Q.Where Does AI Actually Help a Distributor Today?
In four places, all of them boring: reading orders that arrive as email and PDF attachments, cleaning up item and catalog data, drafting the repetitive replies and quotes, and summarizing what happened yesterday. Everything else you are being pitched is a demo.
The test is simple. Does it remove a step a person currently does by hand, every day, without anyone having to change how they work? If the answer involves your team learning a new interface, adopting a new process, or "exploring use cases," it is not ready and neither are you.
It is also worth knowing that the "distribution is behind on AI" line you keep reading is out of date. In the US Census Bureau's Business Trends and Outlook Survey — a fortnightly survey of roughly 1.2 million businesses — the share of manufacturers using AI in a business function reached 23.4% for the reference period 10–23 August 2026, which is marginally above the 23.2% all-sector rate. Wholesale trade reached 20.4%, up from 15.8% four months earlier. (Census changed the question wording in late 2025, so those figures do not join onto older ones as a trend line.)
What follows is the short list that survives that test in a wholesale or light-manufacturing operation, and the one rule that has to sit underneath all of it.
1. Reading Orders That Arrive as Email, PDF and Photos
Order intake is the highest-value place to put AI in a distribution business, because it is the one task where a human is doing pure transcription under time pressure.
Your customers do not all order the same way. Some use the portal. Some email a PDF purchase order. Some send a photo of a handwritten sheet from the back of a restaurant. Some reply to last week's order with "same again, plus two cases of the 16oz." Someone on your team reads each of these and types it into the system.
That typing is the step AI removes. A model reads the attachment, matches each line to your item file, and produces a draft order with the customer, the items, the quantities, the unit of measure, and anything it could not match flagged for a person.
What makes this work in practice, and what most pilots get wrong:
- Match against your real item file, not a generic catalog. The hard part is not reading "2 cs 16oz" — it is knowing which of your four 16oz SKUs that customer means. That requires their order history, not a language model on its own.
- Unit-of-measure confusion is where the money is lost. Cases versus eaches, and "a dozen" meaning twelve units or twelve cases depending on the customer. Every unmatched unit has to stop and ask.
- Leave the original attached. When a claim comes back three weeks later, the person handling it needs the PDF the customer actually sent, next to the order that was created from it.
- Measure the exception rate, not the accuracy rate. You do not care that it got 94% of lines right. You care how many orders came through with zero lines needing a human — that is the number that turns into hours.
2. Item and Catalog Data Nobody Has Time to Fix
The second place is the one no vendor demos, because it is not exciting: your item master is a mess, and AI is genuinely good at the kind of tedious normalization that has sat on someone's list for three years.
Descriptions entered by six different people over eleven years. The same manufacturer spelled four ways. Pack sizes buried inside the description field instead of in a pack-size field. Missing categories, so nothing can be reported on. Duplicate items that split the sales history of one product across two SKUs.
None of this is hard work. It is just an enormous amount of small work, which is exactly what a model does cheaply and a person does slowly and resentfully. Run it as a one-time clean-up with a person approving the merges, then keep it running on new items so the file does not decay again.
This is unglamorous and it is usually the change that makes everything downstream — reporting, the ordering portal, the price list, the website — suddenly work.
Want this looked at for your operation?
Twenty minutes on a call, and you'll know whether this is worth building, buying, or leaving alone. No deck, no pitch.
Book twenty minutes3. The Drafts: Quotes, Replies, and the Same Email Fourteen Times a Day
The third place is drafting. Not sending — drafting.
A quote follow-up. An out-of-stock notification with the substitute suggested. A reply to "can you do better on that price." The weekly email to the customers who have not ordered in six weeks. Each of these is written from scratch by a person who has written it hundreds of times before.
The right implementation puts the draft inside the screen where the work already happens — in the order, in the customer record, in the quote — with the customer's history already in it, and a send button a person presses. Not a separate chat window someone has to remember to open. The moment your team has to leave their workflow to get the benefit, they stop.
4. The Daily Read
The fourth place is the summary nobody has time to write: what changed since yesterday.
Orders that came in below their usual size. Customers who normally order Tuesdays and did not. Items that went short against orders already on the books. A margin that moved on a line and nobody noticed. This is all sitting in your data and none of it gets read, because reading it means running five reports before 8am.
A short written summary, in the same place every morning, is a modest-sounding feature that changes what a general manager notices in a week.
The Rule That Has To Sit Under All of It
Anything an AI agent does that changes stock, a price, or an order stops as a draft that a named person approves — and the approval keeps the name and the date.
This is not a technicality. It is the difference between software your team trusts and software they quietly work around.
The failure mode with agents in an operations business is not that the model is wrong. It is that the model is wrong once, at 4pm on a Friday, in a way nobody sees until the truck is loaded — and after that, nobody trusts any of it again. You spend the next year with a very expensive tool that everyone double-checks by hand.
Draft-and-approve solves this in both directions. The work still gets removed, because reviewing a correct draft takes seconds and typing the order takes minutes. And when something is wrong, it is caught by the person who would have been doing the task anyway, before it reaches inventory or a customer. The audit trail — who approved what, when — is what lets you actually expand the automation later, because you can look back and see the exception rate falling.
| Task | What AI does | What the person still does |
|---|---|---|
| Emailed and PDF orders | Reads the attachment, matches lines to your item file, drafts the order | Approves the draft; resolves flagged lines |
| Item and catalog data | Normalizes descriptions, proposes merges and categories | Approves merges; sets the rules once |
| Quotes and repetitive replies | Writes the draft with history already in it | Edits if needed; presses send |
| The daily read | Writes the summary of what changed | Decides what to do about it |
| Anything touching stock, price or an order | Proposes | Approves, by name, on the record |
Be Honest About the Size of the Gain
Two pieces of evidence are worth holding together before you build a business case.
The Federal Reserve Bank of Atlanta surveyed around 750 corporate executives in late 2025 — median firm revenue $46 million, which is this audience exactly — and separated what firms believed AI had done for them from what could be measured. The finding, in the paper's own words: "We document a productivity paradox, in which perceived productivity gains are larger than measured productivity gains, likely reflecting a delay in revenue realizations." Executives reported a mean labor-productivity gain of 1.8% for 2025. The measured figure for the bucket that includes wholesale trade was "roughly 0.4 percent."
Separately, the Distribution Strategy Group surveyed 233 people in December 2025 and found "63% of distributors remain in 'exploring' or 'piloting' stages, with only 4% having achieved full integration," and — the line worth pinning above your desk — "Eighteen percent of respondents have no formal ROI tracking for AI initiatives."
The honest reading is not that AI does nothing. It is that the gains are real, smaller than the pitch, slower to show up in the numbers than in the mood, and invisible to anyone who did not measure the before. Which is why the project below starts by counting two things.
And one more piece of honesty: there is no rigorous published study measuring what AI does to distribution's actual operating metrics — order processing time, inventory accuracy, fill rate. Everything in that space is vendor case-study marketing with no stated sample or method. If someone quotes you a percentage improvement in order processing, ask where it came from. The answer is usually another vendor's blog.
What This Should Not Be
A few things are worth ruling out plainly, because they are what gets sold.
A chatbot on your website. Your customers are placing repeat orders, not asking questions. A chat widget answers a problem you do not have.
A separate AI product with its own login. If it lives outside the system where the work happens, adoption is near zero after week three. This is the single most common reason an AI pilot quietly dies.
Demand forecasting, as a first project. It is the thing everyone asks for and the thing most likely to fail first, because it needs clean history, and your history is not clean yet. Do the item data first. Forecasting gets much easier afterward, and sometimes stops being necessary.
Replacing a person. In every operation we have looked at, the realistic outcome is that the same team stops doing transcription and starts doing the work that was being deferred — chasing the open quotes, calling the customers who went quiet, fixing the pricing exceptions. That is a better argument than headcount anyway, and it is one your team will actually cooperate with.
Where to Start if You Are Going to Start
Pick the single highest-volume intake path — usually emailed purchase orders from your top twenty customers — and automate only that, end to end, with approval. Run it for a month alongside the manual process. Count two numbers: how many orders came through with no line needing a human, and how many minutes the approval took compared to typing.
If those two numbers are good, you have a real result you can extend to the next intake path. If they are not, you have spent a month and learned something true, instead of spending a year on a platform.
FAQ
Q: Do we need to replace our ERP to use AI in our operation? A: No, and you generally should not try. The order intake, item data and drafting work described here sits alongside whatever system you run today and writes into it. Replacing an ERP to get AI is how a six-week project becomes an eighteen-month one.
Q: Are distributors and manufacturers behind on AI? A: Not any more. In the US Census Bureau's Business Trends and Outlook Survey for the reference period 10–23 August 2026, 23.4% of manufacturers reported using AI in a business function — slightly above the 23.2% all-sector rate — and wholesale trade reached 20.4%, up from 15.8% four months earlier. Most articles still citing a large lag are working from late-2025 data.
Q: How much productivity does AI actually add? A: Less than the pitch and more than nothing. The Atlanta Fed's 2026 study of around 750 executives found executives reported a mean 1.8% labor-productivity gain for 2025, while the measured gain for the group that includes wholesale trade was "roughly 0.4 percent" — what the authors call a productivity paradox. Treat any vendor percentage for order processing or inventory accuracy with suspicion: no rigorous study of those specific metrics exists.
Q: How accurate is AI at reading purchase orders? A: Accuracy on clean typed PDFs is high; accuracy on photos of handwritten sheets is not. The number that matters is not accuracy but exception rate — the share of orders that come through needing no human correction at all. Measure that on your own documents before committing, because it depends far more on your item file and customer history than on the model.
Q: What is the biggest risk with AI agents in a distribution business? A: An agent silently changing stock, a price or an order. The mitigation is structural, not technical: every action that touches those three things stops as a draft that a named person approves, and the approval is recorded with the name and the date.
Q: Should a small distributor build this or buy it? A: Buy anything that is genuinely standard, and build only where your operation is genuinely different — which, for most distributors, is the order intake and the item matching, because those depend on your customers and your catalog. If a product does the job at a monthly price, use the product.
Read us often? You can add Futurise as a preferred source in your own Google account.