Keying a retailer’s orders into your own system. Sending shipping and invoice details back through their portal. Setting up a new trade customer online with the right discount. We automated two of these jobs inside a real UK equipment supplier that fulfils orders for a national retailer, using an AI assistant working through the same screens a person would. Here’s how, so you can try it.
What we automated
1. Daily dropship orders. A national retailer that sells some of the business’s products online sends its orders through the retailer’s supplier portal. Every day someone had to raise each order in the ERP, ship it, then go back to the portal with the carrier, tracking number and invoice number. Now one line, “complete the daily run”, does all three phases. On its first clean run it shipped two orders, invoiced two and raised two new ones in 16 minutes, across both systems. The team reads a summary and deals with anything flagged.
2. New trade accounts. When a trade customer wants to order online at trade prices, a sales colleague emails the director. Now the director types “set up [customer]”. The AI reads the email thread, sets up the account and discount on the online shop platform, sends the login invite, checks it’s live, and leaves a reply as a draft in Outlook. A person reads it and sends it. Eight trade accounts have been set up this way so far.
No code was written for the order run. The AI works in the business’s own browser after a person signs in. It never signs in itself.
Step 1: learn it live, with the person who owns the job
Don’t write the process up front. Sit the AI next to the person who does the job and run a real case, one step at a time, with them watching and correcting. On the very first day, the owner caught two things no written guide would have: a standard description field that was missing, and a button the AI must never press, because a separate system does that job overnight.
Try this:
“I’ve signed in to [portal] and [system] in this browser. Take the first new order only. Before each step, tell me what you’re about to do and which value you’ll use, and wait for me to say OK. Never sign in, never press [risky button], never guess a value. At the end, list anything in our notes that was wrong or missing.”
Step 2: write every correction down, straight away
Keep one plain-English process document: the steps, the rules, the price list, the odd behaviour of the screens, and worked examples. The AI reads it at the start of every run. If a fix isn’t written in, it gets made again next time. Over about five weeks, the order run went from supervised to running a full day unaided.
Give every rule a reason. “Delivery charge = delivery price × quantity, because delivery is priced per unit” is much harder for someone to delete by accident than a bare rule.
Step 3: tell it what to do when it’s unsure
The most important rule in both automations is never guess. If a value isn’t on the order, in the price list or in the system itself, the AI stops on that item, flags it and moves on to the next. One problem order never holds up the rest. In the first five weeks this caught a duplicate order, an order with no shipment behind it, a gap between our price and the retailer’s total, and three stock shortfalls. All went to a person.
It also caught something you might not expect: on the first clean run, the system’s own default prices on both new orders were well above the agreed list price. Because the process names one price list as the only source of truth, both were corrected.
Step 4: save it so one line runs it
Most AI assistants, including Claude, Microsoft Copilot and ChatGPT, let you save a set of instructions so one short phrase runs the whole procedure the same way every time. Point them at your process document so you update one place.
For the trade accounts, the setup on the shop platform did a lot of the heavy lifting. The platform only allows four automatic discounts to be active at once, so most customers share one discount at the standard trade rate, applied to a product group that automatically leaves out delivery, warranties, service and dropship lines.
“Search my inbox for the latest email from [colleague] about [customer]. Read the whole thread, including attachments. Give me the customer name, contact email, rate and any special terms, and anything unclear. Don’t change anything yet.”
Step 5: keep a person on the last step
The AI handles the routine. People make the decisions. Duplicates, stock shortfalls and price gaps go to the process owner. Customer records are only ever added to, never overwritten. Every email is a draft that a person reads before it goes.
Your checklist
1. Pick one job that’s frequent, rule-based and spread across two systems.
2. Time it by hand for a week, so you know your “before”.
3. Run three real cases with the owner watching.
4. Write every correction into one process document, with a reason.
5. Write the risky actions as hard “never” rules.
6. Tell the AI to flag and move on whenever it’s unsure.
7. Name one source of truth for prices.
8. A person signs in to every system. The AI never does.
9. Drafts, never sends.
10. Save a report and a log line after every run.
One safety point before you start: use a business account for your AI assistant and check its data settings before sharing customer or financial details.
Build it yourself, or have us build it
If you want to build it yourself, our guide Orders and Accounts Without the Admin has every step, rule and check from both automations, the problems we hit and how we fixed them, 20 copy-ready prompts and the two procedures we actually use, ready to adapt.
If you’d rather have it done for you, our Order Admin Setup builds it inside your business and connects it to your own systems, so nobody re-keys data. £1,495 + VAT, fixed price, done in three weeks.
We’re not IT people. We’ve sat in the MD’s chair and built the systems behind it, and we’ll talk to you in plain business terms. Not sure where admin is costing you most? Take the free Growing Pains Check.
Leave a Reply