Label and photograph
A unique barcode goes on the unit, then photos of the retail box and the product itself are captured on a phone. This is the only time a person handles the item.
AI operations case study · AI receiving automation
From blind receiving to 30-second AI identification.
How an overnight warehouse test became the largest operational cost savings in company history — conceived, tested and deployed in under two weeks.
Prior operating experience led by Built-N-AI founder Casey Jackson. Company identity is omitted.
Watch it first
DockSnap: AI receiving inside Dolibarr ERP.
Simplified rendering of DockSnap. This demo shows a single reception that did not need review.
Results at a glance
81%Labor cost reduction$23,500 → $4,500 per truckload
15×Labor savings multiplierPer-unit processing cost
<2 wksConcept to deploymentIdea → test → live production
90%+Accuracy on first launchDeveloper predicted 25%
3×Faster time to market3–4 weeks reduced to 1 week
<30sPer-item identificationPhoto → full product record
The challenge
Receiving was a costly black box.
For a company buying truckloads of mixed consumer electronics — mostly open-box and customer-return product — each truck arrived with thousands of units and no manifest worth trusting. Until the product was physically in hand, there was no reliable way to know exactly what had been purchased.
Identifying each item required multiple human touches: reading packaging, cross-referencing part numbers, researching manufacturer data and hand-keying information into the ERP. With 2,000 units on a truck, the process took 3–4 weeks from receipt to sale-ready state, tying up labor, slowing cash flow and creating downstream data-quality problems.
Incomplete or inaccurate product records also meant thinner marketplace listings, affecting impressions, conversion and sell-through.
“We had no idea what we were receiving until it was physically in our hands. The labor cost was enormous and the data quality downstream was hurting our sell-through.”
The insight
A late-night warehouse test.
Five electronics items were pulled from inventory. Using only a smartphone and an AI platform, photographs of the retail box and physical item were submitted directly to AI — a single multimodal step, one photo in, a full structured product record out.
The response came back in under 30 seconds — comprehensive enough to test whether AI could eliminate much of the identification and data-entry work.
Data returned in 15–30 seconds
- Full manufacturer part number
- Manufacturer name
- Full product description and specifications
- A listing title ready for online marketplaces
- UPC codes
- MSRP
- Used / resale market price
- A confidence rate on the identification itself
100% accuracy in the initial five-item test.
No barcode scanner. No prompt refinement. Photographs alone were sufficient for full product identification in the test.
From test to production
Under two weeks from idea to live workflow.
After the proof of concept, a full requirements document was written and handed to the company’s lead developer. His initial expectation was that image-based identification would top out around 25% accuracy — too low for production use.
In under one week, a new Advanced Intake module was built and integrated directly into the company’s Dolibarr ERP platform. On first launch, the system performed at more than 90% accuracy.
“Getting my lead developer excited and onboard was as important as the cost savings. Once he saw what this technology could deliver, the whole organization shifted.”
What the tool does
Photograph the product once. Nobody keys it in again.
The module was not an identification lookup bolted onto an existing process. It replaced the receiving process. A worker labels a unit, photographs it, and puts it down — the product record and the purchase order are built from that, and the review that used to mean re-handling the item now happens on a screen.
AI identifies it
In 15–30 seconds: manufacturer part number, manufacturer, description and specifications, a marketplace-ready listing title, UPC, MSRP, used market price, and a confidence rate.
This isn’t the model guessing — it’s retrieval-augmented generation (RAG): the AI looks up the real manufacturer data, UPC and current marketplace pricing before it ever writes the listing.
Needs review
Entries land in a review queue. A person inspects each one and either approves it or overrides the data on the fly — a judgement call on a screen, not another trip to the pallet. The confidence rate tells them where to look first.
This is human-in-the-loop by design — the AI flags what needs a second look, a person makes the final call before anything hits the ERP.
One button, whole truckload
Once the load is imaged and approved, a single click checks every part number against the database. Anything new is auto-created as a fully loaded product record, attributes and specifications already populated, and the purchase order is generated.
PO created automatically
The purchase order is generated automatically, with every unit on the load as a line item — including the part numbers that were just created. No one builds the PO by hand or keys a single part number. It all happens on its own.
Received into stock
The PO is received against the barcode applied back at capture. The unit lands in stock priced and listable, with no re-handling and nothing keyed by hand.
One physical touch, start to stock.
The old process read packaging, cross-referenced part numbers, researched manufacturer data and hand-keyed the record — several separate touches per unit, each one a chance to introduce an error. Now the product is handled once. Everything after that, review included, happens on a screen against a record that is already populated.
What that changed
Four compounding gains, not one.
- Single touch. Product is handled once instead of at every step of identification, data entry and receiving. Review moved from the pallet to the screen.
- Speed. Identification dropped from minutes of research per unit to well under a minute, and a truckload from 3–4 weeks to about one.
- Data accuracy. Attributes come from the AI response straight into the record, so transcription errors and half-filled fields largely disappear.
- Throughput. The same crew clears far more units per shift, because the bottleneck moved from human research to image capture.
Why it compounds
Better records sell faster.
Richer product data is not a back-office nicety in this business. A complete record — correct part number, real specifications, a proper listing title, a defensible price — produces a stronger marketplace listing, and stronger listings get more impressions and convert better.
So the labor savings and the revenue side pull in the same direction: the faster and more accurately product is identified, the sooner it is listed and the better it performs once it is.
The confidence rate is what makes the review queue efficient rather than ceremonial. It tells a reviewer which of the thousands of units on a truck deserve a hard look, instead of spot-checking blind — and a low-confidence entry can be corrected in place before anything is created downstream.
Before & after
The operating model changed.
Before AI
3–4 weeks to process one truckload
After AI
1 week to process one truckload
Before AI
Manual identification with multiple human touches
After AI
Photo capture → AI identification in under 30 seconds
Before AI
New part numbers built by hand, one field at a time
After AI
Missing part numbers auto-created, attributes and specs already populated
Before AI
PO creation and receiving keyed by hand after identification
After AI
One click creates the PO; it is received against the barcode applied at capture
Before AI
~$23,500 labor cost per 2,000-unit truckload
After AI
~$4,500 labor cost — an 81% reduction
Before AI
Inconsistent product data and thin listings
After AI
Richer records: MFR, P/N, UPC, retail price and description
Before AI
Slower sell-through from incomplete product details
After AI
Faster path to complete marketplace listings and sale-ready inventory
Before AI
Developer predicted ~25% accuracy at best
After AI
90%+ accuracy on first production launch
Business impact
More than labor savings.
Lower direct labor cost
Labor cost fell from about $23,500 to $4,500 per 2,000-unit truckload.
Faster processing
Truckload processing fell from 3–4 weeks to about one week, accelerating the path to market.
Richer product records
AI returned manufacturer, part number, UPC, pricing and descriptive information needed for stronger product records.
Intake pipeline
The workflow connected image capture, product identification and ERP intake with minimal human intervention.
AI adoption shifted
Seeing a skeptical developer become an advocate helped open the door to broader organizational AI adoption.
Founding Client Program — 3 spots
Prove it on your data before you commit.
This workflow was built and proven inside real operations, not sold before it was tested. The first three external clients get founding-client pricing in exchange for a case study and testimonial.
Step 1 — Pilot
$1,200
A 2-week pilot on one truckload or product batch from your own inventory, proving AI identification accuracy on your actual parts before you commit to full integration. Credited toward the full build if you move forward.
Step 2 — Full build (founding-client rate)
$4,500
The full Advanced Intake module integrated into your ERP, scoped to your receiving process. Founding-client pricing for the first 3 external clients only.
Key takeaway
Deep operational expertise + practical AI.
This is what becomes possible when AI is treated as a solution to a real operating problem rather than as a technology project. The entire process — late-night test to live production — took less than two weeks, without an outside AI consulting team or a months-long implementation.
The ingredients were operational experience, a clear problem and the willingness to test an idea.