Autonomous & infrastructure
Multi-partner format polymorphism
Twelve partners, one extraction prompt. Ticket arrival to finished release document in ninety seconds.
Read the caseAll case studies
Case study · Exception handling
Meridian Operations · autonomous pipeline exception pathException-path delivery orders: 80 sec from operator click to draft
The autonomous pipeline handles 92% of requests without human involvement. The other 8% are structurally ambiguous. The specialist gives the operator a structured interface to resolve that ambiguity — rather than starting from scratch on the source data.
0 sec
Operator click to draft delivery order
0%
Exception path coverage
0
Incorrectly issued delivery orders since launch
Built with
The problem
A 'Failed' status with a timestamp, and the operator restarts from the source PDF
Before the specialist, exceptions hit a queue with a status of 'Failed' and nothing else. No reasoning. No candidate options. No way to tell whether this was a data gap, an ambiguity, or a new pattern. The operator opened the source PDF, started over, and worked through the resolution from scratch — sometimes the next morning, depending on shift.
- No context on failureWhen the autonomous pipeline stalled, the operator got a ticket with status 'Failed' and a timestamp. No reasoning. No candidate options. The operator restarted from the source PDF.
- Access sprawlBefore gating, any operator with ticket access could attempt to issue a delivery order. Issuing one commits a real downstream action — it moves freight. Untrained operators occasionally issued duplicates or used wrong terminal codes.
- Shift carryoverExceptions that hit the queue after 4pm sat until the next trained operator came in. The exception queue was a reliable source of morning backlog. Resolution times were measured in shifts, not minutes.
- Audit gapsWhen a manually resolved exception turned out wrong, there was no record of which candidate the operator chose, why, or what input they typed into the form. Post-mortems relied on memory.
Exception queue · 8:02am
6 carried overMO-49118Failed16:42 prev. dayno reasoning recorded
MO-49120Failed17:05 prev. dayno candidates shown
MO-49126Failed19:18 prev. dayno reasoning recorded
MO-49131Failed21:44 prev. dayno candidates shown
MO-49138Failed02:13 overnightno reasoning recorded
MO-49142Failed06:58 todayno candidates shown
Operator opens source PDF · starts overResolution by shift
The pipeline
From inbox to verified record in one pass
Six steps from autonomous-pipeline failure to issued document. The specialist owns the structured interface; the operator owns the missing context.
01Trigger
Pipeline cannot auto-resolve
Autonomous pipeline hits an irreducible exception: no reference match, ambiguous address, missing terminal code, unknown carrier. Ticket lands with the failure code and the pipeline's reasoning.
Azure Functions
02Pick
Specialist gated by feature flag
Operator picks the Delivery Order Specialist. Gated by a feature flag — not granted by default, not derivable from other permissions. Trained operators only.
Claude
03Context
Failed run loaded in full
Specialist loads the failed pipeline run: failure code, reasoning, candidate options. Operator sees 'reference number didn't match — 3 candidates' rather than reconstructing from PDF.
04Resolve
Operator and specialist work the problem
Operator provides the missing context (paper bill, terminal call, prior experience). Specialist asks one clarifying question at a time and records the resolution reasoning.
Claude
05Generate
Draft from resolved inputs
Specialist regenerates the delivery order from the resolved inputs. Operator sees a draft — same fields, same layout, same downstream compatibility — before anything is issued.
06Issue
Operator confirms; reasoning logged to audit
Operator confirms the draft. Specialist issues the delivery order, updates the ticket, marks the exception resolved. Resolution reasoning stored in the audit log, not in memory.
Zoho Desk
Operator has the missing context
Eighty seconds from click to confirmed draft
Operator provides the missing context, specialist resolves the ambiguity in one exchange, and the draft is generated within 80 seconds. The ticket closes in the same shift.
Unknown carrier / new pattern
Resolve today's case; flag the pattern for engineering
Exception resolves for the immediate ticket, but the specialist flags the carrier for engineering follow-up. A pattern the autonomous pipeline has never templated will produce the same exception on the next ticket. Specialist handles today's case; engineering handles the pattern.
Validation review
Resolution confidence per exception type
The specialist handles five exception types. The 'unknown carrier' confidence is intentionally low — it reflects a known limitation about pattern-recurrence visibility, not a model failure.
Field-level confidence
Pass 2 — Claude self-review
HighMediumLow
No reference matchClosest candidates by edit distance
95%High
Ambiguous addressTwo candidates, distinct terminal codes
92%High
Missing terminal codeOperator-provided; format-validated
90%High
Conflicting codes across pagesOperator confirms which page governs
88%High
Unknown carrierNew shipping line, no template yet
59%Low
Routed to human review. Today's ticket resolves on operator-provided details, but the follow-up to template the carrier depends on engineering capacity. The specialist has no visibility into whether the pattern recurs before the template is added.
4 of 5 fields cleared the 0.85 threshold
model: Resolution confidenceThe stack
Boring tech, glued together well
Each vendor handles what it's best at. Aisyst owns the orchestration layer in between.
Claude
Loads context, drives conversational resolution, generates the draft
Zoho Desk
Ticket state, exception queue, audit annotation on resolved tickets
Azure Functions
Autonomous pipeline orchestrator and exception emitter
PostgreSQL
Pipeline run state, audit log with resolution reasoning
Drizzle ORM
Type-safe queries against the run-state and audit tables
Third-party logos are trademarks of their respective owners and appear here only to indicate integration.
Outcomes
What changed when context loaded with the ticket
Same-shift resolution rate climbed because the structured context load eliminated reconstruction time. The operator arrives at the decision, not at the source data.
0 sec
Operator click to draft delivery order
0%
Of inbound requests reach the specialist
0
Incorrectly issued delivery orders since launch
0%
Exceptions resolved within same shift
Watch the exception path rate
The autonomous pipeline should handle 92%+ of requests without intervention. If exceptions climb above 12%, the cause is almost always a new pattern in the data — a new carrier, a new terminal format, a new address structure. That pattern needs a pipeline patch, not more trained operators. Watch weekly; if it trends up, the pipeline's pattern library is behind the data.