Freight Document Processing AI Agent

What this agent does

Reads the document, writes only the record

Rate confirmations, bills of lading and invoices, ingested from supported TMS, document, email or EDI sources.

Field labels and formats normalised in the record only, with the page and the document each value was read from.

Reason

A rate confirmation as a contract — terms extracted and compared, never amended.

A bill of lading as a receipt, a contract of carriage and sometimes a document of title.

The straight-versus-order distinction surfaced, because endorsing the wrong one moves title.

Decide

Declared-value and limitation-of-liability terms flagged for a person, because they set what is recoverable after a loss.

An invoice as an assertion — tested against the contract, with any absence reported.

Out

Electronic bills kept as documents of title: derived copies marked non-authoritative.

Write actions executed only inside the approval boundaries agreed during implementation.

→ Product statement

The agent reads and structures; a named person accepts each field, and the original document stays where it is, unchanged.

Example workflow

One document, arrival to acceptance

AgentHuman

1Document receivedRate confirmation, bill of lading, commercial invoice or supporting paperwork

2Document type establishedWhat the document legally is, what may lawfully be done to it, and the pages each field is read from

3Record preparedFields, page references and confidence

4Controls appliedDocument-type rules, required-element checks, authoritative-copy marking and the confidence threshold

No human action required

Stages 1 to 4 run unaided and alter no document — the record is prepared, and the person's lane opens at the confidence gate.

5DecisionBranches at the confidence threshold

High confidence

Goes to the named person to accept.

Low confidence

Adds a document-desk read first.

A person accepts

The record is held with its page references, its flags and the confidence.

Accept · Correct · Send to the document desk

Accepted — record released▼

6Downstream systems updatedOnly where write access and approval policy allow it

7Outcome evaluatedField accuracy, corrections on acceptance, absent-element reports and downstream rework

Corrections

Every correction made on acceptance is counted in the evaluation.

What should not run autonomously

Human approval stays in control

Outside the boundary — human approval required8 items

Generating, altering or re-issuing a bill of lading.

Amending, signing or accepting a rate confirmation.

Generating an invoice or filling a field left blank.

Deciding which values populate a customs entry.

Automation boundaryAgent acts unaided

✓Read the document and write the structured record from it.

✓Carry each field forward with the page it was read from.

✓Compare extracted terms against the tender and the agreement.

✓Report a required element as absent, and hold it for a person.

Any write happens inside the boundaries agreed at implementation, and never to the document itself.

Setting a declared value or a limitation of liability.

Deciding the country of origin of imported goods.

Holding or emitting an authoritative electronic copy.

Changes to extraction rules, thresholds or approvals.

Example output

One extracted field, annotated

Everything the agent extracts is attached to the page it was read from.

Extraction output · single documentIllustrative example

Document

Extracted field

Value read

Page reference

Confidence

Status

Order bill of lading

Declared value stated on the face of the bill

$50,000.00

Page 1, value box

88%

Flagged, original untouched

As receivedTaken from the page as it was scanned — nothing on this side is inferred, and the document itself is not changed.

What the page shows The value box as printed The negotiability marking The named consignee

Why it stops hereA declared value sets what is recoverable after a loss and a person still decides.

ActionAcceptCorrectSend to the document desk

What the score decidesBelow the configured threshold the field picks up a document-desk read before it.

Value

Where AI adds value

The same four claims, placed at the point in the workflow where each one applies.

Where the value landsValue 01 – 04

Every documentFrom the intake channel or the TMS

03Reading

Read to the rules for the type

Draw on what the document legally is, the pages in front of the agent and your field definitions.

01Approved path

Read it, keep the original

Routine documents come back as records, with page references attached.

02Human review

Send the rest to a person

Absent elements, ambiguous terms and low-confidence fields are marked, so the operator's read starts where risk concentrates.

04Build an evidence trail

The extracted field, the page it came from and the person who accepted it stay with the document.

Integrations

Typical integrations

Five system groups connect to the same agent. Which of them are in scope is decided in discovery.

Document intakeEmail · scanned mail
SharePoint · Box · Dropbox

TMS and shipment recordsMcLeod · MercuryGate
Trimble · Revenova

eBL and trade platformsBolero · essDOCS · WaveBL
Carrier eBL portals

Agent

Freight document processing

Reads the document
Writes the record
Holds for a person

Payables and EDIEDI 210 · 214 · 990
ERP and payables

Observability & evaluationOpenTelemetry · Langfuse
Supported monitoring/evaluation sources

Integration availability depends on the client's existing systems and API access.

Agent controls

Six layers between the model and the record

Controls stack inward. What the stack does not catch is set out in the map below.

L6 · Outermost — last line of defenceInward → L1 · closest to the model

L6Rollback / safe modeNarrow the agent to filing and indexing when evaluation or production signals degrade.Roll back

L5Version monitoringTrack model, prompt, extraction-rule and document-type configuration changes.Track

L4TraceabilityRecord the document, the page, the extracted field, the flags and who accepted it.Record

L3Person acceptsHold the record for a named person; it governs what is accepted, not whether the document is genuine.Gate

L2Policy guardrailsTest records against the configured document-type and required-element rules; a failure returns the record.Restrict

L1Confidence thresholdsRoute low-confidence fields to a document-desk read before the person sees them.Require review

Model coreRecord produced — fields, page references, flags and confidence

L1 – L2Test whether a record may stand

L3Puts acceptance in a person's hands

L4 – L5Keep the field and the page it came from

L6Narrows to filing and indexing when signals degrade

How Nestack evaluates it

Evaluate the reading workflow — not only the finished record.

Coverage runs the whole depth of the workflow, and every layer is cut by slice.

Surface — the record the business acts on

Depth of coverage ▼

E1Final-output evaluationDid each extracted field match the page it was read from?

E2Step-level evaluationDid the agent apply the rules for the document type in front of it?

E3Tool evaluationDid it read and write the correct document and the correct field?

E4Confidence calibrationDo low-confidence fields actually attract more corrections on acceptance?

E5Slice evaluationHow does performance change across specific document types?

E6Business outcomeHow many fields were corrected, or reworked after acceptance?

Floor — the outcome the operations team answers for

Failure modes

Where each failure originates in the agent

Seven ways a read goes wrong, placed at its own stage.

Agent lifecycleDirection of processing →

01 · Retrieval1 mode

BD-03

Wrong copy read

A field is taken from a superseded or duplicate copy.

Stage gathersRate confirmations, bills of lading

02 · Reasoning2 modes

BD-04

Document type misread

A contract is handled under the rules for an assertion.

BD-06

Ambiguity settled silently

An unclear term is recorded as though it were agreed.

Stage proposesFields, page references and confidence

03 · Tool / write2 modes

BD-02

Absence supplied

A blank required element is filled by inference.

BD-05

Derived copy unmarked

A derived artefact circulates without its marking.

Stage writesOnly where write access and approval policy allow it

04 · Output1 mode

BD-01

Page reference lost

A field arrives without the page it was read from.

Stage returnsThe record a person accepts and the business acts on

05 · Change / Version1 mode

BD-07

Silent extraction regression

A model or rule change widens what the agent asserts from a page.

Stage tracksModel, prompt, extraction rules and type config

Sev-1 · a document is written or altered Sev-2 · a wrong field reaches a downstream system Sev-3 · scan degrades, field routes to review

See the complete Transportation failure-mode catalogue →

Affected slices

Overall accuracy can hide a document type

Handwritten bills of lading are the cohort a total under-counts: a small share of the volume, a large share of the corrections. Nestack reports the correction rate by document type, not only in aggregate.

Slice performance — reported separately, not only in aggregateIllustrative example

Slice	Failure rate	Lift	Status
Handwritten bills of lading	5.1%	3.2×	Review
Scanned and faxed copies	4.3%	2.7×	Review
Non-English invoices	3.0%	1.9×	Watch
Digital rate confirmations	1.0%	0.6×	Normal

Bar: correction-rate lift vs. digital rate-confirmation baseline · scale 0–4.0× · tick marks the 2.0× review threshold 2 of 4 slices over threshold

Evidence-linked improvement

A cycle closes on a case, not a correction

An apology does not close a cycle; a case the next release must pass does. That suite is what the next document read into the system is measured against.

Improvement cycle · five stagesSwitchback — the path turns at Improve and returns at Learn

01Detect

Corrections rise in one document type.

02Diagnose

No cause is accepted until the document, the page and the extracted field have been read back against each other.

03Improve

The correction is versioned, with the documents that motivated it attached.

04Verify

A failing case holds the release until it passes.

05Learn

The case is added permanently, and the extraction rules move with it.

Learn → DetectThe return edge. The next document read is measured against a suite one case longer.

Typical build scope

Twelve workstreams across six weeks

The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, sources, reading workflow, evaluation, integration, then production validation and handover.

Workstream Week 1Week 2Week 3Week 4Week 5Week 6

01Document workflow discovery and boundary definition.

02Document, TMS and EDI source assessment.

03Document-type rules and required-element field mapping.

04Document ingestion and field normalisation.

05Extraction logic and page binding.

06Confidence scoring and field routing.

07Acceptance workflow.

08Document-system integration.

09Field-extraction regression cases.

10Guardrails and acceptance controls.

11Document-trail instrumentation.

12Deployment, documentation and Agent Care handover.

12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallel Final scope and sequence confirmed in discovery

Engagement tiers

What each tier includes

Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.

Capability✓ in scope · — not at this tier PilotOne document type, one lane ProductionProduction document systems AdvancedMultiple formats / entities

Introduced at Pilot

Reading to your document rules✓✓✓

A person accepts each field✓✓✓

Extraction-accuracy baseline✓✓✓

Introduced at Production

Reporting by document type—✓✓

Acceptance workflow in your systems—✓✓

Approved record write-back—✓✓

Document-system integration—✓✓

Introduced at Advanced

Multi-format and multi-language rules——✓

Multi-stage acceptance approvals——✓

High document volume——✓

Multi-format document controls——✓

Build price From $5,000 From $8,000 Custom quote

Final build priceConfirmed after discovery based on integrations, workflow complexity, transaction volume, approval controls and deployment requirements.

Separate from buildBuild pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.

What we need from you

What you bring, and what we build with it

Each input maps to a piece of build scope and a week in the delivery timeline.

You bringWe build with it

01Your document types and where each one arrives → Document-type rules and required-element mappingWeek 1

02Representative documents, including the bad scans → Extraction baseline, field normalisation and page bindingWeek 2

03Your field definitions and the record they feed → Extraction-rule mapping and automation-boundary definitionWeek 1

04Access to relevant APIs, feeds or exports → Document, TMS and EDI assessment, then integration setupWeek 2

05Fields you would not want posted → Extraction cases and the evaluation suiteWeek 4

06What an extracted field may never replace → Confidence scoring, field routing, guardrails and acceptance controlsWeek 3

07Named people to accept extracted fields → Acceptance workflow, then pilot and production validationWeeks 5–6

Nothing else is required Deployment, documentation and Agent Care handover are ours.

Delivery timeline

Four phases across six weeks

Each band covers the weeks the work really takes, so week 5 carries evaluation and launch together.

Phase W1W2W3W4W5W6

Discovery W1

Build W2 – W3

Evaluate W4 – W5

Pilot & Launch W5 – W6

Week focus W1Document workflow discovery, type rules and the automation boundary W2Source integration and the extraction baseline W3Extraction workflow, confidence logic and acceptance controls W4Evaluation suite, required-element checks and failure-mode testing W5Document-system integration, pilot documents and targeted corrections W6One document cycle read under the operations team, then handover

Reading the bandA bar sits on the weeks its work is named in and no others. The week 5 overlap is real work, not padding.

At the end of W6Validation finishes on live documents and Agent Care assumes monitoring.

DurationSix-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.

Next step · Transportation AI agent

Build a document agent around what may lawfully be done to each document.

Which documents may the agent only read, which may it compare, and who accepts a field before it moves? Show us where they arrive, and we'll map the workflow and set the boundary.

Discuss this agent →

Nestack Agents · Freight document processingAGT-TR-15 · Agent Care available after launch