Our service desk tool sends an automatic acknowledgement to every email it receives, with a ticket number and a line saying we will be in touch. Two of our largest customers run their own service desks, and several of their teams email ours directly about deliveries and portal access.
On a Friday evening in May, one of those customers changed the address their portal notifications came from to their service desk's main mailbox. At ten past seven a routine notification reached our desk and opened a ticket. Our tool acknowledged it. Their desk treated our acknowledgement as a new request, opened a ticket and acknowledged that. Ours opened another. Every message was polite, correctly formatted and slightly different from the one before, because each carried a new ticket number, so neither system's duplicate check found anything to stop.
By Monday morning our queue held eleven thousand four hundred tickets and theirs held about the same. Our tool, which copies the account contact on every ticket, had emailed their head of logistics around two hundred times. The genuine tickets raised over the weekend, including a depot reporting a broken label printer on Saturday night, were somewhere in the pile. Both service desk managers spent Monday morning deleting.
Each tool had behaved exactly as configured. An automatic reply assumes there is a person at the other end who will read it and stop. Two automatic replies facing each other each assume that about the other, and neither is right.
Our desk now marks every automatic message with the standard header that tells other systems it was generated automatically, and it no longer acknowledges incoming mail that carries the same header or its common equivalents. No sender receives more than five acknowledgements in an hour. A sixth message from the same address in an hour opens one ticket for a person to look at, with the rest attached to it. Mail from the service desk systems of known customers and suppliers goes to a queue that does not acknowledge at all. And a surge of tickets from a single address pages the engineer on call, because the cheapest moment to stop a loop is its first ten minutes.
The loop has tried to start twice since, with a different supplier each time, and stopped at the fifth message.
When we automate a reply, we are writing something that another automation will read. We had only ever pictured a person reading ours.
– Serguey Shinder
Top comments (0)