Here is a contradiction worth sitting with.
Gumroad now runs support that is 100% autonomous, capped at $5,000 a month, and, per Antiwork's own site, reportedly closing 478 of 479 tickets inside 24 hours with a median resolution time of 23 minutes. By any traditional measure, that is a solved problem.
And yet the same company is paying customers to find the bugs its agent can't. "$25 credit if you report a bug and we fix it," Sahil Lavingia posted. "All 100% autonomous ($5,000 monthly budget cap)."
Read those two facts together and the story most people tell about Gumroad breaks. If the agent were as finished as the resolution stats suggest, you wouldn't need a bounty. The bounty exists precisely because the numbers hide something.
Here's the reframe. The $25 credit is not a growth hack. It's a confession.
The $25 credit is not a growth hack. It's a confession. You only pay strangers to surface defects when your own processes stopped catching them.
You only pay strangers to surface defects when your own processes stopped catching them. Gumroad has externalized its QA loop — it is buying, at $25 a head, the failure signal that a functioning internal knowledge operation is supposed to generate for free.
That's either radical honesty or a symptom, and it's worth being precise about which. In Gumroad's case, run by a tiny remote team that treats its own operations as a public experiment, it reads as honesty. For most companies copying the autonomous-support playbook without copying the transparency, the same mechanic would be a symptom of a loop nobody owns.
Because the uncomfortable thing the bounty admits is this: the labor didn't disappear when the tickets got automated. It moved.
The old job was answering the question. The new job is making sure the answer exists, is correct, is structured so a machine can retrieve it, and gets fixed when a customer proves it wrong. That is not the same work. It's not even the same department.
And this is now the default, not the frontier. Fin's 2026 Customer Service Transformation Report found 82% of senior leaders say their teams invested in AI for customer service over the last twelve months, with 87% planning to invest in 2026.
So the question is no longer whether to automate support. It's whether you've staffed the disciplines automation creates.
The first is knowledge management, and the framing that matters is brutal: an agent is only as good as what you feed it. A lack of information, badly structured articles, or out-of-date documentation all produce confidently wrong answers — the kind that degrade customer trust faster than a slow human ever did.
The failure mode is specific. When product ships a feature and the knowledge base hasn't caught up, the agent can't answer, hands off to a human, and within hours the team is fielding a backlog of questions the agent should have handled — right when volume peaks. Customers start to distrust the agent, and the team loses the capacity it was supposed to gain.
That's the rot. It doesn't announce itself. It compounds quietly, one stale article at a time, while the resolution-rate dashboard still looks fine.
The fix Fin describes is a discipline they call new product introduction, or NPI: a feature isn't "shipped" until the agent can answer questions about it correctly. Someone from support joins the product walkthroughs and the marketing kickoff, tests the feature before launch, retires outdated content, and writes the new content in the customer's language rather than internal jargon.
The detail that gives it away: "If you're relying on your customers to find your content gaps and then racing to fill them, you'll always be one step behind."
That sentence describes exactly what a $25 bug bounty is. It's paying your customers to find your content and product gaps because you've decided being one step behind is cheaper than building the process to stay ahead. For a small shop, maybe it is. Just don't mistake it for the finished state.
The second net-new discipline is conversation design — owning how the agent communicates, not just what it knows. Because if nobody trains the agent on tone, structure, and when to escalate, it sounds like an LLM, "because it is one."
Here is the number that should end any argument that this is soft work. Fin A/B tested two opening messages — one warm and conversational, one their older default. The warmer greeting alone lifted CSAT from 72.8% to 78.4%.
Nothing about the answers changed. Nothing about the product changed. A single design decision, applied to the first thing a customer sees, moved satisfaction almost six points. That is the entire ROI case for a role most teams haven't created.
Conversation design covers tone, response structure, handoff logic, interaction flow, and response quality — with a discipline worth internalizing: if it's about information, it belongs in the knowledge base; if it's about tone or handling a situation, it belongs in the agent's instructions. Teams that dump every edge case into instructions end up with an LLM reading paragraphs of rules before it can say a word.
Now to the part that ties the bounty to the deepest problem: measurement.
Gumroad is running fully autonomous support. What is it watching to know the agent is any good? If the answer is CSAT, it's flying blind.
CSAT captures less than 10% of conversations, and the responses skew toward the extremes — the delighted and the furious. The vast majority of customers say nothing and move on.
That's a rounding error you're using to coach a system that now handles nearly everything. As AI takes over more conversations end-to-end, the share of your customer experience sitting outside any human review grows — while CSAT still only reflects the sliver that generates a survey response.
Worse, a single score compresses multiple failures. A thumbs-down might be the product, the policy, or the service quality, and it won't tell you which. It hides the exact friction that autonomous support introduces: repeated explanations, multiple handoffs, technically correct answers that still felt wrong.
Fin's answer is CX Score: evaluate every interaction, AI and human, score it 1 to 5, and surface the reason behind the score. For most teams that's roughly five times more conversation coverage than CSAT alone.
Once every conversation is scored and the reasons are visible, the operational loop changes. You can trace which topics score poorly and why, route the issue to product or support or a specific workflow, fix it at the source, and confirm the fix worked before the next customer hits it.
That is the correction loop, running internally, on complete data, closing before customers feel the gap.
Now look at the $25 bounty again. It's the same loop — surface the defect, fix it, close it — but crude, external, and firing only when a customer is annoyed enough to file a report and claim the credit.
The bounty is an honest substitute for the visibility a scored-every-conversation system gives you for free. It works because Gumroad is small and Lavingia is transparent. It does not scale as a quality strategy, and it should not be mistaken for one.
So here's where it lands. Autonomy is table stakes now — 82% are already in, 87% are doubling down.
The moat was never automation. The moat is the maintenance discipline underneath it: knowledge ops keeping answers current, conversation design keeping the tone human, and full-coverage measurement keeping the whole thing honest.
The companies that treat "ship it and watch CSAT" as sufficient will look fine on the dashboard while their agents quietly rot from the inside — one stale article, one flat greeting, one unsurveyed frustrated customer at a time.
Whoever owns the correction loop owns the outcome. Gumroad, at least, is paying $25 a pop to admit it needs one. Most teams haven't even noticed the work moved.