Telecom Supplier Invoice Reconciliation with LangSmith Fleets

Tue, 21 Jul 2026 23:05:00 +0530

Telecom Supplier Invoice Reconciliation with LangSmith Fleets

A telecom services business buys connectivity, voice, data, and network services from multiple carriers and resells those services to its own customers. Supplier invoice lines must be reconciled against customer billing data to find services where the business is still paying a carrier but may no longer be receiving corresponding customer revenue.

The existing workflow spans manual invoice downloads, SharePoint folders, provider-specific Logic Apps, Azure Function parsers, a reconciliation database, generated reports, spreadsheets, and manual investigation. Carrier formats vary across CSV, PDF, XLSX, ZIP, and voice-data packages. Large files, compressed packages, parser timeouts, inconsistent data, and legitimate one-to-many allocations make the process difficult to run consistently.

AAIC designed a LangSmith Fleet workflow that combines deterministic financial processing with focused agent reasoning. The result is a safer reconciliation model: code handles parsing and exact matching, while agents orchestrate runs and investigate exceptions using approved evidence.

Let’s discuss your finance operations automation use case

Talk to an Agentic AI Expert

Business Problem

The finance operations team needed a way to reconcile supplier invoices across multiple telecom providers without depending on inconsistent manual overlays.

The process had to identify unmatched supplier invoice lines, preserve evidence, and avoid risky automation choices. It needed to:

Accept one provider invoice package through manual SharePoint upload or an enabled provider API adapter.
Validate provider, source type, file structure, checksum, and run identity.
Safely unpack compressed files and reject unsafe or ambiguous packages.
Select the correct provider-specific deterministic parser.
Normalize provider invoice data into canonical supplier invoice lines.
Query only an approved read-only copy of customer billing data.
Match using service ID, provider, billing period or date evidence, invoice evidence, and account or customer references where available.
Route ambiguous, weak, multi-match, no-match, and parser-warning rows to human-reviewable exception investigation.
Preserve reports, manifests, parser warnings, query evidence, match summaries, and investigation evidence in a structured result package.

In-Scope Provider Families

The documented scope includes six provider families:

AAPT.
Telstra.
Optus.
Vocus.
Megaport.
Equinix.

Seven parser adapters are required because Optus has separate PDF and Excel/voice parsing paths.

Agentic Solution

The LangSmith Fleet design keeps sensitive financial extraction deterministic and uses agent reasoning where judgment adds value: orchestration, exception triage, and evidence-based investigation.

The workflow is:

Accept one provider invoice package through SharePoint upload or a provider API adapter.
Validate provider, source type, file structure, checksum, and run identity.
Safely unpack compressed files and reject unsafe or ambiguous packages.
Select the correct deterministic provider parser adapter.
Convert provider-specific data into canonical supplier invoice lines.
Query an approved read-only copy of customer billing data.
Apply deterministic matching using service, provider, billing period, invoice, account, and customer evidence.
Mark exact one-to-one matches as auto_matched.
Route ambiguous rows to an exception-investigator agent.
Produce raw and refined reconciliation reports with evidence for each decision.
Preserve source files, reports, manifests, parser warnings, query evidence, match summaries, and investigation evidence in a structured package.
Place result artifacts in SharePoint and send a text-only Outlook failure notification with artifact links or paths when a run fails.

The exception investigator can suggest a match using approved evidence. It cannot invent invoice rows, alter deterministic parser output, execute arbitrary SQL, or update the database.

Important Safeguards

The design deliberately keeps financial parsing and exact matching out of open-ended agent execution.

Provider invoice extraction remains deterministic.
Unsafe ZIP extraction fails closed.
Mixed Optus PDF and Excel/voice packages fail as ambiguous.
Billing lookup is read-only.
Exact candidate evidence is required for automatic matching.
One-to-many cases remain review items instead of being automatically allocated.
Human-verified matches require invoice-number evidence for the relevant billing period.
Run validation checks paths, run identity, row counts, report columns, no-write evidence, and secret markers.
Database update defaults to db_update_enabled=false and fails closed.
Provider and database secrets stay outside prompts, reports, and logs.

Solution Design

The maintained Fleet design contains:

A reconciliation supervisor.
An exception-investigator subagent.
A reconciliation skill for intake, orchestration, billing lookup, matching, reporting, validation, failure handling, and database-write guardrails.
A telco-parser skill with provider-specific deterministic adapters.
Stable script wrappers and shared deterministic implementation code.
Provider migration tracking and validation documentation.

This creates a practical boundary between software and agent reasoning. Deterministic code handles provider formats and exact matching. Agent reasoning is reserved for controlled exception investigation where the task is to explain and suggest based on evidence, not to mutate financial records.

Before and After Scorecard

The following scorecard defines the operating targets for the rollout and should be replaced with measured production values after the first deployment cycle.

Provider coverage

Before: Multiple carrier-specific flows existed across Logic Apps, Azure Functions, SharePoint folders, spreadsheets, and manual investigation.

After target: Six provider families and seven parser adapters operate behind common intake, parsing, reporting, and validation contracts.

Exception handling

Before: Unmatched rows required manual searches and inconsistent overlay practices.

After target: Exact matches are separated from ambiguous, weak, no-match, parser-warning, and one-to-many cases with evidence preserved for review.

Audit package completeness

Before: Evidence could be spread across source files, generated reports, spreadsheets, parser output, and manual notes.

After target: Every run preserves source, manifest, parser warnings, reports, match summaries, query evidence, and investigation notes in a structured result package.

What LangSmith Fleets Standardized

LangSmith Fleets provides the repeatable operating model for the reconciliation workflow:

Supervisor and exception-investigator responsibilities.
Reusable reconciliation and telco-parser skills.
Provider adapters behind common parsing contracts.
Runtime configuration, deterministic tools, and safety guardrails.
Consistent manifests, reports, evidence logs, and validation rules.
Human-review boundaries for ambiguous financial data.
A controlled migration structure for existing Logic Apps and Azure Functions.
A target integration pattern for SharePoint intake/results and Outlook failure notification.

Impact

The design helps the finance operations team move from a fragmented reconciliation process to a controlled, evidence-first operating model.

The value is not simply automating invoice parsing. The bigger gain is consistency across providers: one run package, one validation model, one matching contract, one exception pathway, and one auditable artifact structure.

That consistency makes it easier to identify revenue leakage candidates, prioritize human review, and migrate provider-specific automation without losing control of financial data.

Build a Similar Reconciliation Workflow

If your finance, operations, or managed services teams reconcile supplier invoices against billing data, AAIC can help design the right split between deterministic processing, agent investigation, human review, and measurable rollout controls.

Discuss your reconciliation automation workflow

Telecom | Applied AI Consulting

Telecom Supplier Invoice Reconciliation with LangSmith Fleets