My AI didn't want to bother me. The customer did
In a conversation on my Aspen website, a visitor wrote:
âThe price of the trip.â
Then:
âCan you contact me by phone?â
And then, just in case there was any doubt:
âCan you call me?â
My email system received the notification, summarised it and gave it a label:
NOISE.
Underneath, it added: âno reply neededâ.
I had managed to automate part of my travel agency. Specifically, the part where someone asks about a trip and we decide thereâs nothing to do.
Progress of a questionable kind.
I have a system that reviews my email and prepares cards to help me decide what needs attention. Customers, suppliers, information, noise. The intention is good: spend less time sorting things and more time dealing with them.
The Aspen enquiry didnât arrive from the visitorâs own inbox. It came from Airtable, the tool that collects notifications from the website. The sender was an automated address, one of those that tells you not to reply.
And there was the trap.
A rule in the software turned messages from automated senders into noise. It was meant to stop the system drafting replies to an address where nobody would read them.
So far, quite sensible.
But âdonât reply to this addressâ had turned into âthis matter doesnât need a replyâ.
Itâs as if the postman brought me a request for a quote and I decided to ignore it because the postman didnât want to go skiing.
The best bit appeared when we read the original card. Its summary said:
âAirtable notification: new lead asking about aprĂšs-ski, dates and prices for a trip to Aspen.â
It had understood.
The enquiry was recognised, summarised and correctly described. Right above it was NOISE. Right below it, âno reply neededâ.
All on the same card. We didnât even have to change screens to find the contradiction.
This looked like a story about an AI that hadnât understood a customer. The evidence shows something more uncomfortable: the system could understand the message and still handle it badly.
The rule surrounding the intelligence had the final say.
And Iâm the one using that system.
At this point I could get dramatic. Say I lost a sale, calculate how much the trip would have cost and write a headline about artificial intelligence ruining my morning.
It would sound great. It would also be made up.
The notification weâve checked contains no phone number or email address for the visitor. The website assistant had asked for a number so we could call, but that information doesnât appear in the conversation we received.
So there was a request for contact. There wasnât, in that notification, a number to call back.
I donât know whether that enquiry ended in a sale, another call or nothing at all. I know my system labelled it as noise. Thatâs enough of a problem to deserve attention, without dressing it up with another one I canât prove.
The software now includes an exception for these Airtable notifications: it marks them as a customer and adds a note. Replying to the automated sender wonât help; you need to look for the contact details inside the message.
If theyâre missing, theyâre missing. No label can invent them.
I havenât needed another artificial intelligence with a more impressive name to understand the failure. Iâve needed to put three fairly ordinary things side by side: what the visitor wrote, what my panel showed and the rule that decided where to put it.
Iâm taking one very specific check away from this. When I test an email system, I wonât just look at whether the summary is well written. Iâll follow an enquiry all the way through to see where it ends up, what action the system suggests and whether I have the details to take it.
Because an inbox can be beautifully organised and still be badly attended to.
And I donât sell organised inboxes.
Iâve been selling trips since the eighties. If someone asks how much it costs and whether I can call them, that isnât noise.
Itâs my job.
What did you think?