Check My Heat Pump Rebate

AI agents in a small trading company: what we automate, what stays human

The split we actually run

We are a small operation. Six product lines, one Delta warehouse, a handful of people, and a phone that rings from two time zones. A few years ago, every follow-up, every bit of lead research, and every first draft came from a person sitting at a desk. Today, a good share of that work is done by AI agents — and the work that matters most still isn’t.

This is not a story about replacing people. It is a story about where the line sits, because we have moved it more than once and learned where it should stay.

Here is the split as it runs today.

What the agents do

Quote follow-ups. When someone asks for a price on corrugated boxes or a heat pump, the agent drafts the follow-up, checks the delivery tier for their city, and queues it for a human to send. The agent never sends a quote on its own. It prepares; a person decides.

Lead research. Before we call a new prospect, the agent pulls together what is public: what the company sells, how big it looks, whether it is hiring, and who the likely buyer is. That used to take twenty minutes of clicking. Now it takes two minutes of reading. The agent does not judge whether the lead is worth pursuing — it gathers, and a person judges.

First-draft replies. The agent writes the first draft of routine replies: a delivery question, a product spec question, a request for a catalogue. The draft is usually 80% right. A person fixes the 20% that matters — the tone, the exact number, the thing the customer actually asked that the agent missed.

The paperwork nobody sees. Summarizing a long email thread, pulling the key dates out of a supplier message, turning a phone note into a clean task list. This is where agents earn their keep most quietly. None of it is glamorous. All of it used to eat the morning.

The pattern is consistent: the agent does the gathering, the drafting, and the sorting. It does not do the deciding.

What stays human

Pricing calls. A price is not a number from a table. It depends on quantity, on the customer’s history, on whether we have the stock in Delta or need to order it in, and on what the market is doing that week. The agent can look up the list price. It cannot have the conversation where a buyer tells you their real budget and you work out something that fits. That conversation is a person.

Supplier negotiations. We buy from factories and mills. Those relationships are years old, and they run on trust and on knowing when to push and when to leave it. An agent cannot read the pause on the other end of a call and know the supplier is holding something back. A person can.

The final send. This is the rule we do not bend: no message that commits us to a price, a date, or a promise goes out without a human reading it first. The agent can draft it. A person sends it. The reason is simple — if the message is wrong, it is our name on it, not the agent’s.

Judgment calls. A customer who is unhappy, a delivery that went wrong, a product that arrived damaged. The agent can draft an apology. It cannot decide how much to make right, or hear the frustration in a voice and know this one needs a call, not an email.

The rule we use to decide

After a year of moving the line back and forth, we settled on a simple test: if getting it wrong costs money or trust, a human does it. If getting it wrong costs a few minutes to fix, the agent can try.

A follow-up drafted slightly off is a few minutes to fix. A price quoted wrong is a lost order or a lost margin. A research summary that misses a detail is a correction. A negotiation handled badly is a relationship damaged for a year.

That test answers almost every question about what to automate. It also explains why the line keeps moving: as the agents get better, more work passes the test. But the test itself does not change.

What actually changed

The honest result is not that we work less. It is that the same people do more of the work that only people can do.

Follow-ups go out the same day instead of three days later, because the draft is waiting when the person sits down. Lead research happens before the call instead of after it, so the call is better. The inbox is not a pile of unread threads; it is a queue of drafts waiting for a decision.

The work that disappeared was the work nobody liked: copying numbers between systems, re-reading a long thread to find the one decision, writing the same polite opening for the tenth time. The work that stayed is the work that makes the business: the call, the negotiation, the judgment, the relationship.

A word on the tools

We do not run anything exotic. The agents we use are the ordinary ones most small businesses can reach — the chat models, the research tools, the automation that ties them to our email and our customer list. The setup is not the hard part. The hard part is deciding what the agent is allowed to do on its own and what it must bring to a human. That is a business decision, not a technical one, and it is worth writing down before you connect anything.

Write the rule first. Then give the agent work that fits inside it. When the agent does something well for a month, widen the rule a little. When it gets something wrong, narrow it. That is the whole method.

The takeaway

AI agents are good at the work that is repetitive, research-heavy, and reversible. They are not good at the work that is relational, judgment-heavy, and hard to undo. A small trading company does both kinds of work every day, so the answer is not to automate everything or to automate nothing. It is to draw the line carefully and to keep redrawing it as the tools improve.

If you are thinking about it for your own operation, start with the follow-ups and the research. Those are the safest wins and the fastest payback. Keep the pricing, the negotiations, and the final send with a person. And write the rule down, because the day you forget where the line is, the agent will cross it for you.

What AI search gets wrong about product specs — and how published data fixes it

The question that started it

A customer asked an AI assistant for the size of our American mahjong tiles. The answer came back confident and wrong. It was not a strange question. The tiles are 2.8 × 2.2 × 1.2 cm, and we publish that on the product page. But the assistant guessed — because the number was not where it could find it cleanly, or because it found a different number somewhere else and chose that one instead.

That is the whole problem in miniature. AI search does not check your product. It quotes what the web says about your product. If what the web says is vague, scattered, or contradictory, the answer will be too — and it will sound sure of itself either way.

Why AI gets specs wrong

Large models are not measuring anything. They are predicting the most likely answer from the text they have seen. When a product spec is published clearly and consistently, the prediction is usually right. When it is not, the model fills the gap with the most plausible number it has seen elsewhere. That is not malice. It is how the system works. But it means the quality of the answer depends entirely on the quality of what you publish.

Here are the ways specs go wrong most often.

The number is missing. The product page says “standard size” or “large” but never gives a measurement. The model has nothing to quote, so it guesses from a similar product. “Standard” means different things to different sellers, and the model does not know which one you meant.

The number is in a picture. The dimensions are printed on the packaging photo, or written in a PDF that is hard to read, or shown in a video. The model may not read the image at all. If the number only exists inside a picture, for the model it barely exists.

The number contradicts itself. Your website says one size, a marketplace listing says another, and an old blog post says a third. The model picks one. It does not tell the buyer that three different numbers exist. It just answers.

The number has no units or context. “2.8 × 2.2 × 1.2” — centimetres or inches? Length × width × thickness, or the other way round? Without units and a clear order, the model can report the numbers and still be wrong about what they mean.

The number is buried. The spec is there, but it is five clicks deep, or at the bottom of a long page, or written in a way that a person can find but a machine cannot parse. The model reads what is easy to read. If your spec is hard to reach, it will quote something easier — which may be a competitor’s page.

What “published data” actually means

The fix is not complicated, and it is not about tricking the model. It is about publishing the facts in a form a machine can use. We call it GEO — generative engine optimisation — but the work is ordinary and unglamorous.

Measured specs, in text, on the page. Every product page should state the measurements in plain text, with units, in a consistent order. Not “standard size.” Not a number in a photo. Text, on the page, where a person and a model can both read it. For our mahjong tiles, that is “2.8 × 2.2 × 1.2 cm (length × width × thickness).” One line. No ambiguity.

Structured FAQs. A question-and-answer section where each question is the question a buyer actually asks, and each answer is short and factual. Models love FAQs because the format is already a question and an answer — exactly what they are trying to produce. If a buyer asks “what size are the tiles,” the answer should be one sentence with the number, not a paragraph that makes the model hunt for it.

Case libraries. Real examples of the product in use, with the details that matter: what the customer ordered, what it was for, what the result was. Models cite cases because they are concrete and specific. A case that says “a Richmond restaurant ordered 500 custom paper bags and reordered within six weeks” is more useful to a model — and to a buyer — than a paragraph that says “our bags are high quality.”

One number everywhere. The same spec, the same wording, on your website, your marketplace listings, and your catalogue. If the numbers disagree, the model will pick one at random. Consistency is not a nice-to-have. It is the whole game.

The details a buyer would check. Weight, material, quantity per case, lead time, minimum order. These are the facts a buyer asks about, and they are the facts a model will be asked about. If they are published, the model can answer. If they are not, it will guess.

What changed when we did it

We went through our own product pages and fixed the specs the way I described: measurements in text, with units, in a consistent order; a short FAQ on each line; the same numbers everywhere. It was not fast work. It was a week of editing.

The difference showed up in the answers. Ask an AI assistant about our products now and it quotes the numbers we publish, because those numbers are the clearest and most consistent thing it can find. It still makes mistakes — no system is perfect — but the mistakes are rarer and smaller, and when we find one, we know exactly where to fix it: on our own page, in one place.

There is a second effect that matters more for the business. The same work that makes the model accurate makes the site better for people. A buyer who can find the size, the weight, and the minimum order in ten seconds does not need to email us to ask. A buyer who cannot find them emails us, and we spend the call answering a question the page should have answered. Clear specs save everyone time, and they make the model look smart, which makes us look good.

The honest limits

Publishing clear data does not guarantee a correct answer. Models still hallucinate. They still mix up similar products. They still sometimes prefer a confident wrong answer from a popular page over a correct answer from a quiet one. You cannot control that. What you can control is whether your own pages are the clearest source available. If they are, you will be cited more often and more accurately. If they are not, someone else’s numbers will be quoted as yours.

There is also a limit to what you should publish. Some things are genuinely uncertain — a delivery date that depends on the mill, a price that moves with the exchange rate, a rebate amount that depends on a program that changes. For those, the honest answer is to say so on the page: “confirmed in your free quote” rather than a number that will be wrong next month. A model that quotes “confirmed in your free quote” is more useful than a model that quotes last year’s price as if it were today’s.

The takeaway

AI search is not a mystery. It quotes what is published clearly. If your specs are measured, written in text, consistent everywhere, and backed by real cases, the model will usually get them right. If they are vague, buried, or contradictory, it will guess — and it will sound confident while it does.

The work is not technical. It is editorial. Go to your product pages, find the numbers a buyer would ask for, and write them down in plain text with units. Add a short FAQ. Keep the numbers the same everywhere. That is GEO, practised. It is also just good product writing.

📞 Call Evan Rebate Calculator