A small warehouse or 3PL can fix stock counts, pick errors and slotting with a spreadsheet and ChatGPT or Claude reading your exports, and often should start there. Warehouse-system add-ons such as Cin7's forecasting, from a $349-a-month base plan, make sense once barcode scanning is in place and volumes justify them.
Scanning, not intelligence, removes most pick errors. An AI add-on cannot correct a wrong item that nobody scanned. So the order of purchase is: location labels and barcodes first, a system that enforces scans second, and only then anything that predicts or recommends. The sections below price each stage and show what the cheap route can already do.
Three jobs, three different tools
"Warehouse AI" covers unrelated problems, and the right tool depends on which one hurts. Stock counts are an accuracy problem: the system says 40 and the shelf holds 37. Pick errors are a process problem: the wrong item or quantity leaves the building. Slotting is a layout problem: pickers walk too far because fast movers sit at the back. Each has a cheap diagnostic you can run this week.
| Job | What the data looks like | Cheap route | Paid route |
|---|---|---|---|
| Stock counts | Count sheet versus system quantity per SKU and bin | Export both, ask an assistant to rank variances and spot patterns by location, shift or supplier | Inventory app or system with scan-based cycle counts |
| Pick errors | Returns and complaint log with reason codes | Paste the log, ask for error rate per thousand lines and top repeat SKUs | System that forces scan-to-pick and flags mismatches at the bench |
| Slotting | Twelve weeks of order lines per SKU | Rank by pick frequency, check pairs often ordered together | System with slotting or forecasting tools, often an add-on |
What the warehouse systems actually charge
List prices below were read from vendor pages in November 2026. They move, and several vendors show discounts for the first year, so verify each page before you decide.
Cin7 Core lists Standard at $349 a month, Pro at $599 and Advanced at $1,199, all with barcode scanning and basic warehouse management for one warehouse location. Advanced warehouse management is an add-on on Pro and included on Advanced. Its AI pieces are separate: product-description generation is in every plan, while ForesightAI forecasting, advanced automations and document recognition are listed as add-ons. So the base price tells you nothing about the AI bill.
Sortly sits at the light end: a free plan with 100 unique items, then Advanced at $49 a month for 500 items, Ultra at $149 for 2,000 and Premium at $299 for 5,000. New customers who pay yearly get 50% off the first year, which works out at $24, $74 and $149 a month billed annually; after that, yearly plans carry a 20% discount. It is inventory tracking with scanning, not a full warehouse system.
Zoho Inventory prices by orders per month and number of locations. Its pricing page lists a free plan with 50 orders a month, one user and one location, then Standard at $29 a month (500 orders, two locations), Premium at $79 (3,000 orders, four locations, 500 bins per location), Plus at $129 (7,500 orders, six locations) and Enterprise at $249 (15,000 orders, ten locations), each per organisation and billed annually. Stock counting and barcode generation appear from Premium up, which matters more for this job than any AI label; the main plan list does not itemise AI features, so check the comparison page if that is what you are buying.
One warning on stock apps for online sellers: Shopify's own Stocky app stopped working on 31 August 2026, with purchase orders moving into Shopify admin. If a guide tells you to use it, it is out of date.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowThe spreadsheet and ChatGPT route, in practice
ChatGPT Plus lists at $20 a month, and Claude Pro is $20 billed monthly. You already own the data: the system or app exports CSV files, and the assistant reads them. Because a large export can exceed what an assistant handles reliably, summarise first with a pivot table, then ask questions of the summary.
Example: ranking count variances
Say a nine-person 3PL cycle-counts 300 bins a week and sees 27 variances. Export bin, SKU, system quantity, counted quantity, counter and date. A prompt that works:
Below is a cycle-count export. For each row compute variance (counted minus system). Then:
1. List the ten largest absolute variances with SKU and bin.
2. Group variances by aisle, by counter and by weekday, and say whether any group has a share of variances well above its share of counts.
3. Flag SKUs with a variance in two or more weeks.
Do not guess causes; only report what the numbers show.
DATA:
[paste rows]An illustrative reply: "27 variances from 300 counts (9%). Aisle C holds 14 of the 27 but only 60 of the 300 counts. Eleven variances are positive (more stock than the system says), 16 negative. SKU BX-204 appears in weeks 2 and 4." What you do next is human: aisle C may hold a fast-moving range, share a bin between two SKUs, or be where returns are put away unscanned. The tutorial on why your count and the system disagree walks through those causes. The assistant narrowed 300 bins to one aisle in a minute.
Example: measuring pick errors
Take 12,000 lines shipped in a month and 36 post-packing errors from returns and complaints: 3 per thousand lines. Paste the 36 rows with reason code (wrong item, wrong quantity, missing, damaged by packing) and ask for counts by reason and by pick zone. Illustratively, 21 wrong-item errors in the zone where two SKUs differ only by size suggests similar-looking products stored next to each other. Moving one of them is free. A forced scan at the bench might catch the rest, which is where a system pays: if a scan rule cut wrong-item errors from 21 to 3 a month, and each error costs about $15 in reshipping, returns handling and goodwill, that saves 18 × $15 = $270 a month, a figure to set against a $349 plan. Use your own cost per error, which is usually higher than owners guess once labour is counted.
Example: a first slotting pass
Export twelve weeks of order lines, with SKU, quantity and order number. Ask the assistant to count lines per SKU, rank them, and report what share of lines the top 20% of SKUs account for. In many warehouses the top fifth of SKUs produce most of the picks, though your figure may differ, and the point is to compute yours. Then ask for pairs of SKUs that appear together on more than 15% of the orders containing either. A sample result: 48 of 240 SKUs generate 71% of lines; SKUs A-17 and A-18 ship together on 34% of their orders. The action is physical: move the 48 to the first two bays at waist height, put the pair side by side, and time a sample of twenty orders before and after. If a picker covers 2.1 minutes per order before and 1.7 after, that is 0.4 minutes saved on 300 orders a day, 120 minutes, two hours of labour daily.
Where a system's AI add-on starts to pay
Forecasting add-ons such as Cin7's ForesightAI predict demand per SKU to suggest reorder quantities. They need clean history, ideally a year or more, and stable item codes. A new seller with eight weeks of data gets little from one. They pay when stockouts and overstock are both costly, for example when a hundred fast SKUs sell out every few weeks while cash sits in slow ones. For earlier help, the approach in AI for ecommerce inventory and reorder points sets reorder points from a sheet.
Document recognition is another add-on: it reads supplier delivery notes at goods-in. Worth pricing if you receive many different suppliers' paperwork; pointless if two suppliers send you a standard sheet.
A decision table by warehouse stage
| Where you are | Buy this first | Leave for later |
|---|---|---|
| Under 100 orders a day, paper pick lists, no bin labels | Bin labels, barcodes and a scan-capable inventory app | Any AI add-on |
| 100 to 500 orders a day, scanning in place, errors above 2 per 1,000 lines | A system that forces scan-to-pick; assistant for monthly error analysis | Forecasting |
| Several clients, shared racking, billing disputes | A system with client-level stock and billing data; see the 3PL billing tutorials | Slotting tools until the data is clean |
| A year of clean history, frequent stockouts | Forecasting add-on trial on the top 100 SKUs | Full rollout across all SKUs |
A 3PL has a billing problem as well as a stock problem
If you run a 3PL, accuracy errors become invoices disputed by clients, so your tool choice should include how easily the system exports storage days, pick counts and special-handling events. An assistant can reconcile those exports against your rate card in minutes; see 3PL billing reconciliation with system data. Clients ask the sharper questions too, and the checklist in questions to ask a 3PL about its AI and automation shows what they will want answered, such as how errors are measured and who sees the data.
A fuller example: a small 3PL over one month
Picture an eight-person 3PL storing stock for six small online brands, shipping about 280 orders a day over 22 working days. Illustrative numbers again. At month end the owner has three complaints: client B says 14 units went missing, client D was charged for a pallet position it says it never used, and the packing bench is behind by mid-afternoon.
Instead of buying anything, the owner exports four files: cycle counts, returns and complaints, order lines and the storage ledger. Each goes to the assistant with a narrow question. The count file shows that 19 of 31 variances sit on client B's racking, all in two bays near the returns table, which points to returns put away without a scan. The complaint file shows 41 errors across 6,160 lines, 6.6 per thousand, and 25 of them are look-alike pairs in client A and client C ranges. The order file shows that 52 SKUs produce 68% of lines yet sit on the top two levels of the furthest aisle. The storage ledger shows the disputed pallet position was assigned to client D on the 3rd and never released after a stock transfer.
The fixes cost almost nothing: a scan-in rule at the returns table, one look-alike pair in each of five locations moved apart, the 52 fast SKUs moved to the nearest bays, and a weekly check for pallet positions with zero stock. The owner then has a clear brief for any system vendor: forced scans for returns, location-level alerts, an export of storage days. That brief is worth more than a demo, because it turns a sales call into a test of whether the product does the things you need.
Questions that expose weak AI claims in a vendor demo
Vendors will show polished forecasts and slotting heat maps. Ask the following and write down the answers.
- Which of these features are in the plan I am pricing, and which are add-ons with a separate price per month or per user?
- How many weeks of my own order history will the forecast or slotting tool need before its output is useful?
- Can I run it on my real SKUs during a trial, and what do I get back that I can compare with my own spreadsheet result?
- What happens when the suggestion is wrong, and can a person override it without breaking the record?
- Can I export storage days, pick counts and error logs in a standard file, so I am not locked in?
- Where does my clients' data sit, and is it used to train anything shared with other customers?
A vendor who cannot answer the first two without checking is telling you the AI is not the core of the product, which is fine, but then do not pay for it as if it were.
How to tell the change worked
Pick three numbers and measure them before and after, using the same definitions each month: count accuracy (bins with no variance divided by bins counted), pick errors per thousand lines, and minutes per order at the bench for a sample of twenty. In the fuller example above, a month later the count accuracy might move from 90% to 95%, errors from 6.6 to 3 per thousand and bench time from 2.1 to 1.7 minutes. If none of the three move, the change was cosmetic. If one moves and another worsens, you have probably shifted work rather than removed it, for example forcing scans slowed the bench but cut errors, in which case the question is whether the error saving beats the minutes lost.
Mistakes that show up in the first month
Counting only the bins that are easy to reach, which flatters accuracy. Changing the definition of an error halfway, which makes the trend meaningless. Letting an assistant "clean" the export by deleting rows it considered outliers; ask it to flag, never delete. And moving fast sellers to a prime location without relabelling, so pickers still go to the old bay out of habit. A week of floor walks with the layout sheet in hand fixes that one.
Goods-in: the quiet source of count errors
Many count variances are born at the dock, when a delivery is checked against the wrong document or put away to a location nobody recorded. A short prompt run weekly on the goods-in log catches the pattern. Export the receipts with supplier, expected quantity, received quantity, put-away location and time between arrival and put-away, then ask the assistant to list suppliers whose receipts differ from expected in more than 5% of lines, and any pallets that sat unlocated for more than a day.
An illustrative reply: "Supplier K short-shipped on 6 of 11 deliveries, by 3 to 12 units; four pallets waited between 26 and 40 hours for put-away, all received on Fridays." Two actions follow, neither involving software. Raise the short shipments with supplier K using the receipt records, and give Friday afternoon a named person for put-away. The assistant found the pattern; the owner picked up the phone.
If you later buy a system, goods-in is where scanning pays back fastest, because every pallet receives a location the moment it is received. That is also where a document-recognition add-on would read the supplier's delivery note. Price it against the real number of different supplier formats you handle, not the number of deliveries.
Keeping the data safe
A stock export can reveal a client's sales volumes, so treat it as confidential. Paste only the columns you need, remove client names or replace them with codes, and use a plan whose data controls suit your contracts. OpenAI and Anthropic say their business plans do not train on your content by default; their consumer plans have a model-training switch in settings. Never let an assistant write back into the live system: let it analyse exports and recommend, and let a person enter changes.
A four-week trial before you spend
Week one: label bins and count the 50 fastest SKUs; record errors by reason. Week two: run the variance and pick-error prompts above on the exports and fix whatever is free, such as moving look-alike SKUs apart. Week three: re-slot the top SKUs and time twenty orders. Week four: rerun the numbers. If errors fell and picking sped up, you have your baseline; if scanning discipline is the remaining gap, shortlist a system and ask the vendor for a trial on your own SKUs, with the add-on priced separately in writing. If you want help deciding what to test first, that is the sort of thing my AI implementation consultation covers.
Warehouse AI questions before you spend
Can ChatGPT run my warehouse?
No. It cannot see live stock or scan a bin. It is useful for analysing exports: count variances, pick-error logs and sales velocity by SKU. Keep the system or sheet as the record and use the assistant to read the data and suggest changes.
Do I need a warehouse management system with AI?
Not at small scale. Below a few hundred orders a day, barcode scanning and a disciplined sheet or inventory app usually fix more errors than any AI add-on. Buy a system for scanning and location control first; treat forecasting add-ons as optional.
What does slotting mean?
Slotting is deciding which product lives in which location. Fast sellers go close to the packing bench at waist height, heavy items low, and items often ordered together side by side. A sales export and a simple ranking is enough to start.
How do I measure pick errors?
Divide wrong-item, wrong-quantity and missing-item picks by total lines shipped over the same period, usually shown per thousand lines. Count only errors found after packing, from returns or complaints, and log the reason code.
Further reads
- How to Run a Stocktake Faster With AI and a Phone Scanner — Count faster with a phone as the scanner.
- 3PL Quote: What 3PLs Charge and How to Compare Every Fee With AI — What 3PLs charge and how to compare quotes.
- 3PL Pricing Calculator: AI Quotes Storage, Pick and Pack Rates — Build storage and pick-and-pack rates into a quote.
- 3PL Client KPI Reports: AI Builds Weekly Reports From WMS Exports — Weekly client reports built from system exports.
- Best AI for Ecommerce Operations: Orders, Stock, Returns, Suppliers — Orders, stock and returns across a small online business.
- AI Robots for Warehouses: Do They Pay Off for a Small Operator? — Calculate payback for warehouse cart-assist robots before signing a RaaS contract, comparing walking time savings against rental fees.
- 3PL Client Onboarding Checklist: AI-Drafted SOPs and SLAs — The 3PL's own onboarding checklist for a new client: what to collect, how AI drafts the client SOP and SLA, and who signs what before the first pallet arrives.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Cin7 pricing page, Sortly pricing page, Zoho Inventory pricing page, Shopify help page on moving from Stocky, ChatGPT and Claude pricing pages, OpenAI and Anthropic privacy pages, read in November 2026.