The Day a Robot Rang Our Robot: Two AIs Negotiated Over SSDs After Closing Time

One Monday evening an AI shopping agent rang our business to compare prices on solid-state drives, and our AI receptionist answered it. Twice, from two different numbers, four minutes apart. If you run a small business with a phone, this post is worth ten minutes, because the interesting part is not the novelty of two machines talking. It is what happened to every other business the agent rang that evening, which was almost certainly nothing at all. The part that unsettled us most is that the written transcript gives nothing away. It reads like an ordinary customer. We only knew because somebody sat and listened to the audio, which is why both recordings are on this page for you to judge yourself. By the end you will know what the law does and does not require either side to disclose, why detection is harder than it sounds, and what being invisible to a shopping agent actually costs. We run an AI receptionist on our own line before selling it to anybody, so this is our own call log rather than a vendor’s anecdote.

What actually happened on our phone line?

An automated caller rang twice on Monday 3 August, at 17:26 and 17:30, asking whether we had two specific NVMe drives in stock and what we charged for them. Nobody was free to answer it. Our AI receptionist, Paula, picked up both times, and we found the transcripts the next morning the way we find every call: waiting in the log, already summarised.

Half past five on a Monday is worth pausing on, because it is not the dead of night. It is the exact moment a lot of independent shops have just locked the door, cashed up or started driving home. The enquiry did not arrive at an unreasonable hour. It arrived at the most ordinary one imaginable, and that is precisely why so many phones would have rung out.

The first call, 5:26pm

The caller opened without preamble and asked for a specific part, by full model number, first time, correctly:

Caller: “I’m looking for two one-terabyte Western Digital WD Blue SN5100 NVMe SSDs. Do you have those in stock?”

That is a real product, incidentally. The WD Blue SN5100 datasheet describes an M.2 2280 PCIe 4.0 drive rated up to 7,100 MB/s sequential read. Whoever or whatever was calling had the exact part in mind, not a vague description of one.

The second call, four minutes later

At 5:30pm it rang back from a different number and ran the same script again. Same product, same three questions, whole phrases delivered identically rather than rephrased. On paper that looks like somebody trying again. In the audio it is the same delivery twice over, which is a different thing entirely, and it is the detail that gave the game away.

What it was almost certainly doing

Our best reading is that somebody had asked an AI assistant to source two specific drives at a target price, and the assistant was working through a list of Manchester computer businesses by telephone, asking each the same three questions: do you carry it, what does it cost, and is there an equivalent at a similar price. When it could not get a price from us, it thanked Paula and left. No pressure, no negotiation, no attempt to be sold something else.

How did we know the caller was an AI?

By listening to it. And that is the part worth dwelling on, because if we had only read the transcript we would almost certainly have filed both calls as an ordinary trade enquiry and never thought about them again.

Credit where it is due: it was a good conversation

Let us start with the thing that surprised us most. The caller was good. Not stilted, not obviously synthetic, not the press-one-for-sales experience anyone over thirty is braced for. It asked a clear question, listened to the answer, responded to what it had been told and closed the call politely. As a piece of conversational engineering it was genuinely impressive, and we would rather say so than pretend we saw through it in the first three seconds.

That matters more than a compliment. If the technology has reached the point where a business owner cannot casually tell, then “we will just know” has stopped being a plan.

The transcript alone would not have told us

This is the thing we did not expect. Read the written record of either call and you get something that looks entirely normal: a caller wants a specific part, asks whether we have it, asks the price, asks about an equivalent, thanks us and rings off. Written down, that is not suspicious. It is an efficient customer.

Every “tell” you could list from the text has an innocent explanation. No small talk? Some people do not do small talk. Exact model number? They had done their homework. Same questions twice? They rang back because the first call did not get them anywhere.

We only knew because one of us sat and actually listened to the audio.

The repetition is identical, not merely similar

When a person rings back and repeats themselves, the words might be close but the delivery never is. The stress lands somewhere different, the pace changes, they trail off or add “sorry, me again”. Read as text, that repetition looks the same as software repeating itself, which is exactly why the transcript hides it.

Heard as audio, the two are nothing alike. What you notice is not that the caller said the same thing, it is that the saying was the same. That distinction does not survive being written down, and it is the single reason we spotted this at all.

It asked exactly three questions and nothing else

Stock, price, equivalent. No deviation, no follow-ups of its own invention, no curiosity about anything outside the task. It did not ask where we are, whether we deliver, how long we have been trading, or whether the price was negotiable. Told we are not a parts retailer, it did none of the ordinary human things, which range from “oh, sorry” to “well, who would have them then?”. It took the answer and left.

On its own, in text, that reads as a person in a hurry. Alongside everything else, and once you have heard it, it reads as a routine filling in fields.

Do not take our word for it, listen

This is precisely why both recordings are on this page rather than just the transcripts. We are not asking you to accept our reading of a conversation you cannot hear.

Play call one, then play call two, and pay attention to the repeated lines in particular. If you would have caught it cold, on a Monday evening, with one ear on the telly, you are sharper than we were. Most people we have played it to were not sure until the second call.

What would make us wrong about this

We should be straight about the limits. We cannot see who operated the caller, we have not traced the numbers, and everything above is inference from two phone calls. It could in principle have been a person working a rigid call list.

What makes us confident is the combination: the exact model number, a second call from a different number four minutes later, an identical delivery rather than a merely similar one, and the complete absence of conversational drift. Any one of those alone would prove nothing. Together they look like software. But we would rather you listened and formed your own view than took ours.

The uncomfortable business consequence

If the transcript does not reveal it, then neither does your CRM. Most businesses never hear their own calls. They see a note, a summary, a line in a system, and that record is exactly the layer where the evidence disappears.

So the practical answer to “how would I know” is, for most firms, that you would not. Which is an argument for caring less about detection and more about being answerable and honest to whoever, or whatever, rings.

What did our AI do when it met one of its own?

It told the truth and refused to guess, which sounds like a low bar until you consider what the alternative would have caused. Here is the core of it:

Paula: “I’m sorry, but we don’t stock or sell NVMe drives here. We’re a repair and upgrade service rather than a parts retailer. If you’d like us to quote for fitting one into a laptop or desktop for you, I can take the details on that.”

It said what we actually are

We are not a parts shop and we are not a walk-in shop. We are an office-based repair and upgrade business that collects and returns. A receptionist that had fudged this to keep the caller happy would have started a chain of disappointment ending with a real person expecting stock we have never held.

It refused to invent a price

This is the failure we most wanted to avoid, and it is the one most likely to happen with a poorly configured system. Everything we do is quoted by a technician after seeing the machine, because the price of an upgrade depends on what is inside it. Paula held that line through three increasingly direct attempts to extract a number.

Consider what a guessed price would have done here. With a person on the phone, an optimistic quote is embarrassing. With an agent on the phone, that invented number goes straight into a comparison table, gets relayed to a customer as fact, and we get chosen on a price we never agreed to. The agent has no mechanism for detecting that we were being vague to be helpful.

It offered the two things we do want

An upgrade quote, or an email address that a human reads. Both were declined, which is the correct outcome. The call took a minute, cost us nothing and misled nobody.

Where it fell short

On the second call she could have offered the email route sooner rather than restating the quoting policy, and her sign-off collided with the caller hanging up. Both are on the training list. We publish the imperfect version because a case study where the system performed flawlessly is usually a case study with the awkward parts removed.

Why did neither AI have to say what it was?

This is the genuinely surprising part, and the timing is almost too neat. The calls happened on 3 August 2026, the day after Article 50 of the EU AI Act became applicable.

What Article 50 requires

Article 50 is the transparency provision. The European Commission’s guidance on transparency obligations states that providers of AI systems intended to interact directly with people must ensure those people are informed they are interacting with an AI system, unless it is obvious, and that the disclosure happens from the start of the first interaction in a clear and distinguishable manner. The text of Article 50 itself sets out the same duty. It became applicable on 2 August 2026 with no grace period.

The machine-to-machine gap

Here is the wrinkle. The obligation is about AI systems interacting with natural persons. The Commission’s guidance is explicit that systems operating through machine-to-machine communication, without direct contact with people, fall outside the scope of that duty.

So on our line, on the day the rule took effect, an AI rang a business and reached another AI. If no human was ever part of the conversation, the disclosure duty arguably never engaged for either side. The rule written to stop people being fooled by machines does not obviously cover machines talking to each other. That is not a loophole anyone designed. It is what happens when a technology arrives in a shape the drafting did not anticipate.

And in the UK there is no equivalent duty at all

Worth stating plainly for a UK audience: the EU AI Act is EU law. Britain has no equivalent statutory transparency requirement, and the UK approach has so far been to apply existing regulators to AI rather than pass a single AI statute. We set the timeline out in our plain-English guide to AI regulation. Practically, that means a UK business deploying a voice agent today chooses for itself whether to disclose.

Why ours discloses anyway

Because being found out is worse than being upfront, and because a caller who works it out mid-conversation feels tricked in a way that costs more trust than the automation saves. Paula says what she is in her first sentence, every call, and mentions that the call is recorded. We wrote about why that transparency turns out to be commercially useful rather than costly in most callers enjoy talking to our AI. The joke of this particular evening is that ours disclosed to a caller with no feelings to hurt, one day into a regime designed to require exactly that, while the caller disclosed nothing to us.

Was the call even legal?

Almost certainly yes, because it was not marketing, and that distinction carries more weight in UK law than people expect.

The rule on automated calls

PECR regulation 19 prohibits automated calling systems being used for direct marketing without the recipient’s prior specific consent. The ICO’s guide to PECR covers electronic and telephone marketing, and the regulator’s position is that AI voice systems count as automated calls. There is no conversational-AI exemption that makes a synthetic voice into a live human agent.

Why a buying enquiry is different

That rule is about marketing to you. A machine ringing to ask whether you will sell it something is not promoting goods or services at us, it is a purchasing enquiry, and PECR’s marketing rules are not obviously engaged. We are not lawyers and this is not advice, but the distinction is real and it is the reason we treat these calls as ordinary business enquiries rather than as nuisance.

What changes if this scales

The interesting question is volume. One agent ringing four shops is a customer doing their homework efficiently. Ten thousand agents ringing every shop in a city, repeatedly, is a load problem, and Ofcom’s rules around nuisance and silent calls exist precisely because automated dialling at scale has previously become a menace. Nothing about those two calls was abusive. It is not hard to imagine a version that would be.

What is agentic commerce, in plain English?

It is what happens when people stop searching for things themselves and delegate the searching to software that can act. We covered the general shift from software that answers to software that does in our guide to AI agents. This was that shift arriving on a telephone in Manchester.

From searching to delegating

This is not a fringe behaviour any more. Ofcom’s Adults’ Media Use and Attitudes research found that just over half of UK adults, 54 per cent, now use AI tools such as ChatGPT, Copilot or Gemini, up from 31 per cent the year before. Once that many people have an assistant to hand, some of them will start asking it to do the ringing round.

The old pattern was a person opening ten tabs. The new pattern is a person stating an outcome, “find me two of these for about this money”, and something else opening the tabs, and now making the calls. The person never sees the businesses that failed to respond. They see a shortlist.

The phone is the last unindexed database

Everything a search engine can read about your business has been crawled for years. What has never been machine-readable is the thing only your phone knows: whether you actually have it today, what you would actually charge, whether you could fit them in on Tuesday. An agent that can dial reaches information no crawler has ever had access to. That is why this is happening on the phone rather than only on the web.

What an agent can find out about you

Only what something at your end will tell it. If nobody answers, the agent learns nothing and infers nothing generous. If a voicemail answers, it learns you are closed. If a capable system answers, it learns exactly what you sell, what you do not, and how to take it further. Our call ended in a “no” and that is still a better outcome than silence, because a clear no is information, and information is what keeps you in the dataset for next time.

Who lost money that evening?

Every business it rang that did not answer, and none of them will ever know. That is the part worth sitting with, because a missed call from a person leaves traces and a missed call from an agent does not.

A missed call and being invisible are different things

When a person cannot reach you, some of them ring back, some leave a message, some try your website, some remember you next week. The relationship survives the failure. When an agent cannot reach you, it records no answer and moves down the list. There is no callback, no message, no second attempt in a week, and no lingering intent to be recovered. You are simply not in the comparison.

The maths is not complicated

We have written before about what unanswered calls cost in ordinary trading terms in missed calls cost more than you think. Agentic enquiries make the arithmetic worse in one specific way: they are perfectly efficient at removing you. A tired human ringing round at half five, catching shop after shop that has just shut, might give up after three and try again tomorrow. Software works the whole list, without irritation, and completes the job that evening.

The bottom row is the one to worry about. Being chosen on an invented number is worse than not being chosen, because you now have a customer with a false expectation and an agent that has recorded you as cheaper than you are.

What should a small business actually do about this?

Four things, none of which require you to believe any particular prediction about how fast this grows.

Answer the phone, including when you are not there

This is the whole game and everything else is detail. The businesses that appear in an agent’s shortlist are the ones something answers for at twenty-five past five, after the door is locked, accurately. That is what an AI receptionist is for, and it is the same reason it earns its keep with human callers who would otherwise have rung a competitor.

Publish the facts a machine will ask for

The three questions our caller asked were about what you sell, what it costs and what is equivalent. If the answers to those exist in a form a machine can read, on your own site, you are easier to include. Prices are the uncomfortable one. We publish ours, which is unusual in this trade, and it is one of the reasons we can be compared at all.

Decide in advance what your system may say

This is the governance question and it matters more as callers get more automated. Write down what the machine may state as fact, what it must escalate, and what it must never do. If you have no idea where to start with that, it is exactly what an AI opportunity assessment is for. Ours may describe what we do, take details and book a callback. It may not price work, promise a timescale or discuss an existing customer’s job. That boundary is why the exchange above was safe.

Do not lie to the robots either

There is a temptation to think that fooling an agent is harmless because there is no person on the line to deceive. There is a person, one step back, and they will turn up expecting whatever the agent was told. The honesty rule does not soften just because the immediate listener is software.

Where we might be wrong about all this

Two calls on one evening is a thin evidence base, and since the written record gives nothing away, our reading rests on what we heard rather than on anything we can print. We would rather flag that than have you find it.

We cannot prove who was calling

We have strong behavioural evidence and no identification. If it turns out to have been a very consistent human sourcing agent, the specific claim in the title is wrong, though almost everything else in this post still holds.

It might not have been shopping at all

An alternative reading is a data-gathering exercise: something building a price and stock dataset for Manchester, rather than buying on behalf of a named customer. That would be less charming and, for the businesses that did not answer, exactly as consequential.

Two calls is not a trend

We are not going to dress this up as a wave. It is the first time we have clearly seen it on our own line. What makes it worth writing about is not the volume, it is that the volume can only go one way, and that the response to it is something you would want to do anyway.

It is also worth remembering how differently people feel about AI in the abstract. The government’s public attitudes to data and AI tracker found 96 per cent of UK adults aware of AI, yet “scary” and “worried” among the words most associated with it. The customer whose agent rang us may well be uneasy about AI in general while cheerfully delegating their shopping to one. Both things are true at once, which is the whole story of this technology so far, and we tell the longer version of it in the story of AI.

Hear it for yourself

Both recordings are above, unedited apart from the file format. We have published them because the written transcript genuinely does not give it away, so asking you to trust our reading of a conversation you cannot hear would be asking too much. Listen to the repeated lines across the two calls in particular. That is the single clearest sign that you are hearing software rather than a person having another go, and it is the one thing that does not survive being written down. If you would rather test the system than read about it, ring 0161 660 0356 and talk to Paula directly. She will tell you what she is in her first sentence, take whatever you actually need, and have it in front of a human by morning.

If you run a business whose phone rang out that Monday evening, that is the real point of this post. Not that two machines had a polite conversation about solid-state drives, but that ours was in the running and yours was not asked twice. Have a look at what we run and what it costs, or ask us a question and a person will answer it. We are Manchester based, and there is a local version of all of this if you would rather deal with somebody down the road.