Stable
ServicesAI & Automation for logistics and trade compliance

Your operations,
running themselves.

Stable builds agentic workflows, document AI, and intelligent process automation for trade and logistics teams, so your people focus on decisions, not data entry.

See what we build
60%+
Reduction in manual entry on AI-processed workflows
<2s
Document extraction turnaround, median, in production
4–6 wks
From kickoff to production for a focused extraction workflow
0
Black-box models. Every decision is logged with its reasoning.
Stable Agent Pipeline · Commercial Invoice Processing
Live · 284 docs processed today

Document queue

PDF
INV-2026-04-081.pdf
Processed
XLS
packing-list-HKG.xlsx
Processed
PDF
INV-2026-04-082.pdf
Extracting…
EDI
856-MAEU-20260422.edi
Queued
PDF
COO-CN-20260422.pdf
Queued

Agent stages

Layout detection
Field extraction
Confidence scoring
HTS classification
Sanctions screening
Routing decision
INV-2026-04-082.pdf
94.2% confidence
Supplier
Shenzhen Brightway Elec. Co. Ltd.
99.1%
Invoice date
2026-04-18
98.7%
Invoice total
USD 47,320.00
99.4%
Incoterms
FOB Shenzhen
81.3% — review
Line item 1 — description
PCB assembly, wifi module, 802.11ax, 2.4/5GHz, 500 units
97.6%
Proposed HTS classification
8517.62.0090
Machines for the reception, conversion and transmission of voice, images or other data: Other apparatus for transmission or reception of voice, images or other data
8517.62.0090
88%
8525.60.1010
7%
8542.31.0001
5%

Routing decision

Auto-file
Cleared for ACE entry
Confidence above threshold · No holds
Review queue
Incoterms — verify
FOB vs. CIF affects dutiable value

Agent log

10:14:02 · layout detected · commercial-invoice-v210:14:02 · extracting 14 fields10:14:03 · 13/14 high confidence10:14:03 · incoterms 81.3% → review queue10:14:03 · classifying line item 110:14:04 · 8517.62.0090 · 88%10:14:04 · screening supplier…10:14:04 · no sanctions match10:14:04 · routing → auto-file + 1 review
Document AIAgentic WorkflowsIntelligent ClassificationException RoutingZero Manual EntryRuns in WeeksDocument AIAgentic WorkflowsIntelligent ClassificationException RoutingZero Manual EntryRuns in Weeks
What we build

Six AI capabilities,
deployed in your stack.

Each capability ships as a production system. Not a prototype, not a pilot. Trained on your data, integrated with your existing tools, and monitored in production.

01

Document AI & Extraction

Invoices, bills of lading, 7501s, packing lists, certificates of origin: we train extraction models on your actual document types. Fields come out structured, typed, and confidence-scored. Human review only when the model flags uncertainty.

02

Agentic Workflows

Multi-step agents that handle intake, routing, and follow-up autonomously. An agent can read an EDI 856, look up the HTS code, check the denied party list, calculate duties, and queue a filing, all without a human in the loop.

03

Intelligent Classification

AI-assisted HTS classification trained on your commodity history. Ranked alternatives with confidence scores, auditable reasoning, and a review queue for edge cases. Your classifiers stay in the loop; the AI handles the obvious ones.

04

Exception Routing & Alerts

When something's wrong (a value discrepancy, a missing document, a hold flag), the system routes it to the right person immediately, with context. No more end-of-day email chains. No more exceptions buried in a queue.

05

Data Reconciliation

Shipment records, invoices, and customs data rarely agree on first pass. We build reconciliation agents that catch mismatches before they cascade, across your TMS, ERP, broker portal, and filing system.

06

Operations Intelligence

Aggregate data across shipments, brokers, and carriers to surface trends your team would otherwise miss: duty exposure by supplier, processing time by document type, exception rates by lane. Built into the workflow, not a separate BI tool.

How we work

From messy process
to running system.

Every automation starts with understanding the real workflow. Not the documented one. We map, train, deploy with guardrails, and hand you a system that improves over time.

01: AUDIT

Map the process we're automating.

We spend two weeks with your ops team tracing the actual workflow. Not the documented one. Where do humans touch data? Where do things break? What's the exception rate? We scope automation around the real process, not an idealized version.

02: TRAIN

Train on your documents and data.

Your data is the training set. We extract, label, and fine-tune models on your actual invoices, customs entries, and internal records. Not generic corpora. Models that understand your commodity mix, your carriers, your customers.

03: DEPLOY

Deploy with guardrails and review queues.

No big-bang cutover. Automation starts in shadow mode, processing alongside humans so you can see what it would have done. We tune confidence thresholds until it's better than your current error rate, then cut it over.

04: IMPROVE

Monitor, measure, and improve.

Every decision the system makes is logged with its reasoning. Exception rates, confidence distributions, and processing times feed back into model updates. We hand over a system that gets better with every shipment.

Recent work

AI in production,
with numbers to show for it.

Metrics are measured in the first quarter after deployment. Not projections. Anonymized where requested.

Trade Compliance · IOROngoing · '25–'26
Document AI

Cut manual document entry by 60% for a global IOR.

A Fortune 500 importer processed hundreds of customs entries daily across dozens of commodity types. We built AI document processing that extracts shipment data, validates against ACE records, and queues exceptions for human review, reducing manual entry processing time by over 60%.

60%
Reduction in manual entry processing time
<2s
Document extraction turnaround, median
Customs Brokerage · DrawbackLaunched Q1 '26
Agentic workflows

Automated duty drawback identification and filing prep.

DrawbackAI analyzes import and export records to surface eligible drawback claims, calculates recovery amounts, and assembles the filing package, work that previously took a specialist several hours per claim. Importers now file claims they would have written off.

~97%
Of eligible claims identified vs. manual review
More claims processed per specialist per day
Financial Services · InsuranceLaunched Q4 '25
Back-office automation

AI-powered back-office for a regulated insurance carrier.

Vigil needed SOC 2-certified automation for annuity and life insurance servicing. We built AI validation workflows, automated compliance checks, and intelligent document handling, all with the audit trail a regulated carrier requires.

SOC 2
Type II certified from day one
70%
Reduction in manual back-office processing
Common questions

What teams ask us
before they start.

No. Messy data is the default state of every operations team we've worked with. We start with a data audit: understanding what you have, what format it's in, and where the quality breaks down. We build the cleaning and normalization into the pipeline, not as a prerequisite you have to complete first.
We set a baseline from your current process first. If your team currently mis-classifies 4% of entries, the bar for the AI isn't 100%. It's better than 4% on the cases it handles, with a reliable way to flag the ones it's uncertain about. Shadow mode lets you measure accuracy against your real workload before any cutover.
Every AI decision includes a confidence score and reasoning trace. Below a configurable threshold, records route to a human review queue automatically. Wrong decisions are logged and feed back into model updates. The system is designed to fail gracefully. Not silently.
That's the common case. Commercial invoices from 30 different suppliers all look different. Bills of lading from different carriers use different field names. We train layout-aware models on your actual document corpus, the more examples you have, the better. Starting with as few as 200–300 labeled examples is enough to build a useful baseline.
A focused document extraction workflow (one document type, one extraction target) can be in production in 4 to 6 weeks. A full agentic workflow that spans intake, classification, validation, and routing typically runs 10 to 14 weeks. We scope it with you before we start so there are no surprises.
Yes. The models, the training data, the pipeline code, and the configuration all sit in your infrastructure from day one. We hand over architecture documentation and a 30-day support window through your first month in production.

Automate the work
holding you back.

Tell us the process in a paragraph. We'll come back in a week with a written scope, a deployment timeline, and a fixed fee. No pilot programs, no proof-of-concept theater.