Rolling Out AI in a Wholesale Business — Where to Start (and in What Order)

Wholesale is, from a distance, a business of repeated decisions. The same eighty customers order from the same two hundred SKUs, on the same discount terms, in roughly the same rhythm, week after week. That repetition is exactly what makes it a good place to put AI to work — and exactly why so many wholesalers get the first project wrong.

The mistake isn't picking a bad idea. It's picking a good idea in the wrong order.

In wholesale, the sequence matters more than the shortlist

Most AI shortlists in a distribution company look similar: demand forecasting, dynamic pricing, a chatbot for customers, maybe route optimisation. All four are legitimate. None of them should be first, because all four depend on data that a phone-and-email order process doesn't produce.

That's the whole argument of this post. Your first AI project isn't the one with the biggest headline number — it's the one that produces the clean data the next three need. Get that wrong and you spend a year building models on top of a manual process, then discover the model was learning your data entry habits rather than your market.

Count the repeated decisions before you look at any tool

Spend a week counting, not brainstorming. For each process, write down two numbers: how many times a day it happens, and how many minutes it takes each time. That's it — the shortlist writes itself from the product of those two columns.

I have used this on enough wholesalers to know what usually comes back at the top. In the alcohol wholesaler I documented as a case study, order intake ran at 120 orders per day at roughly 15 minutes each. That is 30 person-hours every single day — three full-time people, doing one thing, all day. No forecasting model on the planet was going to out-earn fixing that.

The second column matters as much as the first. A decision that happens twice a day and takes an hour is a worse target than one that happens two hundred times and takes four minutes, even though the second one feels trivial. Volume is what an agent is good at; judgment calls that happen twice a day are what your people are good at.

Step one: get orders in as structured data

The first project in almost every wholesale business is the same — move order intake off phone, email and WhatsApp, and into something machine-readable. Not because ordering is glamorous, but because everything downstream reads from it.

There are two ways to do it and they aren't exclusive. A B2B self-service portal moves the typing to the customer, who is the one person who actually knows what they want. An intake agent handles the customers who will never use a portal — it reads the email or the WhatsApp message, resolves "the usual 20 cartons" against that customer's history and price list, and creates the order in the ERP as a draft for someone to confirm.

Most wholesalers need both, and the split is predictable: the portal takes the customers who order often and know the catalogue; the agent takes the long tail who order sporadically and describe products in their own words. In the case study, the portal reached 70% of orders within the first month and processing time fell from 15 minutes to 2 — the remaining 30% is precisely where an intake agent earns its keep.

Step two: the money side

Once orders arrive as data, the next-largest pile of repeated decisions is financial. Payments have to be matched to invoices, credit limits checked before release, counterparties verified before you pay them.

Bank reconciliation is the strongest second project in most distribution businesses for the same reason it is in an accounting office: high volume, a clean definition of success, and it runs safely in draft mode while you calibrate. A wholesaler with several hundred active customers has a statement that nobody enjoys, every single day.

Credit control is the underrated one. An agent that watches receivables, flags customers approaching their limit before the next order ships, and drafts the reminder email is a small build — and in a business running on 30-day terms with thin margins, it is often worth more per hour of development than anything in the warehouse.

Step three: now forecasting and pricing make sense

With two quarters of clean, time-stamped order data behind you, forecasting stops being a research project and becomes an ordinary one. You know what was ordered, when, by whom, at what price, and — crucially — what was asked for and not fulfilled, which is the signal a manual process throws away entirely.

The same applies to pricing. Discount structures in Polish wholesale are rarely documented anywhere except in a sales director's head and a spreadsheet with fourteen tabs. Once every order flows through one path, those rules become visible, and the conversation about changing them can be based on what customers actually do rather than what everyone remembers.

Chatbots I'd still leave to the end, and only for a specific job — answering "where is my order" and "what's my price on this SKU" — which is worth doing precisely because those two questions eat the sales team's day. As a first project it fails for the reason it always fails: it answers questions instead of processing work.

The ERP question to settle before anything else

Every project above writes to or reads from your ERP, so the integration path decides your timeline. This is the single most useful thing to establish in week one.

The spread is real and worth knowing in advance: enova365, Optima and Subiekt each expose a completely different surface, from a documented REST API at one end to a Windows service and a read-only database at the other. None of them are blockers. What does block projects is the access itself — a hosting arrangement where nothing external can reach the ERP, or a vendor relationship where nobody will grant credentials. Sort that out first; it's the item that quietly turns a three-week build into a three-month one.

Build the ERP coupling into one isolated connector module, whatever the system. When your vendor ships a version bump, you want to revalidate one component, not the whole agent.

Rules that keep a wholesaler out of trouble

Three constraints, all learned the expensive way, all worth writing into the scope before development starts.

The ERP stays the source of truth. Stock, prices and customer terms live in one place, and the agent reads them there rather than keeping a copy. A second copy of the price list is a pricing error waiting for a busy Friday.

Drafts before autonomy. New orders, new bookings, new emails — everything the agent produces starts as a proposal a human confirms, with a record of what it saw and why it decided as it did. You relax that per process, after the numbers earn it, not on day one across the board.

One process at a time. The temptation to launch intake, reconciliation and credit control together is strong because they touch the same data. Resist it: each one needs its own calibration period, and running three at once means you can't tell which one is producing the errors.

A realistic first quarter

Weeks 1–2 — count and choose. Measure volume × minutes across your processes, pick the top one, confirm ERP access, and agree what "working" will mean numerically. Not "faster" — a percentage, measured on your data.

Weeks 3–6 — build and shadow. The first process gets built and then runs alongside your team without changing anyone's job. The agent proposes; people work as usual; at the end you compare. The number that comes out of this is the only honest basis for every later decision, and it will be lower than any vendor's benchmark, including mine.

Weeks 7–12 — switch on, then widen. The team starts from drafts and an exception queue instead of a blank screen. You measure the same number again, plus time spent, and only then start the second process — with the first one still running rather than paused for the rebuild.

One quarter, one process, one number you trust. That's a slower story than the shortlist you started with, and it's the reason the second project takes half the time of the first.

---

I build these agents for Polish wholesalers and distributors — order intake from email and portal, reconciliation and credit control on the money side, wired into enova365, Comarch Optima, Subiekt or whatever you run. If you want to know which of your processes would automate first and at what percentage, get in touch — the scoping conversation is free and takes half an hour.