Labs / Field notes · Updated 2026-10-21

An automated caller asked a building-products supplier for a window price

One reported call. The questions that followed were not about the technology. They were about who the caller worked for and whether anything it asked for should ever turn into an order.

By Founder

Published

PlannedDecided, not built.
  1. Information
  2. Identity
  3. Authority

Explore with AI

Field report

01

What was reported

On October 21, 2026, an employee at a building-products supplier answered a call that their phone had flagged as possible spam. They described it afterward as an automated assistant calling on behalf of someone shopping for windows.

According to the employee, the caller asked whether the supplier carried vinyl replacement windows, then vinyl double-hung windows. It asked about one size, about 24¼ by 53½ inches, then the current lead time and the price. It gave no location and did not say who it was calling for, though the employee allowed that they may have missed part of its opening.

The employee answered its questions, then said they could not get to it right then. Nothing in the account says an order, a payment or a name was exchanged.

One employee's recollection is the whole evidence.

02

What the people at the supplier did with it

They did not talk about the technology. They talked about what could go wrong. A window like this takes three to four weeks to arrive, so a card used to place an order could fail well after the order is in. The supplier would need to know where the job site is. Someone asked whether the supplier could require a real phone number tied to the assistant. It cannot: the number comes from whichever service operates the assistant.

Then the question moved from can we answer it to what are we willing to let it do. The proposal was a short statement on the website, much like terms and conditions, saying what the supplier will do and what it will not, and where to go for the rest.

The conversation moved from information to authority within minutes.

03

Information, identity, authority

First came information. The caller wanted to know what the supplier sells, in what material and size, how long it takes and what it costs. Those are good questions, and a business that answers them differently across its phone, its website and its price sheet will have the inconsistent answers repeated, quickly and at volume.

Second came identity. Who is this, and who is it working for? A possible-spam flag was the only check the employee had, and it checks a phone number, not a buyer.

Third came authority. Even if the supplier knew exactly who the caller represented, would it let software place an order? A made-to-order window with a lead time of weeks is not a casual purchase. Most of the effort right now goes into being found and understood. The second and third steps are where a business gets hurt.

Access to information is not authority to transact.

04

What a website can and cannot do about it

A website can state a boundary in plain language and in a form a machine can read: what an automated visitor may ask, which questions get a firm answer and which get call us, that a quote is an estimate and not an offer, and that orders and card payments are taken by a person.

That sets an expectation for honest software and gives a human something to point to. It does not secure anything. A caller that ignores it is still a caller. The controls that do the work sit elsewhere: the person answering knowing what to ask, the payment step refusing what the policy refuses, and a delivery address confirmed by a human who can reach a human. We would not call any of it a protection until it is enforced where an order is actually taken.

A policy on a page does not secure an order flow.

What we asked and what we found

Every entry answers the same eight questions, in this order.

Question

What does one automated sales call tell a business about the difference between answering a question and accepting an order?

Why it matters

An employee at a building-products supplier took a call that their phone had flagged as possible spam. They described an automated assistant calling for someone shopping for windows. We have their account, given in a recorded conversation hours later. We do not have a recording of the call. What it raised is worth writing down because it came in a predictable order: information, then identity, then authority.

State

Planned Decided, not built.

What we built or tested

  • Not built: any test, policy or tool for this. This entry records what was reported and the checks we plan to run.
  • No client system was changed because of this call.

Evidence

Reported by the employee who took the call, October 21, 2026
An apparently automated caller asked whether the supplier carried vinyl replacement windows, then vinyl double-hung windows. It asked about one size, about 24¼ by 53½ inches (recalled from memory), the current lead time and the price. It gave no location.
Reported by the same employee
The caller said it was calling on behalf of someone looking for windows and did not say who. The employee allowed that they may have missed part of its opening. They answered its questions, then said they could not get to it right then.
Caller ID
The phone labeled the call "possible spam." That is a judgment about a phone number, not about a buyer.
Raised by people at the supplier afterward
A window of this kind takes three to four weeks to arrive, so a card used for an order could fail after the order is placed. The supplier would need a confirmed delivery location. Whether the supplier could require a real phone number tied to the assistant was asked; the answer was that the number comes from whichever service operates the assistant.
Proposed in the conversation
State on the website what the supplier will do (describe what it sells and what is available) and what it will not (take card payments or place orders on an assistant's behalf), and send anyone who wants more to call, write or visit.
Hypotheses stated in the conversation, not shown
That the call came from a personal assistant working for a shopper, that the supplier was reached because of its website, and that calls like this will become common within a year.

What we learned

  • The questions arrived in an order: information (what is sold, in which size, how long, at what price), identity (who is this, and who is it working for), then authority (may software place an order at all).
  • Telling a stranger what you carry costs nothing. Taking money from something you cannot identify is a different act with different consequences.
  • A website can state the line in plain language and in a form a machine can read. It cannot enforce the line.
  • The controls that matter sit where an order is taken: the person answering the phone, the payment step, and a delivery address confirmed with someone reachable.

Limitations

  • This is one verbal account of one call, relayed within hours. The call was not recorded.
  • We do not know what placed the call, who sent it, how it found the supplier, or whether anyone intended to buy. We do not know whether a price or lead time was given.
  • The size was recalled from memory and was not described as a frame size, a rough opening or a glass size.
  • The concern about a failed card is a worry about what could happen. Nothing in the account says it did.
  • Nothing here measures how common such calls are. We do not name the supplier or the employee.

What changes next

Ask several current AI assistants to find, compare and price a replacement window from a building-products supplier's public site and compare their answers with what the supplier says on the phone. Draft a plain-language transaction statement and a machine-readable equivalent and review both with the supplier's staff before anything is published. Write a short intake script for unidentified automated callers and try it with the staff who take those calls. Results are published here with the method and limits.
Failed tests are published as results. A state on this page describes today, not plans.

Next in Labs

← All of Labs

Explore with AI

Get Upstream

One email a week: what we asked the AI, what it got wrong, and how we checked.

Our research reaches 17,000+ followers across our channels. Upstream is the weekly version.

Weekly. Unsubscribe anytime. Privacy