Spare Hand Automation
Use cases

Workflows we've built, and what they changed.

Real builds, explained in plain English: the problem, what the automation does, where a person stays in control, and the result.

  • Receiving
  • Import documents
  • ERP integration
  • Working prototype

Import documents to inventory receiving, checked line by line

Built with a workflow engine and Claude, connected to Microsoft Dynamics 365 Business Central, and tested end to end on a clean shipment and a shipment with planted errors.

Watch the 2-minute demo Recorded in a test environment with sample data.

The problem

When an import container arrives, someone compares the commercial invoice, packing list and bill of lading against the purchase order, line by line, then keys the received quantities into the ERP. Short shipments, wrong items and documents that disagree are easy to miss until they turn into an AP dispute or an inventory error.

What we built

Drop the shipping documents in a folder. AI reads every line exactly as printed, plain rules compare the documents to each other and to the open purchase order, a person approves each line, and the approved quantities are staged on the PO, ready to post. Every decision is logged.

How it works

  1. 1

    AI reads the invoice, packing list and bill of lading and extracts every line exactly as printed. It never fills in or corrects a number.

  2. 2

    Rules match each line to the open PO and flag short shipments, items not ordered, documents that disagree, and weight totals that don't line up.

  3. 3

    The reviewer gets one form: approve or reject each line, with the reason beside every flag. Items not on the PO are never offered.

  4. 4

    Each line is re-checked in the ERP right before writing. Approved quantities are staged for receipt. A person clicks Post.

The import documents to receiving workflow in n8n: documents in, Claude extraction, PO matching, review form, Business Central write-back and audit log
The actual workflow (sandbox): documents in, AI extraction, matching, a person's review, then the write-back and audit log. Click to enlarge.
The result 3 of 3 shipment errors caught before anything was received

The test shipment carried three problems: 12 chairs short, an item that was never ordered, and an invoice billing 50 units where the packing list showed 48. All three were flagged on the review form. The reviewer rejected the disputed line, the extra item was never offered for approval, and only the approved quantities reached the ERP. When the ERP briefly refused two simultaneous updates, the workflow logged them as failed, wrote nothing, and the rerun went through clean.

Safeguards built in

  • • Line-by-line approval; unanswered lines expire
  • • When documents disagree, the lower quantity is recommended
  • • Never more than what's still open on the PO
  • • Every line re-checked in the ERP right before writing
  • • Refused writes are logged and flagged, never guessed
  • • Posting the receipt stays a manual click in the ERP

Built with

  • n8n workflow engine
  • Claude AI
  • Business Central
  • Google Drive
  • Google Sheets
  • Email approval forms
  • Inventory
  • Email requests
  • ERP integration
  • Working prototype

Email requests to inventory moves between locations

Reads move requests, checks stock at the exact location, confirms with the sender, and stages approved moves for a person to post. Unclear requests get a question, not a guess.

Built on Microsoft Dynamics 365 Business Central and tested end to end with five fictional requests.

Watch the 2-minute demo Recorded in a test environment with sample data.

The problem

Sales, warehouse and customer service teams email requests like “please move 20 units of item X from the main warehouse to the east location.” Someone reads each email, works out the item, quantity and locations, keys a reclassification into the ERP and replies. Requests get missed, misread (wrong item, wrong direction) or entered twice, and nobody can say later who asked for what.

What we built

One inbox for move requests, open to approved senders only. AI extracts each request; plain rules check it against Business Central; the sender gets a confirmation, a question, or the quantity actually available. A reviewer approves each line, approved moves are staged as journal lines for a person to post, and every request is logged.

How it works

  1. 1

    Watches a dedicated request inbox.

  2. 2

    Accepts requests from approved senders only. Anyone else gets a polite decline, and nothing is processed.

  3. 3

    AI reads the email and pulls out item, quantity, unit and from and to locations. It only extracts; plain rules decide.

  4. 4

    Checks every line against Business Central: the item exists, the description matches, both locations are allowed, and there is enough stock at the from-location, minus anything already waiting to be posted.

  5. 5

    Replies to the sender: a confirmation restating exactly what was understood, a clarifying question when something is unclear, or the quantity actually available when the request is too large.

  6. 6

    Sends clear lines to a reviewer, one Approve or Reject per line. No answer means no move.

  7. 7

    Re-checks stock right before writing, then stages each approved move as a journal line in Business Central.

  8. 8

    A person reviews and posts the journal. Nothing posts automatically.

  9. 9

    Logs every request and every line, including declined, unclear, over-stock and rejected ones.

The email requests to inventory moves workflow in n8n: Gmail trigger, sender check, Claude parsing, stock checks, reply, review, write-back and log
The actual workflow (sandbox): request email in, sender check, AI reading, stock checks, a reply, a person's review, then the write-back and log. Click to enlarge.
The result 5 of 5 fictional requests handled correctly

Stock changed only for the two approved lines, and only after a person posted them. The log holds 6 rows for 5 emails, each with its outcome and a plain-English note.

The test run

  • Clear request

    10 units: confirmed, approved, staged, and posted by a person.

  • Two-item request

    Both lines confirmed. The reviewer approved one and rejected the other; only the approved line was staged.

  • Unknown sender

    Politely declined. Nothing processed.

  • Vague request

    “Send some of the Athens items over to east.” The workflow asked instead of guessing, listing the missing quantity and source location and naming the two items that match “Athens.” Nothing moved.

  • Too-large request

    25 requested, 10 on hand: the sender was told only 10 are available. Nothing moved.

Safe by design: what testing caught

  • • Asks instead of guessing when a request is unclear.
  • • Blocks duplicate moves: a replayed approval was refused because the same move was already waiting in the journal.
  • • Checks stock at the exact location, not company-wide. In testing, a company-wide total would have shown hundreds of units that were not actually at the source location.
  • • A reply loop (the workflow answering its own confirmation email) was caught in testing and closed three ways before go-live.
  • • Stock is checked twice: when the request arrives and again right before anything is staged.

Built with

  • Business Central (sandbox, demo company)
  • n8n
  • Claude (Anthropic)
  • Gmail
  • Google Sheets
  • Inventory
  • ERP integration
  • Working prototype

Inventory cycle-count reconciliation with AI

Built with a workflow engine and Claude, connected to Microsoft Dynamics 365 Business Central, and tested end to end.

The problem

Physical counts were reconciled against the ERP by hand in a spreadsheet. Each variance had to be researched one by one, and adjustments were keyed back into the system manually. Slow, and easy to get wrong.

What we built

An automation that pulls on-hand quantities from the ERP and the physical counts from a shared sheet, has AI explain every variance, routes the important ones to a person for approval, and writes the approved adjustments back into the ERP, with a full audit trail.

How it works

  1. 1

    Counts are compared to the ERP, item by item. Missing or unknown items surface as exceptions.

  2. 2

    AI rates each variance, explains the likely cause from the record, and recommends an action.

  3. 3

    Flagged lines go to a reviewer as a simple approve/reject form, line by line. Unanswered lines are never approved.

  4. 4

    Approved adjustments are written into the ERP, pre-validated. A person clicks Post.

The cycle-count reconciliation workflow in n8n, grouped into trigger, data collection, AI analysis, human approval and ERP write-back
The actual workflow (sandbox), in five stages: trigger, data collection, AI analysis, human approval, write-back and audit log. Click to enlarge.
Close-up: Claude analyzes variances, flagged lines go to a review email, and only approved lines are posted to Business Central
Close-up: AI analyzes each variance, flagged lines go to a person by email, and only approved lines reach the ERP. Click to enlarge.
The result $74,050 false write-off caught before it reached the ERP

In the final test, six variance lines went through the workflow. Four were flagged for review: the reviewer approved two and rejected two, including a $74,050 “not counted” line that a fully automatic system would have written off. Exactly the approved adjustments were created in the ERP, matching the AI's analysis.

Safeguards built in

  • • Line-by-line approval, never all-or-nothing
  • • Unanswered items expire; they're never auto-approved
  • • Posting stays a manual click in the ERP
  • • Every line logged with its outcome

Built with

  • n8n workflow engine
  • Claude AI
  • Business Central
  • Google Sheets
  • Email approval forms
  • Accounts payable
  • Freight invoices
  • Working prototype

Freight invoice audit and recovery

Checks every carrier invoice against the contracted rates, flags overcharges and duplicate bills, and drafts the recovery claim for a person to send. Clean invoices go straight through.

Built with Google Drive, Gmail and Google Sheets and tested end to end with two fictional carrier invoices.

Watch the 2-minute demo Recorded in a test environment with sample data.

The problem

Carriers bill for every load. Each invoice should be checked against the contract: the base rate, the fuel surcharge, and every extra charge. Most teams don't have time, so overcharges get paid, duplicate invoices slip through, and the window to dispute them closes.

What we built

A workflow that reads each carrier invoice and checks it line by line against the rate card on file. AI only copies what is printed; plain rules flag the differences. A reviewer decides what to dispute, approved disputes become a draft claim email, and every invoice, line and decision is logged.

How it works

  1. 1

    Picks up new carrier invoices (PDF) from a Drive folder.

  2. 2

    AI reads each invoice and copies every line exactly as printed. It only extracts; it never judges a number.

  3. 3

    Looks up the contract for that carrier and lane: base rate, fuel surcharge cap, and the extra charges it allows.

  4. 4

    Plain rules check every line: base rate above the contract, fuel above the cap, extra charges the contract doesn't allow, and invoice numbers already paid.

  5. 5

    Clean invoices go straight through. No email, no review.

  6. 6

    Flagged invoices go to a reviewer in one email: every flag, the amount at stake, and the claim deadline. They choose Dispute or Don't dispute for each line. No answer means nothing is disputed.

  7. 7

    Approved disputes become a draft claim email in the reviewer's own Gmail, with no recipient. A person reads it and sends it.

  8. 8

    Logs every invoice, line and decision, including charges that were accepted.

The freight invoice audit workflow in n8n: invoice in, Claude reading, contract lookups, audit rules, review, draft claim and log
The actual workflow (sandbox): invoice in, AI reading, contract lookups, the audit rules, a person's review, then the draft claim and the log. Click to enlarge.
The result 4 of 4 planted billing errors caught

On one test invoice, a $1,347.00 duplicate was stopped before payment and $310.00 in overcharges was found, about 30% above the contract value. The clean invoice passed with zero flags and needed no review.

The test run

  • Clean invoice

    $1,037.00, matching the contract line for line. Zero flags, no email, no draft, logged as clean.

  • Base rate over the contract

    Linehaul billed at $950.00 against an $850.00 contract rate. Over by $100.00.

  • Fuel above the cap

    Billed at 26% ($247.00). The contract caps fuel at 22% of the contracted linehaul, $187.00. Over by $60.00.

  • Extra charge not in the contract

    A $150.00 redelivery fee. The contract allows only detention and liftgate.

  • Duplicate invoice

    The same invoice number was already processed on 2026-10-21. Paying it again would pay $1,347.00 twice.

  • The reviewer's choice holds

    In a second run the reviewer chose Don't dispute on the redelivery fee. The draft asked for a $160.00 credit and left the fee out.

Safe by design

  • • AI only reads and copies. Plain rules make every flag, and a person decides every dispute.
  • • No answer means nothing is disputed. Silence never approves a claim.
  • • The claim is a draft with no recipient. Nothing is sent to the carrier automatically.
  • • Every flag shows the claim deadline from the contract (180 days in the test), so disputes go out in time.
  • • Every invoice, line and decision is logged, including charges that were accepted.

Built with

  • n8n
  • Claude (Anthropic)
  • Google Drive
  • Gmail
  • Google Sheets

More use cases on the way

New builds are added here as client projects go live. Want yours to be next?

Tell us about one workflow

Prefer to talk? Choose a 20-minute call on the form and we'll send you times.