Mike HeapMike is an experienced Product Manager who focuses on all the “non-development” areas of My AskAI, from finance and customer success to product design, copywriting, testing and more.
Table of Contents
- What does Sofar Sounds do?
- Which helpdesk does Sofar use?
- How did Sofar train their AI customer service agent?
- When did Sofar decide to turn on 'direct replies' to customers?
- What was the biggest thing Sofar did to improve their AI agent's resolution?
- How does Sofar customize their AI agent setup to work for their business?
- Custom Answers for the cross-cutting questions across artists, hosts and audience
- Guidance: voice, clarifying questions, and escalation rules
- What impact is Sofar's AI customer service agent having now?
- Where do Sofar go from here?
Created time
May 26, 2026 01:35 PM
Title length (<60)
Author
Last optimised
Ecomm?
Tags
Image
Publish date
May 25, 2026
Video
Slug
my-askai-sofar-sounds-case-study-2026
Featured
Featured
Type
Case study
Ready to Publish
Ready to Publish
Sofar Sounds runs ~750 monthly Zendesk tickets through My AskAI: 85% AI CSAT, ~16 hours saved every month, and 26% of tickets resolved by the AI so the small support team can respond faster to the rest.
Sofar Sounds is built around real nights out: three artists, an unconventional venue, an audience told the location only hours before showtime. Every ticket the support team sees is from someone whose plans for the week depend on the answer. An audience member trying to find the right address, a host preparing a living room for forty strangers, an artist about to fly in for a show.
I've been on enough Sofar-style support setups (small team, high-emotion inbox, every ticket about an event that's already in motion) to know what makes this one different. The bar this team set themselves is reply speed and reply quality across that whole window, with no flexibility on either.
Across the ~750 conversations we now handle every month inside Zendesk Tickets: 85% AI CSAT, ~16 hours saved a month, and 26% of tickets resolved directly by the AI on its own. The remaining ~74% still need a human today, and honestly, that's the right setup at this stage.
Either way, the AI's job for now is to take the recurring easy questions cleanly off the queue. That lets the small team get to the harder three quarters of the inbox faster.
Here's how it came together.
What does Sofar Sounds do?
Sofar Sounds organizes intimate live music gigs in small and unconventional venues around the world. Founded in 2009 in London by Rafe Offer, the company now operates in over 300 cities, with tens of thousands of gigs hosted since launch.
The model has three audiences (and this matters more for the support story than it sounds; keep an eye on it as we walk through how the inbox actually gets handled). Artists apply via the website and play in groups of three per show, sharing equal billing.
Audience members join via the email list, buy or are invited to a show, and learn the exact location close to showtime, with a house etiquette of arriving on time, staying to the end and listening attentively without phones.
That three-way split is what makes Sofar's support inbox unusual. A ticket from a host losing access to a venue forty-eight hours before a show has nothing in common with a ticket from an audience member who can't find their confirmation email or an artist asking about the payout for last weekend's gig.
The support team is smaller than the events team, and every conversation it touches is about a show that is already in motion. (Fun fact: the venue is usually shared with audience members 36 hours before the gig starts, so the support window for "I can't find the address" is genuinely measured in minutes.)
Which helpdesk does Sofar use?
Sofar runs support inside Zendesk: Support / Tickets, one inbox carrying the three audiences together. Email is the dominant channel; the team handles the conversation, attaches the right city, the right show, the right artist or host, and moves on. Our Zendesk Tickets integration is the seam we sit in.
A meaningful share of the tickets in that inbox still needs a human today. The User Data API isn't connected to Sofar's ticketing backend, so the AI can't see a customer's actual order, the show they have a ticket to, or the host's upcoming gig calendar.
Actions and Tools aren't set up either; both would let our agent process a refund, transfer a ticket, or update a host's gig details directly. Until those pieces are in place, anything that needs a real lookup into Sofar's systems is a human ticket by definition.
That leaves a clean job for our agent in the meantime. More than a quarter of the inbox is made up of questions that can be answered quickly and easily from the help center and the public website: what the etiquette is at a show, how the artist application process works, where to find country-specific information.
By having the AI take those tickets off the queue, Sofar's team has effectively added support capacity beyond what their small team alone could cover. The other three quarters, the ones that still need a human, get a faster reply as a result.
How did Sofar train their AI customer service agent?
Sofar already had a working knowledge base inside Zendesk and a public website that explained the model in detail. The training plan we walked them through was mostly about wiring those into the agent and adding the layer of curation the inbox actually needed.
- The Zendesk help center was the first source connected via our Knowledge connector, capturing the canonical answers Sofar's team had been linking to for years: ticketing, refunds and transfers, host policies, artist guidance, country-by-country city pages.
- The public website was added via website sync, pulling the how-Sofar-works pages, host info, artist info and the city FAQs into the same agent. That's the marketing-side context the help center doesn't always carry.
- A large library of Custom Answers was built on top. These are defined knowledge snippets, written by Sofar's team and embedded directly into our agent, that the AI reaches for when the recurring cross-cutting questions across the three audiences come up in shapes the help center doesn't fully match. More on this in the customization section below.
- Self-Learning was turned on across the lot. Every time a human agent closes a ticket the AI didn't fully resolve, our platform compares the AI's draft reply to the human's actual reply, identifies the difference and writes a new knowledge article from that delta. In the last 30 days, ~500 ticket responses leaned on a Self-Learning-drafted article. On a base of ~750 tickets, that's around two thirds of the inbox running on either a directly resolved AI reply or an escalation message built from one.
A few things were deliberately left out. Sofar's team didn't lean on historic-ticket training, didn't switch on AI Tagging, and (as flagged above) there is no User Data API connection to the ticketing backend.
The setup was kept narrow on purpose. The goal was a triage layer that could be trusted, with the broader automation stack saved for later phases.
When did Sofar decide to turn on 'direct replies' to customers?
Sofar ran our agent in Internal Notes mode for two weeks before flipping the switch to direct replies. Those two weeks let the team pressure-test the Handover & Escalation Guidance rules against a representative sample of the real inbox before any customer saw an AI reply.
(For what it's worth, this is the pattern I recommend to most teams whose tickets are time-sensitive. A two-week notes-mode shadow is enough to surface the rules that need reshaping before a single customer sees an AI reply.)
Inside those two weeks the team watched a single question: when our agent proposed a reply, would a human send it? Where the answer was a confident yes (a public-policy question, a refund window stated unambiguously in the help center, a city-page lookup) the rule that produced that draft was promoted.
Where the answer was anything other than a confident yes (a host-side venue question, an artist payout query, an audience ticket transfer with even a hint of ambiguity) the rule was reshaped to escalate rather than answer.
The decision rule the team ended up with is the one that's still running today. If the AI's drafted reply is correct and safe (no personally identifying details required, no high-stakes sender, no within-forty-eight-hours-of-a-show urgency), the AI fires it directly. Anything else gets a clean handover to a human, with the AI's working out attached as context so the agent doesn't start from a cold inbox.
What was the biggest thing Sofar did to improve their AI agent's resolution?
Here's the catch: the biggest move Sofar made was to tune our Handover & Escalation Guidance so the AI escalates ~74% of tickets to a human by design. Pushing resolution higher would have been the wrong objective at this stage.
Two mechanisms are doing the work, together:
The first is the Handover & Escalation Guidance itself. We define Guidance as a set of natural-language rules (kept short, under 75 words each) that control how the AI behaves: tone, terminology, scenarios it's allowed to answer, and the conditions under which it must hand a conversation to a human.
Sofar's team wrote those rules heavily on the handover side. Anything from a host within 48 hours of a show, anything from an artist about a payout, any audience ticket-transfer request with even partial ambiguity, anything where the AI's confidence on the draft reply sits below the team's bar: handover.
When our agent escalates, the conversation is summarized inside the same Zendesk ticket and control passes to the human, who picks up with the AI's working already laid out for them. It's an escalation that doesn't waste the AI's work; it just stops short of letting the AI claim a resolution it shouldn't (which is exactly the failure mode we built handover guidance to prevent).
The second mechanism is Self-Learning quietly raising the floor on the 26% of tickets our agent does resolve. Every escalation is a free training pass. The human agent's reply, against the AI's drafted reply, becomes the input to a new knowledge article.
The next time a similar ticket arrives, our agent is a little better; the rules don't have to relax for the resolution number to compound. On a ~750-ticket monthly base, ~195 tickets land in the "AI resolves it" bucket and ~555 land in the "AI escalates with full context" bucket; both buckets are fed by the same Self-Learning loop.
I'll show the math so you can pressure-test it: 16 hours saved a month works out at roughly five minutes a ticket on the AI-resolved share (195 tickets × 5 minutes = ~16.25 hours). The savings come from the resolved bucket; the escalation bucket's value shows up in the CSAT number above instead.
How does Sofar customize their AI agent setup to work for their business?
Two customizations carry the weight: the first is a large library of Custom Answers - the knowledge snippets the AI draws on when answering recurring questions across the three audiences. The second is Guidance: the natural-language rules that shape how the AI behaves, including how it speaks to each audience, when it asks a clarifying question, and when it escalates straight to a human instead of trying to answer.
Custom Answers for the cross-cutting questions across artists, hosts and audience
Custom Answers in our platform are reusable knowledge snippets, written by the team and embedded directly into the agent. Their job is to give the AI the canonical text to reach for when a particular question comes up. Routing rules and behavior-shaping guidance sit one layer over the top.
The point is consistency: the reply is on-brand and reuses Sofar's exact wording rather than improvising from the help center.
Sofar's library is large rather than exhaustive on purpose (a pattern we see across most of our well-run customer rollouts; fewer, sharper Custom Answers beat a long tail of single-use ones). It covers the recurring questions that span the three audiences in shapes the help center doesn't fully match: the ticketing mechanics, the way the artist application and equal-billing model works, what to expect when you host a show, the show etiquette guests are signed up to.
Each new pattern the team spots (whether Self-Learning surfaces it from the agents' replies, or it comes out of a manual review) turns into either a new Custom Answer or a tightening of an existing one. The team treats the library as a living document.
Guidance: voice, clarifying questions, and escalation rules
Where Custom Answers are the what of the AI's replies, Guidance is the how. We split Guidance into three types: Communication & Style, Context & Clarification, and Handover & Escalation. Sofar's team writes rules in each of them, kept short (under 75 words per entry) and in plain language.
Communication & Style Guidance sets the voice the AI uses, and Sofar shifts that register depending on who the customer is. With artists the rules push the AI closer to a peer tone, treating them as creative collaborators. With audience members the tone leans warmer and more service-focused, especially when the ticket is about a show that is happening soon (we've watched the same pattern across our hospitality customers: high-emotion inboxes can't afford a flat "Hi there" opener).
With hosts the rules tighten up again. Hosts are operational partners, and a missed nuance can mean a real-world venue problem.
Context & Clarification Guidance controls when the AI asks a question back instead of answering. Before replying about a ticket, our agent is told to confirm city and date; the same question carries different answers depending on the use-case.
Before answering anything about hosting, the AI confirms whether the question is about a past, current or upcoming show. These rules don't add length to the AI's replies; they prevent the wrong answer being delivered with confidence.
Handover & Escalation Guidance is the routing layer covered in detail in the previous section. The rules hand a conversation to a human whenever a ticket falls outside the "correct and safe" decision rule. It's the part of Guidance doing the largest share of the work on Sofar's setup today, and (in our experience across high-emotion inboxes) it's what keeps the resolution number scoped to the share of the inbox our agent can be trusted with.
What impact is Sofar's AI customer service agent having now?
- 85% AI CSAT score across the AI-handled tickets, the headline number for a triage-led setup.
- 26% AI resolution rate, scoped to the share of the inbox the AI can be trusted with today. Handover & Escalation Guidance routes the rest to humans with full context.
- ~750 tickets handled per month: ~195 resolved directly by the AI, ~555 escalated to a human with AI-prepared context.
- ~16 hours saved per month, measured against the AI-resolved share at roughly five minutes a ticket.
All of this on sub-1k tickets a month, the band where "every ticket is a real night out" is the bar, and where the trade-off between AI resolution and CSAT is the one Sofar's team is most willing to defend.
Where do Sofar go from here?
The team is exploring more Custom Answers for the seasonal peaks (festival season in the summer, Christmas and New Year ticket transfer surges, the artist-touring ramps that bunch in Q1 and Q3) and continuing to widen the Self-Learning corpus as the inbox closes more tickets.
Tasks for the recurring high-frequency workflows (audience-side ticket transfers, host-side gig postponements, artist-side payout queries) are on the radar as the natural next step once the data layer is in place.
A useful reframe sits underneath all of that: "100% automation" isn't the goal here, and chasing it would undo the CSAT number the team is most protective of. A realistic ceiling for Sofar once the data layer and the actions are in place probably sits somewhere between 60% and 80% AI resolution while CSAT stays above 80%.
That's high enough that the AI is doing serious work, capped well short of the point where it would start answering tickets it shouldn't. The next twelve months are about widening the slice our agent can be trusted with, while keeping the discipline that makes the inbox feel like a real conversation.
If you'd like to see how the per-ticket economics compare against Zendesk's own AI, our Zendesk Tickets ROI calculator is a useful starting point, and our pricing is on a single page if you want to model the cost at Sofar's volume.
If you'd like to see more customer stories like Sofar's, browse all our case studies here.
