What 36,000 customer contacts say we should build next.

Five problems that are ours to fix show up in every month of the year.

All figures come from Zendesk, our customer support system, unless labelled otherwise: they count contacts, not bookings or app users. Merged duplicates, spam and test tickets are removed. October runs to the 27th. No customer is identifiable on this page.

Upgrade volume follows railway disruption, not the product.

Seven in ten upgrade contacts are the refund form for a cancelled or downgraded train. Take those out and upgrades generate about 480 contacts a month.

• Rail ticket contacts halved over the year, from 719 to around 400 a month. Most are change requests through the self-serve form.
• Superfare holds at 250–300 a month. The exception is September (472), caused by the Microsoft email outage. It has the most fixable mix of any product.
• Secret Fare and Train Swap are under 1% each. Too small to identify themes from support data alone.

*October is 1–27 Oct. Products come from the product field, filled in from tag families where it's blank. Rail change-request forms count as rail tickets.

Upgrades are a disruption story. Superfare is about inflexibility and delivery.

From the automatic reading of each customer's first message, which exists for Feb–Mar and Jul–Oct 2026 only. The 12-month themes below use agent tags, which cover the whole year.

Booking mistakes are the largest problem that is ours to fix, and they're growing.

Each small chart shows Nov 2025 → Oct 2026. Counted from agent tags, one contact per theme. Tags under-count, because agents don't tag everything they see.

Booking mistakes, from the agent's refund reason

Hand-triaged refunds only, so it covers only refunds that needed a human decision.

• Wrong booking. Wrong date overtook wrong time in September. CS's reading of the app is that the date is easy to miss in four places: the search results, the add-a-trip popup, the auctions screen and the place-a-bid screen.
• Standard Premium confusion never went away. Customers who learned that a Seatfrog upgrade means First Class have stopped checking.
• The refund form itself generates contacts. About 80 a month are filled in with the wrong transaction ID or details.
• Superfare: email is the only way a customer receives their ticket. There is no in-app copy or other route. Even outside the September outage, 20–50 people a month say their ticket or confirmation never arrived.
• Missing receipts and past trips ran at about 30 a month until May, then dropped to almost zero. We haven't checked why; it could be a fix or a tag change.
• Buy Upgrade Now problems spike rather than run steadily (March and May). The last-seat display bug was fixed on 21 September.

Some of these have been open for over a year.

Status read from Linear on 28 October 2026.

Pain points the team sees over and over

• Customers run two or three auctions at once, and the wrong trip sits next to the right one.
• "You lost the auction" notifications don't make clear which trip they refer to.
• Two-seat bids where the customers are seated apart, or told late that only one seat exists.
• Customers can't find how to contact us, and ask for a phone number.
• Refer-a-friend discounts not arriving.

CS observation, not measured.

"Booked the wrong thing, then couldn't fix it" is the loudest product complaint.

2,318 rated comments over the year: 377 unhappy, 1,941 happy. Every unhappy comment was read and coded by hand.

A comment can carry more than one theme. Counts are hand codes, so allow ±2–3. Service handling (speed, tone, a decision the customer disagreed with) appears in 64% of unhappy comments. 28 of the 81 "disagreed with the decision" comments are the fare rules themselves, delivered by an agent. 39 unhappy customers objected to being sent to the train operator.

Satisfaction ratings are binary (good/bad). Quotes are verbatim, trimmed with "…", and anonymised: names, emails, phone numbers and booking references removed.

Human first reply has held at about 21 hours all year.

• Contacts the AI touched are rated lower in most months. This isn't a fair score for the AI: it's given the hardest conversations (refunds and date changes, which often end in a "no"), and most of them end with a human.
• Few customers complain about the AI itself. Of 79 unhappy comments where the AI opened the conversation, at most 5 criticise the AI. 15 happy customers said the AI's first reply didn't help and a human then fixed it.

First reply time is Zendesk's reply_time_in_minutes, which counts agent replies only. 15–22% of contacts have no reply recorded (mostly forms that never needed one), so treat these as directional. "Resolved by the AI" uses Zendesk's automated-resolution tag. The September monthly review quoted 9% from a narrower field-based measure; both point the same way.

Upgrade refunds per sale peaked in June and have eased since.

• Why upgrades were refunded. 14,726 refund-form requests in 12 months. From June, when the form started recording the reason, about two thirds were cancelled trains and one third downgrades from First on the day (Jun–Oct: 4,640 cancelled, 2,150 downgraded). Both are railway or operator events.
• Superfare and Secret Fare: about 80–120 refunds processed a month (refund service logs, May 2026 onward only). 439 Superfare change or refund requests weren't eligible under the fare rules.
• Rail tickets: about 1,100 refund contacts and 380 voids in the year, through the refund partner's form.

Sales and refunds come from the booking database (sale table), read 28 Oct 2026, not Zendesk. Before June a single legacy option covered both refund reasons, so cancellations and downgrades can't be split for earlier months.

• Window: 1 Nov 2025 – 27 Oct 2026, UK time, by ticket creation date. 36,167 contacts after removing spam, merged duplicates and test tickets. That's about 5% below the published monthly reviews, which included merged duplicates.
• Product attribution: the product field is set on only 24–37% of tickets, so products were filled in from tag families. 10% stays unattributable.
• Tags under-count, and some were created during the year: Standard Premium tags start in 2026 (combined with the 2025 experiment tag), AI tags from February 2026. The intent classifier has gaps in Nov–Jan and Apr–Jun.
• Satisfaction percentages come from ratings data and aren't interchangeable with the Explore dashboard score (Jul 88.4%, Aug 84.5%, Sep 85.0%).
• Quotes are anonymised: names, emails, phone numbers and booking references removed.
• This page reports problems and numbers. It doesn't propose fixes.