This is the fourth post in a series working through one industry at a time. This week it’s B2B distributors: wholesale and supply businesses that sell to other businesses rather than to consumers.
I spent thirty-one years in a steel service centre, which is itself a distributor in most respects. It buys, stocks, and sells material to other businesses, with all the account management, inventory, and logistics that implies. On top of that, I’ve built the prospecting pipeline and the retrieval systems that sit behind finding and serving accounts. So while every business is different, this one is closer to home than most.
My standard caveat applies. Neither I (nor anyone else!) can tell you from the outside which of these fits your business or in what order. What I can do is point at where my experience says AI earns its keep in a business that runs on a book of accounts.
The Work That Eats the Week
A distributor runs on two assets, its book of accounts and the inventory on its shelves, and that changes where the money leaks. There’s no equivalent here of the field service company’s missed call, that dramatic moment when a job walks out the door. The losses are quieter: a prospect list that never gets worked because the reps are busy quoting, or a steady account that tapers off and only gets noticed at quarter end. An order sits in a shared inbox waiting for someone to key it in, and a line of stock quietly goes stale while the cash inside it sits still.
Ask where the inside hours actually go and most of them land on the work around selling, not the selling itself. Your inside team spends its day building and working the prospect list, keying in orders that arrive as emailed purchase orders, PDFs, and BOM spreadsheets, turning each request into a quote against tiered and account-specific pricing, and answering the same stock, lead time, and substitution questions over and over. Most of that is coordination and data handling rather than judgement.
What I’ve Already Covered
Two of the best places to point AI in a distributor are patterns I’ve covered earlier in this series, and they apply here almost directly. Orders and quotes that arrive as purchase orders, PDFs, and spreadsheets can be read, mapped to your catalogue, and drafted against your own pricing logic for a person to confirm and release. That’s the same logic-engine build I described for steel RFQs and professional services proposals, only now it is reading a distributor’s paperwork and pricing rules. And retrieval over your own documents, the pattern from “When AI Reads Your Documents” (in this case the product catalogue, spec sheets, cross-reference guides, and past order notes) gives a counter or inside sales rep the right spec or equivalent with a citation instead of a walk over to the one person who knows the line. That’s exactly what Docora does, and if either of those use cases fits, the earlier posts cover them in depth.
That leaves the three areas specific to how a distributor makes and loses money, and they deserve the rest of this post: finding accounts, keeping them, and managing the stock your cash is tied up in.
Use Cases
Finding and qualifying the accounts worth pursuing. This is where I can speak from a build I actually ran. A pipeline searches for companies matching a profile, enriches them from public sources, qualifies them against the criteria that matter to you, and then drafts a first pass of outreach for a person to review and send. The reason it runs affordably at scale is that almost none of the work needs a frontier model. The heavy filtering happens on small, inexpensive models, and the capable model is saved for the last step, where the wording actually matters. I walked through those step-by-step model choices in “Fitting the Right Model to the Task”. The Coles Notes version: built this way, the pipeline costs hundreds of times less than pointing a top-tier model at every record. That difference is what makes running it regularly, rather than as an occasional campaign, a realistic proposition.
For a Canadian distributor there’s a catch that belongs in the same breath. Outbound prospecting runs into CASL, which unlike the American rule has no blanket exemption for business-to-business email. A pipeline that scrapes contact lists and fires off messages without threading in consent, sender identification, a physical mailing address, and a working unsubscribe isn’t saving you time, it’s building you a liability. The practical version of this build bakes the rules in: message only contacts whose information is conspicuously published and relevant to their role, carry the required identifiers, honour unsubscribes inside the required window, and keep the audit trail. This is a real advantage of a considered build over a generic scraper, since the scraper will happily help you break the law faster. Of course, you should always confirm the current CRTC guidance and your own counsel before any campaign since the implied consent rules may be tightening.
Keeping the accounts you already have. The counterpart to finding new accounts is spotting the ones drifting away, and here I want to be careful, because it’s easy to oversell. Knowing which accounts have gone quiet doesn’t need AI at all. Your order history is structured, transactional data, and a plain query against it will tell you who has slowed down. Most distributors already have the query; what they’re missing is the frequency. Today the drift gets caught whenever a rep finds time to pore over their accounts (which in a busy month is never) and so the taper runs on for a quarter before anyone sees it. An automated job changes that: it runs every week without being asked, watches for the pattern, and alerts the rep while the account is merely drifting rather than gone. Where AI is worth the effort is the harder next step: reading the unstructured material around those transactions, such as the rep’s notes, the email threads, and the quotes that never converted, to suggest why an account drifted and to draft an approach for the rep to consider. The detection is just a scheduled job. The interpretation is unstructured work, and that’s the part AI is actually good at. Either way, this is only useful wired into your real order data, which makes this a build rather than a subscription.
Managing the stock your cash is tied up in. This is the use case I lived with longest. In a steel service centre the inventory is a commodity, its value moves with the market, and material bought at the wrong time or held too long shows up directly on the bottom line. Most distributors carry some version of that risk. Run too lean and the order goes to whoever has it on the shelf, and eventually the account follows it. Carry too much and the cash is tied up in stock that may be losing value while it sits.
The first job here is predictive ordering. A system that reads your own sales history, seasonality, and the demand patterns of your actual accounts can propose what to reorder, in what quantity, and when. For commodity lines it can go a step further: a job that watches the benchmarks and indexes your buyers already follow and flags when the trend says to order ahead of a rise, or to trim inventory aggressively ahead of a fall. The second job is catching stock going stale, flagging the lines that are slowing down before they become dead stock while there’s still time to move them. That might mean offering them to the customers who actually buy that line, repricing, or (when possible) returning them to the vendor. That’s the same pattern as the inventory agent I described in the steel post, which watches for off-cuts and offers them to suitable customers before they become scrap. Some of this is forecasting arithmetic that predates AI, and the same honesty applies as with the drifting accounts: if a report answers the question, use the report. AI’s real job here is reading the unstructured signals around the numbers, such as market commentary, price announcements, and supplier lead-time chatter, and putting a recommendation in front of your buyer regularly, with the reasoning attached, without anyone having to remember to ask.
Where AI Is the Wrong Tool Here
At this point, I still recommend keeping a person involved on every send. I covered the legal reason above, but even without CASL, outbound messaging that nobody reviews will eventually say something to a prospect that a rep would never have said. In a relationship business, that can easily cost more than the time it saved.
The same restraint applies to the drifting-account signal. It should prompt a rep’s judgement, never trigger automatic discounting, because the reasons a good account slows down are often relationship reasons a model can’t see. Maybe the buyer changed or there was a service issue nobody logged. Or a competitor simply earned the business. A rep can find out. A discount fired at the wrong problem just trains the account to wait for discounts.
Nothing on the inventory side should place orders on its own either. A call on where the market is heading is a judgement about risk, and getting it wrong ties up cash for months or sells off stock you were about to need. The system proposes, with its reasoning visible, and your buyer decides — by design.
Order and quote automation deserves a similar caution. Pointed at a messy catalogue or inconsistent pricing rules, it produces wrong orders and wrong quotes faster than a person ever could, and in this business a wrong substitution shows up as returned material and a damaged relationship rather than a typo.
And then there’s the thing I say for every industry. If your ERP data and your pricing rules are a mess, AI on top of them will hand you clean-looking wrong answers with great confidence. Fix the data and the process first.
Where to Start
I can’t tell you from the outside which of these fits your business, or in what order, and the honest answer usually depends on what your prospect list, your order history, and your inventory turns actually look like. That’s what a Discovery Day is for: one day, in your business, working out the one or two places this would earn its keep and saying plainly where it wouldn’t, including where the honest answer is to fix the data first.