Nasik Abbas

Case study · PadelOS · Summer 2026

Answers in seconds, not hours

PadelGPT is the AI assistant inside PadelOS, the software padel clubs use to run bookings, payments and their player app. Over the summer of 2026 we measured it against the clubs' own support staff, on the same kinds of questions.

The numbers

  • Speed

    9.7 secondsvs4.8 hours

    Median time to a first answer, from PadelGPT and from club staff, across 15 clubs in nine countries.

    Check the source

  • Missed

    9.1%

    of the 1,457 messages players sent to club staff never got a reply.

    Check the source

  • Quality

    28 of 40

    questions where a reviewer, not told which reply was which, rated PadelGPT's answer as good as the club's, or better.

    Limit: One club, 40 questions, first replies only, in English and Dutch. With a sample this size, the true share is likely between 55% and 82%.

    Check the source

  • Capacity

    31–44%

    of staff messages PadelGPT could have fully answered today, from the club's own information.

    Limit: An estimate. Every request was judged twice, by two different methods; the range runs from where both agree to the higher of the two. A hand check of 28 requests agreed with each method on 22.

    Check the source

The problem

Players message their club about bookings, prices, open matches, lessons and refunds. Staff answer when they can, around the rest of their work. Over the summer, 48% of those messages waited more than four hours for a first reply, and 9.1% were never answered at all.

Players also write when the club is busy or closed: 40% of PadelGPT conversations started outside office hours.

Player · Saturday 13:20Hi, can I move my booking to 17:00?

Club staff · Monday 09:22Hi, our apologies for only noticing your question now. I hope you were still able to play!

A request to move a booking, first answered after 44 hours.

Player · Saturday 22:49[Location], tomorrow 9.30

PadelGPT · 17 seconds laterAvailable at [location] tomorrow at 9:30: a double court, 60 or 90 minutes for £24, and a single court for £10. Which court and duration would you like to book?

A late-night availability question, answered from live data in 17 seconds.

What we did

We read everything: all 1,457 messages players sent to club staff between June and August, and every PadelGPT conversation since it launched on 1 July.

Then we asked one question of every staff message: could PadelGPT have answered this? That is a judgement, and a judgement can flatter. So every message was judged twice, by two different methods, and we report the range between them rather than the better number. A person then checked a sample by hand.

Quality was judged by a person, blind: a reviewer compared PadelGPT's first reply with the club's real reply to the same questions, without knowing which was which.

The hard part was not the AI. It was being fair: measuring the AI and the people on the same work, and reporting the less flattering number when two methods disagreed.

What we found

  • Speed. PadelGPT gave a first answer in a median of 9.7 seconds, and nine in ten arrived within 23 seconds, at any hour. Staff took 4.8 hours.
  • Quality. PadelGPT's answer was as good as the club's, or better, in 28 of 401 questions.
  • Reach. Players wrote in eight languages, and 29% of conversations were not in English. PadelGPT answers in the player's own language.
  • Use. 461 players held more than 660 conversations with it in its first two months, rising from 224 in July to 442 in August. At clubs where both channels were live, 36% of all contacts already went to PadelGPT.
  • Capacity. 31–44%2 of the messages staff received could have been fully answered by PadelGPT as it is today. With the planned admin assistant, which prepares an action such as moving a booking for a staff member to approve with one tap, that rises to 92–95%3.

What it is worth

The estimated saving is about 43 hours4 of staff time a month across these clubs, and 13,863 hours5 of player waiting over the three months.

What happens next

  • Clubs fill in the few pieces of information players ask about most.
  • PadelGPT answers first, and passes anything that needs a person to staff with the details already attached.
  • An admin assistant prepares staff-only actions, and a staff member approves each one before it happens.
  • Decisions that need judgement stay with people. In 8.2% of the messages staff answered, they made an exception or offered goodwill, and that is theirs to give.

My role

PadelGPT is the product I lead, and this study was mine to run. PadelOS published it in full, with every chart and its method, at padelos.com.

For how the AI itself was evaluated — replaying real conversations safely and judging replies blind — see the technical case study.

About the numbers

Every number on this page comes from the report PadelOS published. Figures marked with a note are estimates or small samples, and the note says how far to trust them.

  1. 28 of 40: One club, 40 questions, first replies only, in English and Dutch. With a sample this size, the true share is likely between 55% and 82%.
  2. 31–44%: An estimate. Every request was judged twice, by two different methods; the range runs from where both agree to the higher of the two. A hand check of 28 requests agreed with each method on 22.
  3. 92–95%: A projection for an assistant that is planned, not built. Every admin action would wait for a staff member to confirm it; the rest stay with staff.
  4. about 43 hours: An estimate on mid assumptions of five minutes per staff reply. Across the low and high assumptions it is 26 to 69 hours a month.
  5. 13,863 hours: An estimate, at the top of the range: it counts the waiting on requests that the more generous of the two methods judged fully answerable.

Public source: