

ResolveOps is an operations system, not the entire support stack.

Not a helpdesk replacementNot the runtime authorization productNot a generic agent frameworkNot generic RAG