A VP I know spent three days last spring going through 900 customer accounts by hand, one at a time, trying to get a read on how those customers felt about his company. Three days of a senior person's life. And the answer he came back with was already a month stale.
Everything he wanted was sitting in his own business. His reps had been told all of it, out loud, in shops and back offices, over hundreds of visits. It just never landed anywhere he could reach.
So he did what everybody does now. He pasted the question into a chat tool and got back something fluent, confident, and useless, the kind of thing that reads like a McKinsey deck about an industry it has never set foot in. Then he decided the hype was ahead of the product.
I run AI training for distributors, manufacturers, and rep agencies most months now, and that moment is the single most common thing I have to un-teach. The output was bad. The model was fine. Nobody briefed it.
Think of it like a new intern
This is how I explain it in every room.
You hire an intern. Sharp kid, reads fast, works weekends, never complains. Day one you walk over and ask whether you should push harder in Denver. No account history, no sales data, no idea who your WDs are, no clue what happened on the last three visits out there. They'd give you an answer, because interns always give you an answer, and it would be a confident, well-written guess.
Nobody would blame the intern for that. You'd say you didn't set them up.
The model is the intern. The prompt is the question you asked. And the context is everything the model can see when it answers, which in most people's case is nothing at all. We spent two years teaching everybody to ask better questions. Asking was never the part that was broken.
Why more context makes it worse
The field caught up to this in the last year and gave it a name. Anthropic calls it context engineering: curating what goes into the model's window at each step, as opposed to prompt engineering, which is just writing a sharper instruction. Their framing is that a model has a finite attention budget, closer to human working memory than to a hard drive.
That part surprises people, and it's the part worth sitting with, because the instinct once you understand context is to shovel everything in.
Chroma's context rot study ran 18 leading models and found performance drops off as inputs get longer, not gradually but in cliffs. On a long chat-history benchmark the models did worse with the full history in front of them than with only the relevant excerpts. Drew Breunig has a good breakdown of how long contexts fail, with four named failure modes: poisoning, distraction, confusion, and clash. Confusion is the one I see most in our world, where some irrelevant thing that happened to be sitting in the pile gets picked up and used to build a worse answer.
So dumping your CRM export into a chat window usually makes things worse, not better. The work is sorting: the right material, in a shape the model can use, at the moment it's needed.
Which, fine, is a trommel. We put it in the company name for a reason and we do like to torture the metaphor.
What good context looks like in the aftermarket
Generic advice about this stops at "give the model relevant information," which helps nobody. Four things make context relevant when the question is about your field team.
The vocabulary. Walk into a shop and ask who they run for friction and every person in the building knows you're asking about brakes. A general-purpose model reads friction as conflict. Multiply that across jobber, WD, blitz, ride-along, counter guy, program group, and private label, and you have a system quietly mistranslating half of what your reps say. A right answer and a fluent wrong one look identical on the screen.
The relationships. An installer buys through three different WDs. A rep agency carries ten lines into one distributor. Your field guy visited that shop last week riding with a completely different WD than the one who owns the account. That's a many-to-many reality, and most systems flatten it to a single owner field because that's what the template allowed. Wrong structure going in, wrong answer coming out, no matter how good the model is.
The recency and the source. A note from a visit nine days ago and a CRM field somebody last touched in 2019 are not the same quality of evidence. Strip that distinction out and the model weighs them the same, then hands you a recommendation built on a stale record.
The lens. People miss this one entirely. One shop visit means four different things to four different people. Marketing wants voice of customer, category management wants what's moving and what's dying, pricing wants where you're getting beat. A single generic summary serves none of them.
And the honest limit, which is the reason this is real work rather than a clever prompt: you can't point a chat tool at a big database and expect an accurate analysis. It will make things up, confidently. Even done properly we don't get 100% of the entities right, usually more like 80%. The fix is getting the intelligence captured at the conversation, tagged, tied to the right account, and structured so a model can read it before anybody asks a question.
Why not just do this in Salesforce
I got a fair question on a call this week, and I get it a lot. How many customers does Tromml have? About fifteen. Salesforce has more than 150,000. So why wouldn't you just do all of this in Salesforce?
I love that question, because the answer is the whole thesis.
CRMs track what happened. We capture what was said. And a CRM ships as a template you customize, which means every piece of context above is work somebody has to do. Do it inside a horizontal platform and you're hiring a consultant, opening a project, and building the aftermarket's data model yourself, on your budget.
The platform won't meet you halfway, because it can't. A small slice of their customer base is in this industry. Nobody there is building a way to pull a rep agency's activity into your system, or to move intelligence between a distributor and a supplier, because the other 99% of their customers don't have that problem. Same reason the general-purpose agents bolted onto general-purpose platforms keep landing flat.
So we don't compete with Salesforce. We complete it. Salesforce stays your system of record. We're the system of action on top, and what we bring is the context underneath the answer.
That's also the unglamorous reason our vendor reports work. One of our customers took about 250 notes across a single trade show, and out of those notes we pulled roughly $9M in pipeline. No model trick involved. The notes got captured in the moment, tied to the right account, nested under the right distributor, tagged, and stamped with when they happened. We built the context before anyone asked the question, so the question had something to land on.
It's also why our training spends more time on this than on prompts. A prompt library gets you through Tuesday. Knowing what your model can and can't see changes how you work for good. All of the material we teach from is free on our resources page, whether or not you ever talk to us.
Try it yourself in fifteen minutes
Take a real question you'd like answered. Which accounts in this region should I visit next month, and why.
Ask it cold. Read what comes back.
Then ask again, but paste in your last ten visit notes for that region, twelve months of sales for those accounts, and two lines on what you're deciding and what's worrying you. Read that one.
The distance between those two answers is your context gap, and once somebody sees it they stop asking me which model is best.
Honestly, the way I measure a win here hasn't changed in three years. Did it save the person time. Did it show them what needed attention. Did it build intelligence that compounds instead of dying in the parking lot. Context is the input to all three, and this industry's biggest competitor is still waste.
If you want to know what's already sitting in your own data, that's what a free Signal Analysis is for. Send us a slice of your existing calls, notes, or CRM data and we'll show you what can be pulled out of it. Start small, see if it's useful, go from there.