State model
A durable saga record owns step status, retries, deadlines, compensation and custody. Source systems retain legal business state.
A failed standing order may involve funds, limits, timing, destination state or a platform incident, each requiring a different remedy.
The workflow diagnoses the failure from authoritative events, explains the customer impact and offers only safe, policy-permitted recovery actions.
Organisations, systems and operating conditions are intentionally anonymised and recomposed. The design demonstrates engineering and banking-domain reasoning; it does not represent a named client estate, vendor product or measured production result.
Typed commands, state changes and receipts cross an event spine without surrendering domain ownership.
Permitted workThe AI layer may diagnose, prepare and coordinate. Core servicing effects, remediation and complaint outcomes remain policy-controlled actions.
Consistency ruleKeep one durable case state and explicit custodian while source effects reconcile asynchronously.
Hard boundaryThe model is not a system of record, identity provider, policy authority or proof that an external effect occurred.
The journey can change account, payment, case or document state across systems that do not share one transaction boundary.
A new payment cannot be proposed until the original outcome and duplication risk are known.
A deterministic outer workflow contains model-led work inside typed, observable calls. Dashed messages remain proposals until policy or a human grants authority.
A durable saga record owns step status, retries, deadlines, compensation and custody. Source systems retain legal business state.
Mutating calls pass through typed capability contracts. Events notify downstream services, but command acceptance and effect readback remain distinct.
Use idempotent local transactions and eventual cross-system consistency. Unknown outcome is a first-class state, not an exception hidden by retry.
These roles are deliberately vendor-neutral. Each can be independently owned, versioned and replaced.
Runs a versioned state machine with explicit waits, deadlines, retries, compensations, human tasks and terminal states.
Exposes typed, allow-listed capabilities with schema validation, delegated identity, idempotency, timeout semantics and effect readback.
Classifies unknown outcomes, replays idempotent work, runs compensations and restores from the last verified checkpoint.
Builds time-qualified projections from source events and reconciliations without becoming the legal system of record.
Appends request, versions, policy result, model proposal, approval, action receipt, readback, correction and custody events under one correlation key.
Evaluates identity, purpose, capability, amount, risk tier and policy version; returns allow, deny, step-up or human-review with reasons.
Durable records carry provenance, authority, effect and custody without turning a transcript into an uncontrolled memory store.
Keep one durable case state and explicit custodian while source effects reconcile asynchronously.
The selected design is not universally superior. It is the safer fit for this boundary and failure cost.
Keep consequential state transitions deterministic and use models inside bounded steps.
Let the model choose the complete path and recovery sequence.
Cost acceptedThe shell reduces flexibility, but makes deadlines, retries, permissions and recovery testable.
Use direct calls for narrow same-domain reads; mediate cross-domain, reusable or mutating capabilities.
Force every call through one integration pattern.
Cost acceptedMediation adds a hop and ownership overhead; direct calls increase coupling and duplicate controls.
Use local commits, idempotent steps, compensations and an unknown-outcome state.
Attempt one atomic transaction across independent banking systems.
Cost acceptedSagas expose temporary inconsistency and require recovery logic, but fit systems that cannot share one transaction boundary.
Use event-fed projections for scale and direct readback for consequential effects.
Fan out to all systems of record for every interaction.
Cost acceptedRead models introduce lag and reconciliation work, but reduce source load and make cross-system views feasible.
Use append-only events plus a rebuildable current-state projection.
Overwrite the case row with its latest status.
Cost acceptedReplay and storage are more complex, but point-in-time reconstruction and correction lineage remain possible.
Compile stable decision logic and retain retrieval for explanation and residual ambiguity.
Ask a model to interpret the source document for every request.
Cost acceptedRule compilation needs controlled change, but creates repeatable decisions, regression tests and clear exceptions.
Retries are bounded by knowledge of business effect; unknown outcome remains visible, owned and independently reconciled.
Actual thresholds belong to accountable service owners. The design exposes the equations and observables that those owners must baseline.
peak_action_rate = requests_per_second x steps_per_requestreconciliation_load = timed_out_actions + mismatched_readbacksrecovery_capacity = open_unknown_outcomes / owner_resolution_rateTransfer only case-relevant evidence between teams and retain communication history according to the case class.
A design is production-ready only when teams can prove what happened, recover it and change it safely.
partial completion and retry test
timer and deadline preservation
case handoff acceptance
customer communication readback
Failure event, account state, incident status, remedy options, confirmation, action receipt and readback.
The customer or an authorised service agent approves any replacement payment.
Automate reversible steps first. Expand authority only after duplicate, timeout, compensation and operator-recovery tests pass under peak load.
Servicing operations, product, complaints, legal and records owners set the operating controls.