Build each picking or packing procedure as a one-page training card for a single task. Film or photograph your best operator doing it, describe the steps aloud, give the notes and photos to an AI assistant to draft five to seven numbered steps, translate where needed, and have a supervisor test the card on someone who has never done the job.
The card is the standard operating procedure, shrunk to what a person can follow at a bench. Long manuals sit in a drawer, whereas a card with three photographs gets read at the moment of doubt. AI saves the time spent on writing and layout, but it cannot see your floor, so the observation and the sign-off stay with people.
Which five tasks to write first
Do not document the whole building. Start with the tasks where errors cost money or new starters struggle, and choose them with data. Pull a month of returns, complaints and re-pick records and group them by the task that caused them. In a typical small operation the same few appear again and again.
A sensible first set for a pick-and-pack site, in order of how often they matter:
- Pick a line from a bin and confirm it by scanning.
- Pack a single-item order in the right box.
- Pack a fragile or liquid item.
- Seal, label and stage for the courier.
- Process a return and put the item back into stock.
Five cards is enough to start. A site with 8 to 12 staff can write them in about a day of supervisor time, and the evidence from the first month tells you which five to write next. The pick-error tracing tutorial shows how to find the task behind a recurring mistake, which is a good way to choose.
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 nowCapture the task the way your best operator does it
The raw material is observation. Pick the person whose error rate is lowest and ask them to do the task as they normally do, narrating as they go. You are looking for the small habits that make them accurate: which hand holds the scanner, where the tape gun lives, when they check the label.
The photo shoot list
Take photos with a phone held at working height, not from above. For each task, shoot a first frame showing the starting position, one frame per decision point, and a final frame showing what correct looks like. Typical counts are four to eight photos per card. Use a made-up test order so no real customer details appear, and crop out faces and anything else that is not part of the task.
- Start state: the empty tote, the open box, the scanner in its dock.
- Decision points: where the person checks a label, chooses a box size or inspects an item.
- Finish state: the sealed box with the label in the right position.
- Wrong state: one photo of a common mistake, such as a label across a seam, marked with a red cross.
The wrong-state photo is the most useful of the lot. People recognise what not to do faster than they absorb a rule, and it makes the card concrete in a way words do not.
Narration as a source of notes
Record a one-minute voice note on the phone while the operator works, saying each step and the reason for it. Transcribe it with the phone's built-in dictation or an AI tool. Spoken notes catch the unwritten parts, such as "I always check the box for old tape before I use it", that a typed brief would miss.
From notes and photos to a draft card
Give the AI the transcript, the photo list with a one-line description of each photo, and a fixed format. The format matters: the same layout on every card lets people learn where to look. This prompt works in ChatGPT, Claude or Gemini; if you upload the photos themselves, ask the assistant to describe what it sees and correct it where it is wrong.
You are helping me write a one-page training card for a warehouse task. Use the transcript and photo notes below. Output: a title in the form "Task: ...", a one-line purpose, then 5 to 7 numbered steps. Each step is one action in under 15 words, starting with a verb. After the steps, add "Check before you move on" with 2 to 3 yes/no checks, and "Common mistake" with one sentence. Do not add steps that are not in my notes. If a step is unclear or missing a detail, list it under "Questions for the supervisor".A reply for a fragile-item card might look like this (illustrative):
Task: Pack a fragile single-item order. Purpose: the item arrives unbroken. 1. Scan the order label and confirm the item. 2. Choose the box with 5 cm of spare space on every side. 3. Lay a 3 cm layer of paper fill in the base. 4. Wrap the item in two turns of bubble wrap and tape the end. 5. Place it in the centre and fill the gaps to the top. 6. Close the flaps and shake gently; nothing should move. 7. Seal with three strips of tape and apply the label. Check: Does the item stay still when shaken? Is the label flat? Common mistake: filling only the base and sides, leaving the top loose. Questions: Which box codes count as fragile-rated?
What would you fix? Step 2 gives a measurement the operator never mentioned, so check it against how they actually pack: the figure may be right, or the assistant may have invented it to sound precise. Step 7 combined three actions, so split it if the card has room. The open question about box codes is the useful part, because someone has to answer it before the card is issued.
A card layout people read
Put the layout in a template once, then reuse it for every task. A page that works at a pack bench has the title at the top, the photos down the left, and the steps beside them, with no paragraph longer than a line. Here is what each zone carries.
| Zone | Contents | Limit |
|---|---|---|
| Header | Task name, owner, revision date, language | One line |
| Steps | Numbered actions, verb first | 5 to 7 |
| Photos | One per decision point, numbered to match the steps | 4 to 8 |
| Checks | Yes/no questions before moving on | 2 to 3 |
| Common mistake | One photo with a red cross and one sentence | 1 |
| Safety line | Only what your safety adviser has approved | Optional |
Print at A4 or letter size on a laminated sheet at the bench, with a smaller copy on a lanyard card for walking tasks such as picking. Add a QR code that links to a short video of the same task for people who prefer to watch. The screen recording tutorial shows how to produce that video and its steps with one recording.
Writing the same card in every language on the floor
If your team speaks more than one language, a card in the wrong one is a card nobody reads. Translation is where AI saves the most time, as long as you treat the output as a draft. Run it in three steps: translate, translate back, and have a person check.
Build a glossary first
Warehouse words travel badly. A "tote", a "bay", a "pick face", an "SKU" and a "cage" may each be called something different in your building than in a dictionary. Write a list of ten to twenty terms with the word you use and, where you know it, the word your team uses in their own language. Give the glossary to the AI with the card.
Translate the card below into [language]. Use the glossary for every term it contains and do not translate product codes or numbers. Keep each step to one short sentence in the imperative. After the translation, give a back-translation into English, line by line, so I can check meaning.Compare the back-translation with your original. A step that says "place the item in the centre" and comes back as "put the object at the middle of the box" is fine. A step that comes back as "put the item in the tray in the middle" has changed meaning, so flag it. Then ask a fluent colleague to read the translated card aloud at the bench. If they hesitate on a phrase, rewrite it.
Photos and icons carry the meaning
Where words let you down, the pictures hold the message. Keep the layout identical across languages so that step 4 is in the same place on every card, and add simple symbols: a tick for a check, a cross for a mistake. A person who reads English slowly can still follow a card by photo number.
The supervisor's sign-off
An AI-written card is a draft until a supervisor has checked it against the floor. Make the review a short list so it happens every time and does not rely on judgement alone.
- Does each step match what the best operator does, step for step?
- Is every measurement, box code and scanner prompt on the card correct today?
- Does each photo show the right equipment and no customer data or faces?
- Is anything safety-related on the card approved by the person responsible for safety?
- Do the translated versions match the English, checked by a fluent reader?
- Has someone who has never done the task followed the card cold?
The last test beats the others. Hand the card to a new starter or someone from another department, say nothing, and watch them do the task. Note every pause and every question. Each one is a place where the card is unclear, and fixing five of them takes ten minutes. After the second test, the card is ready to issue.
Keep safety out of the AI's hands. Manual handling, forklift rules and anything involving a hazard should come from the person accountable for safety, written in their words. The forklift checklist tutorial explains how to use AI for the form layout while the content stays with the competent person.
A card that went stale in a month
An illustration of the commonest failure. A site issued a packing card that told people to use the medium box for orders under 2 kg. Three weeks later the supplier changed the box size and the stock room put the new boxes on the old shelf. People followed the card, the items rattled inside, and breakages rose from one in 400 orders to one in 150 before anyone connected the two.
The card was accurate on the day it was issued. What was missing was an owner and a trigger. Add two lines to every card: the owner's name and "Review when any box, scanner, label or layout changes". Put a short review into the monthly supervisors' meeting, and read one card aloud each time. A stale card is worse than none, because it looks authoritative.
A second card, filled in: picking one line from a bin
A packing card shows the idea; a picking card shows how short the steps should be. Here is a filled-in card for the most common task on the floor, as the assistant might draft it and a supervisor would tidy it (illustrative).
- Read the location on the scanner screen.
- Walk to the bay and find the bin with that exact code.
- Scan the bin label before you touch the product.
- Take the quantity shown, no more and no fewer.
- Scan the item barcode; wait for the green tick.
- Place the item in the tote, label side up.
Check before you move on: did the screen show a green tick, and does the tote count match the screen? Common mistake: scanning the item barcode without scanning the bin first, so a wrong item in the right bin passes. The photo for step 3 shows a hand holding the scanner at the bin label, and the mistake photo shows a hand reaching into a bin with the scanner still on the belt. Step 3 is the one that prevents most mispicks, which is why it exists as a separate line.
What the card says when the task goes off script
Cards written for the normal path leave people stranded when something is different. The bin is empty, the barcode will not scan, the item looks damaged or the box is torn. Ask each operator what they do in those cases and add a short "If this happens" block to the card, with no more than three lines.
| Situation | What the card says |
|---|---|
| Bin is empty but the screen shows stock | Flag the location on the scanner, tell the supervisor, move to the next line |
| Barcode will not scan | Try twice, then key the code once; if it fails, set the item aside with a red tag |
| Item is damaged | Do not pack; place in the red tote and report the order number |
| Order has more than one box | Mark each box 1 of 2, 2 of 2 and check the label count |
This block is where the assistant's questions help most. Ask it, "What situations might make a person stop following these steps, and what do I need to answer for each?" Then answer only the ones that happen on your floor. A card that handles the three commonest exceptions prevents most of the shortcuts people invent when the card fails them.
The cold-read test, logged
Record the test on a one-line log so the review is evidence and not an opinion. For the fragile-item card, a trial might read as follows (illustrative): reader had never packed fragile items; paused at step 2 to ask which box; placed the fill but not on the base; shake test passed; time 4 minutes 10 seconds against 2 minutes 30 for an experienced packer; changes made: step 2 now names the box code, step 3 photo shows the base layer. A second trial with a different person confirms the changes worked. Two logs per card, one before and one after the changes, is enough to prove the card does its job.
Rolling the cards out in a week
A small site can launch five cards in a working week without stopping the floor. Day one, observe and photograph. Day two, draft the cards with the assistant and a supervisor. Day three, translate and check. Day four, trial each card on a starter or a volunteer and fix what they struggle with. Day five, print, laminate and place them, and brief the team on where each lives and why.
Brief people as a group for ten minutes, and explain that the cards come from what the best operators already do. A card that feels like a surveillance tool will be ignored. A card that reads like a colleague's tips will be used. Invite changes in the first fortnight, and log each one so people see their suggestions land.
The cards also feed other planning. When you know a new starter can pack a standard order in their first morning, the 75% productivity assumption in a labour forecast can rise, and agency cover costs less.
Measuring whether the cards help
Track three things from the week before launch. First, errors per 1,000 order lines: mispicks, wrong boxes and damaged goods, taken from the same records you used to choose the tasks. Second, days for a new starter to reach standard pace, measured by the supervisor on a simple tick sheet. Third, supervisor minutes spent on repeat questions, from a quick tally in the first month.
A quick sum shows what is at stake. Suppose a site takes ten new starters a year, each needing eight hours of supervisor coaching to reach standard pace. If the cards cut that by six hours, that is 60 supervisor hours a year, about $1,440 at $24 an hour, against roughly 18 hours to build twelve cards. The saving repeats every year, while the cost occurs once plus a small review. Replace the figures with yours before you quote them to anyone.
If the numbers do not move after two months, read the cards again as a new starter would. Usually the steps are fine, and the photos are too small, the card is in the wrong place or nobody was told it existed. Move the cards to where hands are, not where walls are free.
Keeping cards alive without a project
Set up three habits and the library maintains itself. A shared folder holds the editable master of each card and the translations, named by task and language. One supervisor per card family owns updates. And every card carries a QR code to the master, so the printed copy can be checked against the current version in ten seconds.
When a process changes, edit the master, ask the AI to produce the translated versions again with the same glossary, check them as before, and swap the printed copies in one pass. Remove the old ones the same day. For a wider approach to writing procedures from rough notes, see the general SOP tutorial, which this one narrows to the bench-level cards a warehouse needs.
Questions about warehouse training cards
How many steps should one card have?
Aim for five to seven. A card read at a pack bench is glanced at between actions, so every extra step lowers the chance it is followed. If a task needs more than eight, split it into two cards, such as 'prepare the box' and 'seal and label', and give each its own photo set.
Can I trust AI to translate a card?
Use it for a first draft, then have a fluent speaker who works in your warehouse read it. Ask the AI to translate back into English as a check. Warehouse words such as bay, tote and pick face often translate literally into something odd, so keep a short glossary of your terms.
Should the card show faces or customer labels?
No. Crop photos to the hands, the product and the equipment. Faces raise consent issues, and customer labels, order numbers and addresses are personal data you should not send to an AI tool or print on a wall. Use a dummy order with made-up details for photographs.
Who owns the cards after launch?
Name one supervisor per card family, such as picking or packing. Their job is to update a card when a box, a scanner or a layout changes, and to review each card on a set rhythm. A card with no owner goes out of date within months.
Further reads
- Warehouse Slotting: AI Analyses Pick Frequency From a WMS Export — Better slotting shortens the pick card for the fastest movers.
- Cleaning Business SOP: Write Procedures and Checklists With AI — Another floor-level procedure set written with AI.
- 3PL Pricing Calculator: AI Quotes Storage, Pick and Pack Rates — See how pick and pack tasks are priced for clients.
- AI for Warehouse Loading: Plan Pallets and Trucks From Orders — Next card family: loading and despatch.
- Cycle Count Template: AI Plans Counts by ABC Class — Count discipline that backs up the pick-confirm step.
- Dock Scheduling for a Small Warehouse: Slot Rules and AI Reminders — Goods-in cards pair with a tidy dock schedule.
- Fast Food Employee Training: Reference Cards and Quizzes With AI — Build crew training cards, ten-question quizzes and a day-one-to-day-thirty checklist from your franchise manual with AI, keeping brand figures exact.
- ISO 9001 Documentation: Hire a Consultant or Draft It With AI? — Which ISO 9001 jobs AI can draft for a small manufacturer, which ones need a consultant, and a costed worked example of the split.
- 3PL Client KPI Reports: AI Builds Weekly Reports From WMS Exports — A weekly one-page report per 3PL client, built from WMS exports: three KPIs defined exactly, charts that do not mislead, and AI commentary a person signs off.
- 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.
- Driver Onboarding Checklist: From Offer to First Route With AI — A stage-by-stage checklist for onboarding a new driver, with AI-drafted letters, reminders and induction packs and care for personal data.
- AI for a Small 3PL: Six Jobs Ranked by Payback — Six small-3PL jobs ranked by payback, with an effort-versus-payback table and a worked billing example.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: the author's own method; the cards and figures shown are illustrative examples.