EDI fails quietly. There's no user staring at an error page — a document just doesn't arrive, or arrives wrong, and the first symptom is a buyer's phone call or a deduction three weeks later. Error handling is therefore the difference between an EDI pipeline and an EDI liability. Here's the design that works, layer by layer.
Layer 1: 997 reconciliation — detecting silence
The 997 functional acknowledgment is EDI's delivery receipt: the partner's system confirming your transmission arrived and parsed. Two disciplines make it useful:
Send yours instantly. Every inbound document gets a 997 back within minutes, automatically. Partners monitor this; test programs grade it.
Reconcile theirs continuously. The high-value alert is not a rejection — it's absence. Every outbound document opens an expectation: a 997 within the partner's window (minutes to hours). A document still unacknowledged at the deadline may never have arrived — an AS2 hiccup, a VAN queue, a certificate expiry — and no one on the partner's side will ever tell you, because they don't know the document existed. Silence detection is the single most valuable monitor in EDI: an unacknowledged 856 caught within the hour is a retransmission; caught at the dock, it's a chargeback.
A rejected 997 (syntax failure) should be rare with pre-validated documents — treat any occurrence as a map bug, not an operational event.
Layer 2: validation and quarantine — catching the wrong
Syntax is table stakes; the real filter is business validation against each partner's companion guide, applied before transmission outbound and before ERP entry inbound. Documents that fail don't proceed and don't vanish — they land in a quarantine queue with three properties:
- Visible — a dashboard state, not a log line. Quarantine that nobody watches is deletion with extra steps.
- Diagnosed — the queue entry says what failed: unmapped item, missing required REF, hierarchy shape, serial format. "Rejected" without a reason just moves the investigation cost downstream.
- Owned and aged — every entry has a resolver and a clock. An 850 in quarantine is a customer order not being fulfilled; its age is an SLA.
Inbound quarantine deserves equal respect. A PO with an unmapped item that limps into the ERP as a broken order is worse than one held at the gate — the gate version gets fixed before the warehouse acts on garbage. And when partners reject at the business level, the 824 application advice is their quarantine notice to you: parse it, route it, treat it like a P1.
Layer 3: alerting that respects humans
Error handling dies by two failure modes: silence, and noise. The design rules:
- Alert on states, not events. "Document unacknowledged 2 hours past window," "quarantine item aged past 4 hours," "zero 850s received from a partner who orders daily." That last one — expected-volume monitoring — catches whole-connection failures that per-document checks miss.
- Escalate by business impact. An unacknowledged ASN for freight arriving tomorrow outranks a quarantined 832 catalog line. Severity comes from document type and clock, not arrival order.
- Weekly patterns, not just daily fires. The same item failing cross-reference every Monday is an engineering ticket. Error handling that only firefights never shrinks its own workload.
Layer 4: retry discipline — recovering without new damage
Retransmission is where recovery creates fresh incidents. Three rules:
- Idempotency first. Before resending, determine whether the partner got the original (997 status, transport receipts like AS2 MDNs). A duplicated 850 acknowledged twice can become two shipments.
- Control numbers by the book. Retransmissions need correct ISA/GS control-number handling per the partner's duplicate-detection expectations — some partners key on control numbers, some on document content.
- Cap and escalate. Automated retry handles transient transport failures (a down endpoint, a full mailbox). Anything that fails twice with the same error is systematic — a human problem wearing a retry loop.
What good looks like
A healthy pipeline reads like this: documents validate before they move; every transmission is acknowledged or alarmed within the hour; failures sit in a diagnosed, owned queue with SLAs; volume anomalies page someone; and the weekly error review closes root causes instead of celebrating catches. Most of that is platform behavior, not heroics — EDISQ ships it as the default: automatic 997s both directions, partner-rule validation before transmission, quarantine with diagnosis, and monitoring across every connection, all inside the per-document price. Details on the product page.
FAQ
What does a 997 actually confirm?
That your document arrived and parsed against the X12 syntax — nothing more. It doesn't mean the partner accepted the business content. Business-level rejection arrives later as an 824 or a human email.
What is 997 reconciliation?
Continuously matching every outbound document against the acknowledgment it should receive within the partner's window. The alert condition is absence: a document with no 997 may never have arrived, and nobody will tell you.
What is a quarantine queue?
A holding state for documents that fail validation — inbound or outbound — so they're visible and fixable instead of lost or transmitted broken. Every document in quarantine has an owner and an age.
How do you retry EDI safely?
Idempotently. Retransmit with correct control-number handling and check whether the partner received the original before resending — duplicated 850s become duplicated shipments, which are their own chargeback.