Before paying for supply chain AI software, a small importer needs four things: clean purchase-order and stock data, a way to capture real supplier lead times from emails, a live link to the stock your sales channel shows, and reorder suggestions that show their working. Without those, any forecast just automates guesses.
The lists you find when you search for this are mostly software directories and feature roundups, written to compare products. A small importer needs the opposite order: decide what the business must be able to do, then see which product can. Your problems are narrower and more expensive than a planning team's: a 60 to 80 day lead time, minimum order quantities, a deposit tied up for months, and supplier updates that arrive as emails.
Why platform roundups answer a different question
A roundup compares breadth: forecasting models, network planning, visibility dashboards, integrations by the dozen. An importer with a few hundred products makes one decision at a time: when to order each product, how many to order, and how much cash that locks in for the next three months. Software helps only if it improves one of those decisions, and each decision depends on an input you probably have in a messy state today.
So the useful question is not "which platform ranks first" but "what does this tool need from me, and what does it hand back?" Problems tend to surface in a fixed order: first your data, then your lead times, then supplier emails, then stock counts, and only after those the reorder itself. Pricing may arrive only after a sales call, so you also need a way to compare totals and a trial that proves the tool before you sign.
Check your data before you check any software
Every forecasting tool starts from five sets of data. If you cannot produce them in an afternoon, the software will not fix that; it will import the mess and present it more confidently.
| Data | Where it usually lives | What breaks without it |
|---|---|---|
| One code per sellable item, plus the supplier's own code | Sales channel, supplier price lists | The tool cannot match a shipment to a product |
| Purchase orders with order, confirmed, shipped and arrived dates | Email, spreadsheets, accounting | Actual lead time cannot be calculated |
| Supplier terms: minimum order, pack sizes, payment terms, quoted lead time | Proforma invoices, emails | Suggested quantities ignore minimums |
| Stock on hand by location, plus stock in transit | Sales channel, warehouse sheet | Orders are doubled or missed |
| Weekly sales by item, with stockout days marked | Sales channel exports | Demand looks lower than it was when you were out of stock |
The last row is the subtle one. If a product sat at zero for nine days, its sales for that month understate demand, and a tool that does not know about the stockout will forecast lower and reorder lower. Mark those days, or ask the vendor how the tool treats them.
The quick test: can you work out the real lead time of your last ten orders in fifteen minutes? If not, start with that sheet, not a demo. Tidying product codes and supplier sheets is the same job as in cleaning messy data, and an AI assistant is good at flagging duplicates and near-matches for you to confirm.
Lead-time tracking: quoted days against actual days
A lead time is not one number. Production, booking, transit, port clearance and delivery to your warehouse each slip separately, and suppliers quote the version that sounds best. What matters is the number from the day the supplier confirms your order to the day the goods are received, measured on your own history. Here are six illustrative orders for one supplier, each quoted at 65 days.
| Order | Quoted days | Actual days | Slip | Where it slipped (illustrative) |
|---|---|---|---|---|
| A | 65 | 71 | +6 | Production waited for the deposit |
| B | 65 | 69 | +4 | Port congestion |
| C | 65 | 78 | +13 | Factory holiday, then a missed vessel |
| D | 65 | 66 | +1 | Nothing material |
| E | 65 | 74 | +9 | Inspection failed, rework |
| F | 65 | 68 | +3 | Customs hold |
The average actual lead time is 71 days and the worst is 78. Every order arrived later than quoted, by six days on average. A reorder point built on 65 days would have been wrong six times out of six. The planning figure I would use is the average of your own last six to ten orders, with safety stock sized to cover the gap to your worst case.
What to ask a vendor: does the tool calculate lead time per supplier from your purchase-order dates, or only accept a number you type? Can it separate stages, so you can see that three of those slips happened before the goods left the factory? A tool that takes a typed lead time is a calculator with your optimism built in. To build the tracking sheet yourself first, building a supplier chase sheet with AI shows the columns and the chase routine.
Supplier emails: the fields an AI reader must extract
Most lead-time truth arrives as emails and attachments: production finished, vessel booked, shipment split, new dates, a revised invoice. This is where AI genuinely helps, because reading messy text into a table is something language models do well. It is also where a quiet error costs most, because a wrong date looks as tidy as a right one.
The fields worth extracting are few: your purchase-order number, the event (production finished, vessel booked, departure, arrival, split shipment, invoice, payment request), the date as written, the quantity affected, and the sentence it came from. Here is an illustrative email and a prompt that handles dates safely.
Subject: PO 2291 - production finished, shipping update
Hi [your name],
Production for PO 2291 finished on 03/04. The vessel is booked: ETD 12/04, ETA 14/05.
The 400 matte black units are held back and will ship separately, ETD 20/04.
Invoice INV-8831 attached. Please arrange the balance payment before loading.
Best regards,
[supplier contact]
Read the supplier email below. Return a table with one row per fact:
purchase-order number, event, date exactly as written, date as day and month name,
quantity affected, and the sentence it came from.
If day and month could be swapped (both numbers 12 or below), write AMBIGUOUS in the
date column and say what other evidence in the email points to. If a fact is not stated,
write NOT STATED. Do not calculate or assume anything.
A good answer flags the two ambiguous dates, notes that 14/05 can only be day-first, and asks you to confirm. A careless one reads 12/04 as 4 December but 14/05 as 14 May in the same table, so the goods arrive almost seven months before they leave. The check that catches it is a rule, not an eye: arrival cannot precede departure, and transit time outside a plausible range gets flagged for a person. Ask any vendor what validation runs after the AI reads an email, and whether the source sentence is shown beside every extracted date.
Two further habits help. Ask suppliers to put your order number in the subject line, and keep the proforma invoice in the same thread, so that the numbers on it can be compared with your order; checking a proforma invoice against a purchase order with AI covers the mismatches. And keep a person approving anything that changes a plan, such as a new arrival date on a purchase order.
You can build a basic version without buying a platform. Zapier's Professional plan lists 750 tasks a month at $19.99 billed annually or $29.99 billed monthly, and Make starts at about $9 a month on credits; AI steps can use more than one task or credit each, so check each vendor's current rules before sizing a plan. Sixty supplier emails a month with three automated steps each would be about 180 tasks, well inside 750 even if the AI step counts for more than one.
Inventory sync: six questions about what the software reads and writes
Inventory sync sounds like a tick-box feature and behaves like a long list of rules. Put the questions below to every vendor, and write down the answers rather than trusting a demo.
| Question | A good answer | A warning sign |
|---|---|---|
| Which fields do you read from my sales channel, and which do you write back? | Reads orders and stock; writes only receipts I approve | "Everything syncs automatically" |
| Is stock in transit counted, and where does it show? | Per order, with an expected arrival date | Only on-hand stock is visible |
| How often does it sync, and what happens when two systems disagree? | A stated interval and a stated rule for conflicts | "Near real time" with no detail |
| How are bundles, multipacks and variants mapped? | Components are deducted when a bundle sells | Each bundle is its own product |
| Does it handle more than one location, or a third-party warehouse? | Counts by location, with transfers | One stock pool only |
| Does it create purchase orders in my system, and can I export them? | The order exists in both systems and links back | I retype it into my sales channel |
Shopify is a useful reference point because its purchase-order pages are public. In its admin you create a purchase order for a supplier, mark it Ordered once the supplier confirms, then create an inventory transfer from it to track the shipment and receive the stock. Shopify's former Stocky app stopped working on 31 August 2026, its history could not be imported, and Shopify's purchase orders do not email themselves to suppliers, so anyone who relied on Stocky now starts from the built-in orders and adds something around them. Whatever you buy has to fit that pattern or replace it. When a vendor says its tool "creates purchase orders in Shopify", ask to see one appear on Shopify's own purchase-order screen in the demo; an order that lives only inside the tool, with incoming stock adjusted some other way, is a different thing. Stocky alternatives for purchase orders and AI reordering compares the routes.
One reorder from start to finish: the sums a tool must reproduce
Take one product (all figures illustrative): a 500 ml insulated bottle in matte black. It sells 24 a day on average. You hold 2,310 units and nothing is on the water. The supplier quotes 65 days, your own history says 71 on average and 78 at worst, the supplier sells in lots of 1,200 units, which is also its minimum order, and the unit cost is $3.85. Safety stock is 14 days of sales: seven days to cover the gap between your average and worst lead times, and seven as a cushion for a busy week. For the maths of reorder points in general, setting reorder points with AI has the basics; here is the importer's version.
The reorder point is daily sales multiplied by lead time, plus safety stock. With 71 days it is 24 times 71, plus 336, which is 2,040 units. You hold 2,310, so you are 270 units above it, which is 11 days of sales. Order within 11 days. Now see how the answer moves when the inputs do.
| Scenario | Daily sales | Lead time used | Reorder point | What it means |
|---|---|---|---|---|
| Baseline | 24 | 71 days | 2,040 | Order within 11 days |
| Supplier email says production starts 6 days late | 24 | 77 days | 2,184 | Order within 5 days |
| Seasonal forecast adds 25% to sales | 30 | 71 days | 2,550 | Order today; already about 8 days past the trigger |
| Supplier's quoted lead time used instead | 24 | 65 days | 1,896 | Looks safe for 17 days, then lands with 192 units left if the real lead time is 71, and 24 if it is 78 |
Three lessons sit in that table. A supplier email can move the order date by six days, so a tool that never reads emails cannot keep this number honest. A forecast change can flip the answer from "wait" to "now", so the tool must let you enter the reason for an override and show the arithmetic that follows. And the quoted lead time looks fine until the goods arrive, which is exactly why you test with your own history.
The quantity comes next. Ninety days of sales is 2,160 units; rounded up to whole 1,200-unit lots, you order 2,400. At $3.85 that is a $9,240 order. With a 30 per cent deposit, $2,772 leaves your account at ordering and $6,468 before loading, and freight and duty come on top at arrival. A tool should show that cash figure beside every suggestion, because a correct reorder you cannot afford is still a problem.
This is your demo script. Load these inputs, or your real ones, and ask the vendor to show the reorder point, the order date, the quantity and the cash needed, then explain each. If the tool cannot reproduce 2,040, or cannot say why it shows something else, you cannot trust it on the other four hundred products.
A scoring sheet for any vendor demo
Score each product from 0 to 2 on the ten rows below, using what you see on your own data. The first three are gates: any zero there ends the evaluation, however good the rest looks.
| # | Criterion | 0 | 2 |
|---|---|---|---|
| 1 (gate) | Calculates actual lead time per supplier | Only a typed number | Calculated from your last 12 months in the demo |
| 2 (gate) | Shows its working on each suggestion | A bare number | Demand, lead time, safety stock, minimum order and cash needed |
| 3 (gate) | Counts stock in transit | Not counted | By order, with expected arrival |
| 4 | Reads supplier emails | No | Extracts dates and quantities and shows the source sentence |
| 5 | Syncs with your sales channel | File upload only | Two-way with documented rules |
| 6 | Handles minimums, pack sizes and container fill | No | All three |
| 7 | Records why you overrode a forecast | No | Reason and audit trail |
| 8 | Human approval before any order or email goes out | Sends automatically | Required by default |
| 9 | Exports all your data | No | Full export on the plan you would buy |
| 10 | Total price at your volume, in writing | "Contact sales" only | Includes usage, AI credits and integrations |
A score of 1 means partly. Out of 20, I would want at least 14 with no gate failures before paying. The exact threshold is yours to set, but write it down before the demos so a charming salesperson cannot move it.
What a four-week trial must prove
A free trial is only evidence if you set the test. Four weeks is enough.
- Week 1: load and reconcile. Import products, orders and stock. Compare the tool's lead times with your sheet for five suppliers. Every difference gets explained before you go on.
- Week 2: backtest. Pick five products that stocked out in the past year. Ask what the tool would have suggested 90 days before each stockout. Would it have warned you in time?
- Week 3: run in parallel. The tool suggests orders and you order your usual way. Compare dates and quantities weekly, and note which of you was right.
- Week 4: the email test. Forward 20 real supplier emails, including awkward ones. Count correct fields, and note every silent error, meaning a wrong value shown without a flag.
Decide the pass rule in advance. A reasonable one: zero silent date errors across the 20 emails, backtest warnings early enough for at least four of five stockouts, and reorder points within a few days of yours on products you understand well. Where you disagree, work out who was right before blaming the tool.
Pricing: what a stockout costs against the monthly fee
The sum that makes a price decision sensible is short. Suppose each bottle contributes $8.00 after landed cost and fees. At 24 a day, a stockout costs $192 a day in lost contribution, and a six-day stockout costs $1,152, before counting wasted advertising. A hypothetical tool at $149 a month costs $1,788 a year, so it has to prevent about one and a half six-day stockouts a year to pay for itself on this one product. Run that sum on your three best sellers; the answer will tell you whether you are shopping for a problem you have.
Prices in this category are built in different units: per user, per product, per order, per integration, or by AI credits. Ask for one number: the total per month at your volume, with usage caps, plus what happens when you exceed them. Ask also what the contract says about leaving: can you take a full export of your data, and is there a minimum term?
A cheaper first step: a sheet, a mailbox label and an AI assistant
If your gate checks show that your data is the main problem, buy nothing yet. Build the sheet: one row per purchase order with quoted and actual dates, one row per product with stock and weekly sales. Label supplier emails, run them through an extraction prompt like the one above, and paste the approved results into the sheet. An AI assistant then does the reorder arithmetic in the sheet, and you check its output against your own sums, as in the worked reorder.
That set-up costs little, and it teaches you what you actually need. After two or three months you will know which step still eats your time: lead-time chasing, email reading or stock counts. Then you can buy a tool for that step and measure it against the sheet. Delay alerts and reorders without an ERP shows the next step up, with alerts that tell you when a supplier update changes your plan.
The importer who buys first and cleans later usually ends up paying twice. The importer who cleans first can run the trial in the previous section and buy with evidence.
Questions importers ask before buying
Do I need an ERP before I add supply chain AI?
Not necessarily. With a few hundred products, one sales channel and a purchase-order sheet, an AI assistant and some automation may be enough for a long while. An ERP earns its place when several people raise orders, stock sits in more than one place, or accounting needs costs to flow from receipts. Buy it for those reasons, not because a roundup listed it first.
Can AI predict port delays for me?
Some visibility tools estimate arrival dates from vessel-tracking data, but for a small importer the dependable gain is simpler: reading the updates your supplier and forwarder already send, and flagging where they differ from your plan. Treat any delay prediction as something to test against your own past shipments, and keep a buffer sized from your own lead-time history rather than a vendor's average.
If my sales are steady, do I still need forecasting?
Probably less than you think. With steady demand, most of the value sits in accurate lead times and in seeing stock in transit, which you can get from clean purchase-order dates. Forecasting earns its fee when demand swings with seasons, promotions or marketplace changes, so test it on a product that behaves that way.
Where should supplier emails go so an AI can read them?
To a shared mailbox or a label, not a personal inbox, so the routine survives staff changes and holidays. Keep supplier invoices and packing lists in one folder. Before connecting any mailbox to a tool, read the vendor's data-handling terms, decide who can see what the assistant extracts, and check whether attachments are read as well as message text.
Further reads
- AI Supplier Management: Track Prices, Lead Times, and Risk — Track supplier prices, lead times and risk over time.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Shopify's help pages on purchase orders and inventory tracking; Zapier and Make plan pages for list prices; standard reorder-point and lead-time definitions.