If you sell a physical product through a network of dealers, you've probably gotten at least three emails this month telling you to "leverage AI." Most of them are selling something. Here's the less exciting truth: for a business like yours, AI is a great tool for about four jobs and a waste of money for about forty others. The trick is knowing which is which before you spend anything.
We've turned down more AI work than we've taken, usually because the honest answer to "should this be AI" was no. Here's the test we actually use, and where it lands for a physical-product business specifically.
A simple test for whether AI fits
Before anyone writes a line of code, three questions settle most of it:
- Is there a pile of unstructured text, photos, or history nobody has time to read? Reviews, support threads, years of notes — the kind of thing a person could technically read, but won't.
- Is a decision being repeated dozens or hundreds of times, mostly by gut feel? One-off judgment calls don't need a model. The same call made every week does.
- Is a wrong answer cheap and catchable? Not safety-critical, not expensive to undo, reviewed by a person before it matters.
Three yeses, it's worth prototyping. Zero, spend the budget on headcount or a better CRM instead. Most real situations land somewhere in between, which is exactly why this has to get answered project by project — not as a blanket "yes, we need AI" from a slide deck.
Where it earns its keep (physical-product / dealer-network edition)
Four places this pattern shows up again and again for a hardware company that sells through independent dealers instead of direct to consumers.
Deciding who to approach next
We built a lead-scoring tool for exactly this: a hardware company whose growth person was deciding who to call next by gut feel and inbound interest, out of thousands of unscored prospects. The v0 was a transparent, weighted blend of a coverage-gap score (does this prospect fill a real hole in the territory, or sit next to someone we already work with) and a look-alike score (how similar is this prospect to our best existing partners, measured on their own public profile). The AI/ML piece was reading public reviews for language that matched the exact problem the product solves — and it took three failed approaches to get that part right. Comparing review text to hand-written examples miscalibrated. A general-purpose zero-shot classifier produced confident false positives at scale. Only having a language model actually read each review and reason about it, rather than score it by similarity, held up under scrutiny. The output was a ranked list with the reasoning attached to each row — usable by a non-technical salesperson, not a black box.
Reading what customers already say
Reviews, support emails, and warranty notes hold your product roadmap and your earliest defect signals — nobody has time to read all of it by hand, and a summarization or classification pass over that pile is close to the ideal AI use case: high text volume, low risk if a single classification is slightly off, and a human still reads the interesting cases it surfaces.
Triage of repetitive questions
Dealer FAQs and order-status questions repeat constantly and rarely need judgment — a good candidate for AI handling the first pass, with a person still on the other end of anything that doesn't fit the common cases. That's a meaningfully different thing from a chatbot that tries to replace the person entirely, which is usually where this goes wrong.
Forecasting, once you have the history
Demand or inventory forecasting can be genuinely valuable for a seasonal, regional business — but only once there are a few seasons of clean sales data to learn from. Before that, a forecast is just a guess with a confident-looking chart around it.
Where it doesn't (yet)
Just as important as the four jobs above are the places we tell clients not to spend the budget, at least not yet:
- Replacing relationships. Dealers sign on because they trust a person, and no model changes that.
- Safety-critical guidance. Install or load instructions for anything that could hurt someone need human-reviewed content, not freeform generation.
- Fixing a process nobody's written down. If you can't explain how you decide today, AI will just automate the confusion.
- Tiny datasets. With 15 customers, a spreadsheet and a phone call beat a model — there's no pattern at that size for anything to learn.
- "AI-powered" as a marketing label. It doesn't move sales, and it invites skepticism from buyers who've been burned by vague AI claims before.
How to start small
The same four steps work for almost any candidate that passes the test above:
- Pick one decision you make repeatedly.
- Build a transparent v0 — a weighted score you can explain in one sentence, not a black box.
- Show it to the people who'd actually use it.
- Add a trained model only after you have outcome data to learn from.
The best AI project we've shipped is usually the most boring one — a ranked list, a sorted inbox, a flagged outlier — not the thing that looks impressive in a boardroom demo. If you've got a decision you're making on gut feel right now, we're happy to take a look.
Three yeses, it's worth prototyping. Zero, spend the budget somewhere else.