AI agents are tireless testers. They will sign up, build a report and press Share fifty times without complaining. What they can't do is tell you whether a problem would make a real customer give up. An agent has never paid you, never churned on you, and never felt the small irritation that decides whether someone comes back tomorrow.
That is why every Swarm finding carries a label: medium confidence until people confirm it. This month we made confirming it one click.
One button for the whole run
At the top of a run's Findings tab, next to the summary of what the agents found, there is now a single Confirm with 5 people button. The same button sits on every finding, so you can act from wherever you are reading.
Pressing it makes one draft study about every open finding in the run. Earlier versions asked about findings one at a time, which meant several small studies for one run. One study per run is simpler to review, cheaper to field and easier to read afterwards.
What the draft study contains
The button builds an AI-moderated interview, already filled in:
- 5 everyday people from the User Evaluation pool.
- One question per open finding, worst first, up to 12. Each question describes the page and the steps, says what the agents found, and ends with "Did that happen to you? Start with yes or no." A follow-up asks what they expected to happen instead.
- Your site as the stimulus, so participants look at the real thing while they answer.
- A device screener, so people answer on the kind of device the agents used.
- A title that ties it back to the run: "Confirm:" followed by the run's name.
Then it opens the study in the normal study editor, on the interview guide, for you to read.
Nothing is charged until you launch
Making the draft costs nothing. You can edit any question, remove a finding you already know about, change the number of interviews, or leave the draft for later. The price is shown before you launch, at the same per-interview prices as any other study: everyday people are $39 per completed interview, so 5 come to $195, with recruiting, the participants' payment, the AI interviewer, recordings, transcripts and the study report included. The full list is on the pricing page.
If you want to hear how the questions sound first, Run a pilot first lets you take the interview yourself. A pilot is free, isn't counted, and never reaches the study's results.
Press the button again and you get the same study, not a second one. Once it exists, the Findings tab says Review the study while it is a draft and Open the study after launch.
Why 5 people
Five is not a magic number, but it is a useful one for this job. The question here is narrow: did this specific thing happen to you, and did it matter? For that, five people on the right device answer quickly and cheaply, and the answers are easy to read. If four of five people hit the same broken Share dialog, you have your priority. If none of them do, the agents may have found something odd about their own setup, and you have saved an engineering sprint.
For bigger questions, such as why people choose a competitor or what they expected from onboarding, you would run a fuller study. Our guide to AI-moderated versus human-moderated interviews covers that choice.
How this changes the way you read a Swarm run
The point of Swarm is not to replace research with agents. It is to point real research at the right places, faster. A typical loop now looks like this:
- Run agents on a journey, such as signup and the first report.
- Read the triaged findings, with the screenshot behind each one. See triaged findings and the flow map.
- Fix what is obviously broken. A 500 from your own API needs no confirmation.
- Confirm the rest with 5 people. Layout issues, confusing steps and slow pages are where human judgement matters.
- Act on what people confirm, with their clips and quotes to show the team.
Agent findings stay labelled as agent evidence throughout the product, and they stay at medium confidence until people confirm them. That rule is the same for everyone, and it is the reason the button exists.
Who runs the interviews
The people are real, screened participants from the User Evaluation pool, not AI personas. The AI runs the conversation and the analysis, and the answers come from humans. If you want to know where that line sits and why it matters, read synthetic users versus real participants.
Try it
Confirm with 5 people is on every Swarm run with open findings, on every plan. Swarm runs are included in each plan every month (1 on Free, 5 on Pro and 25 on Team), and interviews are priced per completed session on any plan. Start with how Swarm works, or see how AI-moderated interviews run.
