Warehouse AI: what it does today and how a smaller distributor starts

Warehouse AI is software that reads a warehouse’s own records to suggest where each item should be stored, the order in which to pick it and when to refill a pick location, and that reads counts, damage and shipping papers from cameras and scans. Most of it runs on data already held in a warehouse management system (WMS), the software that tracks every location and movement in the building, or in an ERP, the system that runs orders, purchasing and accounting. A person approves what it proposes. This guide covers what AI does in a warehouse today and how a smaller distributor in Ontario or Quebec can start with one job.

Derik Lawlis Founder, ThriveAI

What warehouse AI means

The name covers three kinds of software, each with different data needs:

• Rules and optimization. A slotting template or a route calculation follows logic someone set up, and needs complete item and location records.
• Machine learning. A model finds patterns in history, such as the weeks when an item sells faster, and forecasts from them. It needs a long sales history.
• Language and vision models. These read documents and images, such as a packing slip or a photo of a damaged pallet. They need samples of your own documents to be tested on.

Few wholesalers use AI yet. In Statistics Canada’s survey for the second quarter of 2026, 19.2% of all businesses had used AI to produce goods or deliver services in the previous 12 months, and wholesale trade was among the three lowest industries at 7.9%. The other two were agriculture, forestry, fishing and hunting, and construction, which AI in agriculture and AI in construction cover. The wider planning picture is in AI for supply chain.

What AI does in a warehouse today

Each task below is offered today in warehouse or ERP software.

Slotting: where each item lives

Slotting decides which location, or slot, holds each item, usually to shorten trips for the items picked most often. Microsoft’s warehouse slotting in Dynamics 365 totals the demand on open orders by item and unit of measure, assigns it to pick locations by quantity, unit, physical dimensions and fixed locations, and creates the work to move stock into place. The demand can be imported from another system, and Microsoft notes that “Whatever is in the slotting demand is used in the next step, regardless of where it came from.” That is where a forecast of next month’s picks fits. Every re-slot moves stock, so a supervisor approves the plan first.

Pick paths: the order of the stops

A pick path is the order in which a picker visits locations, a version of the travelling salesman problem of finding the shortest route through a set of stops. Microsoft is previewing a spatial location feature in Dynamics 365 that stores X, Y and Z coordinates for each location and reorders pick lines so that “workers follow an efficient route through the warehouse.” For aisle layouts, Microsoft says its fast option “typically produces routes that are close to optimal,” and it adds that preview features “aren’t meant for production use.” The gain depends on the lines per trip. A picker who fetches one order at a time has no route to improve, so batching orders comes first.

Replenishment: refilling pick locations from demand

Replenishment moves stock from reserve storage down to the pick face, the location pickers take from. Among the replenishment strategies in Dynamics 365, min/max keeps each pick location between limits you set and typically runs once a day, and wave demand creates refill work when a wave, a batch of orders released together, needs stock the pick location lacks. Machine learning can supply the forecast behind those limits. Microsoft’s Demand planning includes auto-ARIMA, ETS, Prophet and XGBoost, plus a best fit option that “automatically selects the best of the available algorithms for each product and dimension combination.” AI for inventory management covers forecasts and reorder points.

Labour planning: people per shift

Labour planning matches the work coming in with the people on shift. Blue Yonder, a supply chain software vendor, describes engineered labour standards as “a physics-based model that calculates exactly how long a task should take based on distance, weight, and process steps,” and says its AI-driven forecasting can predict “Workload Surges” weeks in advance. A smaller warehouse can start from its own scan history, since the time between scans estimates how long each task takes.

That use sits close to an example in Ontario’s guide to electronic monitoring: “an electronic sensor to track how quickly employees scan items at a grocery store check-out.” An Ontario employer with 25 or more employees on January 1 must have a written electronic monitoring policy in place before March 1 of that year.

Vision for counts and damage

Computer vision is software that reads camera images. Amazon describes Project P.I., which scans products in an imaging tunnel before they ship and uses “a combination of generative AI and computer vision technologies” to catch damage or the wrong colour or size, with associates reviewing what it flags. Gather AI sells drones that fly a warehouse on their own and capture “3D case counts, damage detection, and LPN-to-location verification,” where an LPN, or licence plate number, is the barcode that identifies a pallet or case. Both descriptions come from the vendors. A smaller warehouse can start with a phone photo of each inbound pallet, which gives every damage claim a dated picture. Computer vision in manufacturing covers camera setup and where it fails.

Reading packing slips and bills of lading

A packing slip lists what a supplier shipped. A bill of lading is the contract of carriage, and Ontario’s Carriage of Goods regulation says it must name the consignor and the consignee, give the particulars of the goods, and record whether the carrier received them “in apparent good order and condition.” Language and vision models read both from a scan or a phone photo and fill in a draft receipt for a receiver to confirm. Microsoft’s Azure Document Intelligence has prebuilt models for invoices and receipts and none for these documents, so reading them there means a custom model trained on labelled samples. Microsoft says “You need only five examples of the same form or document type to get started.” Text extraction from images covers testing it on your own files.

The packing slip is also the middle document in a three-way match, where the purchase order, the receipt and the supplier’s invoice must agree before the invoice is paid. AI invoice processing covers that match, AI in procurement covers drafting purchase orders, and a single layout such as the purchase order template keeps your own orders consistent. Outbound freight and customs paperwork are in AI for logistics.

Autonomous mobile robots

An autonomous mobile robot (AMR) moves carts, totes or pallets through a building on its own, steering around people and obstacles with its sensors. OTTO Motors was the industrial division of Clearpath Robotics, an Ontario company that Rockwell Automation acquired in October 2023. Rockwell describes OTTO as “autonomous technology for material handling inside manufacturing and warehouse facilities,” and OTTO lists models that move loads from 150 kg up to 1,900 kg. Slotting and replenishment decisions still come from your WMS or ERP, and a fleet needs a map of the building and a safety review before it works beside people.

Warehouse AI at a glance

What the WMS or ERP must provide

Each task depends on records a warehouse already keeps, and on whether they are complete:

• Item records. Unit of measure, case and pallet quantities, dimensions and weight. Slotting cannot fit an item without its size.
• Location records. An ID, a size and either coordinates or a walking sequence for each location. Microsoft’s route sorting skips any job whose put location lacks valid coordinates.
• Order history. Shipped lines by item and date, covering at least a full year so that seasonal items show their season.
• Stock and receipts. On-hand quantities by location, and receipts against purchase orders with dates, which also give real supplier lead times.
• Scan history. Who scanned what, where and when.

If bin locations live in a free-text ERP field or in one person’s memory, the location data is the first project. AI for ERP covers reading the ERP you run, and legacy ERP automation covers older systems on your own server. A plant can pair this with production scheduling software with AI.

How a smaller distributor starts

1. Pick one job. Choose a count you already track, such as mis-picks per week, receipts keyed per day, or hours spent on cycle counts, where a few locations are counted each day in place of one full count a year.
2. Pull and fix the data. Export the item, location and order history for that job and fill the gaps it shows.
3. Run it beside the current process. For a few weeks, compare the software’s suggestions with what your team did. AI evals explains how to test on past work.
4. Keep a person approving until the numbers hold.

Document reading and replenishment need no new equipment, so they are cheaper first tests than robots or drones. Cost of AI breaks down the spending, and AI strategy and AI implementation cover choosing the first project and moving it into daily use.

Approvals and an audit trail

An audit trail records what the system proposed, which data it used, who approved it and when. It answers later questions, such as why a receipt shows 48 units where the packing slip said 50.

• Approval first. A named person approves new slots, new stock limits and each receipt before it posts. Human in the loop covers designing an approval step people actually read.
• The source stays attached. Each value links to the document it came from. The Canada Revenue Agency says records and supporting documents must generally be kept “for a period of six years from the end of the last tax year they relate to.”
• Changes are tested first. A new rule or model runs on past weeks before it runs on today’s orders.

An agent that acts on its own, such as one that releases replenishment work, needs these controls most. AI agents for business covers the guardrails, and private AI for business covers where your data sits while a model reads it. The guides cover other sectors too, including AI in mining.

How ThriveAI helps

ThriveAI is an AI engineering company in Ottawa that builds private AI systems for manufacturers and distributors in Ontario and Quebec, on their own data. It brings what your ERP, email, drawings and spreadsheets hold into one database that belongs to your company. Every answer shows where it came from, and it drafts routine work, such as order entries and purchase orders, for someone on your team to approve.

ThriveAI’s systems read your ERP as it is, including older versions installed on your own server, and every connection only reads data. The platform is designed to keep each client’s data on its own server in Canada. You choose a model on that server or a hosted model under a written zero data retention agreement, and a hosted model may process requests outside Canada. Derik Lawlis, the founder, leads every project and stays close to the build. About ThriveAI covers the company.