AI Agent for Wholesalers and Importers: Repeat Orders on WhatsApp Without Double Entry

The common mistake in wholesale is treating a sales rep's personal WhatsApp as a company asset. It looks entirely reasonable: the customer knows a person by name, writes to them directly, gets an answer even in the evening, and the order comes in. What is not visible is that three of the business's failure points sit in that same phone. Order flow depends on one person's availability, every order is captured twice, once in the chat and once when it is typed into the system after hours, and when a shop stops ordering nobody notices, because its ordering cycle is not recorded anywhere. Those three costs are paid every month, and all of them can be calculated.

The three costs of an order that arrives as a chat message

Double entry. Take your number of repeat orders per month, say 250, and multiply it by the time that really passes between the first message and a clean line in the system: clarifying units, verifying a suspicious quantity, looking up a SKU, typing. Six minutes per order is conservative, and that is already 25 hours a month. If the hour of the person doing it is worth 110 shekels, that is roughly 2,750 shekels a month nobody sees as a line in a report, because it is spread across five people and mostly happens between eight and ten in the evening.

Unit and quantity errors. In an industry where the same SKU sells as a single unit, as a case of 12 and on a pallet, "send me 10 cola" is an ambiguous sentence. Every single error triggers a return trip for a delivery vehicle, a credit note, an apology call and sometimes a return of goods that cannot be sold again. The cost of one error is almost always larger than the margin on that order line, which is why typing at ten at night is more than just wasted time.

Silence nobody measures. A B2B customer almost never leaves by announcement. They simply stretch out: from an order every three weeks to one every five, and then nothing. Without a place that records each customer's ordering cycle, that decay surfaces a quarter later, when you look at revenue and discover four shops have dropped off. We analysed this mechanism in depth in reviving leads and customers you forgot, and in wholesale it hurts more, because every shop is recurring revenue rather than a single sale.

What a repeat order actually requires from the agent, step by step

A repeat order looks like a short conversation, and in practice it is a sequence of seven steps that all have to happen the same way every time:

• Identification. The phone number the message came from, matched against the customer record, and from there the shop name, the delivery address, the price group and the payment terms. An identified customer should not have to introduce themselves every time.
• Anchoring in the previous order. Instead of an open question, "the same order as last time, or with changes?" That shortens most orders to two messages.
• Building the lines. SKU, description, quantity and unit. The agent asks units or cases rather than assuming, accepts an order dictated as a voice note in front of the shelf, and recognises the shorthand name the shop uses from the catalogue you approved.
• Sanity check. A quantity that deviates from that customer's history gets one verification question, not a silent order ten times too large.
• Trading rules. Minimum order value, the delivery day for that area, the cut-off time for next-day delivery, a promotion valid until a set date.
• Summary for customer approval. A clean list in one message, and an explicit confirmation before the order goes in.
• Handover inward. A card in WhaleBiz CRM with the lines, the files and the full history, and a human who approves it before it becomes a delivery.

An AI sales agent that captures and qualifies enquiries runs that sequence on WhatsApp, on Telegram or in the website chat, in Hebrew, Russian and English depending on the language the customer opened in, at the same response time at seven in the morning and eleven at night. The less discussed part is the other side of it: when every order is collected in the same structure, you get real data for the first time on who orders what and at what pace, instead of a sales manager's impression.

Price lists, credit and stock: where the red line runs

In wholesale there is no such thing as "maybe I misunderstood". A wrong order leaves on a truck, and a price stated by mistake turns into an argument over an invoice. So the boundary has to be defined at sentence level, not at intention level.

The agent does not confirm that an item is in stock, does not confirm outstanding balance or credit limit, does not approve a special price that is not on a price list, does not approve a return or a credit note, does not commit to an exact delivery time and does not interpret payment terms on a disputed invoice. All of those are financial commitments by the company, and they stay with a human. Anyone who insists gets an honest answer: the line has been recorded, supply is subject to confirmation, and the person handling it will come back to them.

What the agent does close on its own is most of the traffic: a repeat order from an approved catalogue, questions about packaging, an alternative SKU, minimum order value, delivery days per area, sending a price list or product photo as a file, following up on a shipped order, and opening a structured application for a new customer. A question beyond the boundary goes to a human with the full context, and the customer is told when someone will get back to them. That is the same logic we described in a digital consultant for complex products: take away from the expert everything that is not expertise, without pretending there is no expert.

The rules that are actually worth enforcing

Most wholesalers have two written rules that are almost never enforced, because enforcing them on the phone with a regular customer is an unpleasant conversation.

The first is the minimum order value. A shop that orders three items and asks for delivery generates a delivery cost higher than the margin on that order. The agent enforces this without tension: it states the amount still needed to reach the minimum, suggests items the customer usually orders, or offers to attach the order to the next delivery run in that area. The exact same conversation, scripted in advance, lands differently when it comes from the system rather than from a rep who is afraid of losing the customer.

The second is the cut-off time. If orders for tomorrow's delivery close at 16:00, every order that arrives at 18:00 and is handled anyway teaches the customer that the deadline is flexible, and makes the pickers work twice. An agent that replies immediately, records the order, notes that it will go out on the next delivery run and repeats the cut-off time enforces the rule without refusing and without losing the order.

The customer who went quiet for three weeks: follow-up by ordering cycle

The follow-up that matters in wholesale is not on a new lead, but on an existing customer whose ordering rhythm has broken. It only works if there is order history at customer level, and that is exactly what gets created when orders flow through the agent rather than through screenshots.

Three rules cover most cases. First, personal cycle: if a customer orders on average every 21 days and 32 have passed, they get a message that refers to their last order, not a generic broadcast. Second, an item that disappeared from the basket: a shop that ordered a product line every time and stopped, while the rest of the order continues, probably bought it elsewhere, and that is a call for the sales manager rather than an automated message. Third, seasonality and promotions: a reminder before a season and before a promotion closes, based on the dates you defined.

Two disciplines prevent spam: every message refers to a real data point from that customer's history, and any reply stops the sequence and hands off to a human. The number of sequences and automations is unlimited and there are no trigger meters, so following up on 400 shops does not make the plan more expensive.

Four ways to take an order from a shop, and what each one costs

The row worth stopping on is the second one. B2B portals fail not because of their features but because of adoption: a shop owner who orders from seven suppliers will not remember seven logins, and will go back to writing on WhatsApp. An AI agent is the only route that adds order without asking the customer to change a habit.

Where this meets the system you already have

There are three levels of connection, and it is worth knowing in advance which one you are on.

No integration. The catalogue, the price lists and the trading rules are loaded as approved documents the agent works from: up to 15 documents on Standard and up to 45 on Business. Orders accumulate as cards in WhaleBiz CRM with Excel export, and entry into your own system is done once, from a clean approved list rather than out of a chat thread. That alone removes most of the double-entry cost.

Automatic push. The Business plan includes API access and webhooks: every approved order is pushed to a destination you define, whether that is your order system, a middleware layer or a folder your system reads. This is the level at which the typing disappears entirely.

Two-way sync. Live stock, outstanding balances and price lists updated from your system in real time are custom development rather than part of a plan, which is why they are quoted before work starts and do not appear as a monthly price. Projects like that start from 10,000 shekels, and changes to existing systems are priced at 350 shekels per hour. The details are on the custom development and integrations page.

What decides which level you start at is not the budget but the state of the data: a catalogue with valid SKUs, units of measure and pack multipliers, price groups, minimum order values, delivery days per area and payment terms. Anyone who does not have that list clean discovers during setup that this, and not the bot, is the real problem.

What it costs and which plan to start on

A wholesaler starts on Standard, 490 shekels per month plus a one-time setup fee of 1,990 shekels, before VAT. That is the plan with WhatsApp through the official Meta API, sending files such as a price list or a product photo, and analytics with Excel export and import, and it includes 1,800 AI messages per month. The Start plan, 190 shekels per month plus 990 shekels setup, does not fit this industry for one reason: it has no WhatsApp.

A company that wants a separate agent for several divisions, replies on Instagram and Facebook, in-chat payment from a new customer with no credit line, or API access and webhooks for pushing orders, belongs on Business, 990 shekels per month plus 2,990 shekels setup, with 4,000 messages per month. The full comparison is on the pricing page.

The timeline matters as much as the price. Setup on Standard is up to 14 business days and on Business up to 21, and the count starts from the moment we have received all materials, not from the moment of payment. The setup fee is paid in one payment before work starts, the subscription begins on the day the agent goes live, and every ongoing adjustment after that comes out of the plan's monthly change quota: 5 changes on Standard and 10 on Business. Fixing a fault, for example an agent that stops replying or an order that was not saved, is not counted against the quota and carries no charge.

Frequently asked questions

Can the agent see real stock levels and promise an item is available? Without an integration into your inventory system, no, and it will not pretend otherwise. It records the line and says supply is subject to stock confirmation, because a message that says "in stock" and turns out to be wrong costs more than an order delayed by an hour. Without any development you can load a list of out-of-stock items as an approved document. Live stock sync is custom development, and on Business, API access and webhooks push every approved order to a destination you define.

How do you prevent unit and quantity errors? In three layers. Every SKU has a unit of measure and a pack multiplier, so the agent asks units or cases instead of guessing. A sanity rule against the customer's history produces one verification question when the quantity deviates. And before the order goes in, a structured summary is sent for approval: SKU, description, quantity, unit. All three layers are configured during setup from your own catalogue.

We run different price lists per customer group. Will the agent know which price to show? Only if the price lists were loaded and only if the customer is identified by phone number against the customer record in WhaleBiz CRM, from which the price group follows. Anyone writing from an unrecognised number, or asking for a special price that is not on a price list, does not get a number: the agent collects details, says a quote will be sent for approval and hands off to a human. A price that commits the company comes from a price list you approved or from a person.

Customers send voice notes and photos of an empty shelf. How is that counted? Agent replies are counted, not incoming messages, and a reply to a voice note, a photo or a file counts as one message, the same as a text reply. Standard includes 1,800 messages with overage at 0.38 shekels, Business 4,000 with overage at 0.32 shekels. The files are stored on the order card.

Which plan do you start on and how long does setup take? Standard, 490 shekels per month plus a one-time setup fee of 1,990 shekels, up to 14 business days, because that is where WhatsApp through the official Meta API and file sending live. Start, 190 plus 990 setup, does not fit: no WhatsApp. Several agents, Instagram, in-chat payment or API and webhooks require Business, 990 per month plus 2,990 setup, up to 21 business days. Prices exclude VAT and the subscription begins on go-live day.