An AI research agent: user research without a research team

What an AI research agent is

An AI research agent is an AI agent, such as Claude or ChatGPT, connected to a research tool so it can run studies with real people: write the questions, publish the study, read the answers and report what it found. The connection works through MCP, the Model Context Protocol, the open standard that lets AI agents use other software. More than 10,000 MCP servers were active by the end of 2025, and Claude, ChatGPT and Cursor all support it.

The agent is not the participant. It does not invent customers or simulate answers. Real people answer, by voice, screen recording or text, and the agent does the work around them that a researcher would do.

The most underrated way to grow

Knowing your audience better than your competitors do is the most underrated way to grow. Most teams spend their energy elsewhere: shipping features faster than ever, chasing the latest growth hack, copying what worked for someone else. John Cutler has a name for the teams that measure themselves by output: feature factories. When anyone can build anything, building more is not what sets you apart. Building the right thing is.

The companies that do it are rare, and they pull ahead. In Forrester’s 2025 survey, only 6% of companies qualified as customer-obsessed, and their leaders reported 41% faster revenue growth than the rest. Y Combinator has told founders the same thing for years: the most important tasks for an early-stage company are to write code and talk to users. Your AI agent already helps you write the code. With Feedbaq, it can help you talk to your users.

Research is a job most teams cannot staff

Good research takes real skill and real time, and a startup running fast and lean rarely has either to spare. It means turning a vague worry into a question a study can answer, writing questions that do not lead, finding people, listening closely, asking the follow-up, and then reading everything without seeing only what you hoped to see. Companies hire specialists for this for a reason.

Even companies that invest in design rarely have enough researchers. The most common ratio Nielsen Norman Group found was one researcher for every five designers and fifty developers, and about one in seven companies in User Interviews’ 2025 survey had no dedicated researcher at all. At a startup the number is usually zero. In Maze’s 2026 survey, time and bandwidth was the biggest barrier to doing research, and only 39% of product managers ran research themselves. The slowest part comes after the answers arrive: in Lyssna’s 2025 survey, slow manual work was the top problem in making sense of research, and most teams took one to five days to do it. So the study that would have changed the roadmap never happens, or happens once and ends up in a folder.

What your agent does at each step

Feedbaq is a tool for collecting feedback from your users. You write a few questions and share them as a link. People answer by voice, screen recording or text, and you get every answer with a transcript. Your AI agent can set it up and analyze the answers for you.

Your agent does that through Feedbaq’s MCP connection, and it can do the researcher’s work at every step of the loop while you make the decisions. Here is each step, what it asks of a researcher, and what the agent does there, with the Feedbaq tools it calls.

1. 1

   Decide what to learn

   What it takesTurning a vague worry into a question a study can answer, and choosing who can answer it.

   Your agent, with FeedbaqStarts from what it already knows: your product, your roadmap, your notes, your support tickets. It proposes the question and who to ask.

2. 2

   Write the study

   What it takesNeutral questions about real past behavior, in the right order, in the right format.

   Your agent, with FeedbaqWrites the questions, chooses voice, screen, text or choice for each, and sets AI follow-ups with instructions.

   create_studyupsert_questionset_image

3. 3

   Set it up and publish

   What it takesA clear welcome, a sensible length, a limit on responses.

   Your agent, with FeedbaqWrites the welcome and thank-you pages, applies your brand, sets a response limit and publishes the study.

   update_studypublish_study

4. 4

   Reach people

   What it takesFinding the right people and getting them to answer.

   Your agent, with FeedbaqDrafts the message for your users, customers or waitlist. You send the link. Participants need no account.

5. 5

   Listen and ask more

   What it takesHearing what matters and asking the follow-up in the moment.

   Your agent, with FeedbaqFeedbaq’s AI follow-ups ask about each answer while people respond, so every answer arrives with its why.

6. 6

   Make sense of it

   What it takesReading every answer, finding the patterns, weighing them fairly, keeping the evidence.

   Your agent, with FeedbaqReads every full response with transcripts, groups what people said, counts it, and quotes them with a link to each answer.

   list_responsesget_response

7. 7

   Decide, and go again

   What it takesTurning findings into a decision, and knowing what to ask next.

   Your agent, with FeedbaqSuggests what the findings mean for your plans and drafts the next study. You decide.

   create_study

A study from start to finish, in one chat

In practice, the loop takes two questions from you and one email you send. The study runs itself in between.

The findings are what a good researcher would bring back, and they came from the people themselves. The decision, to fix the import before touching the price, stays with you.

Findings you can check

An AI finding you cannot trace to what someone actually said is a rumor, so every finding from your agent should point back to its source. AI is fast at reading many transcripts and finding patterns across them. In a 2026 survey by Condens, 71% of researchers said AI made their analysis much faster. It also makes mistakes: when Teresa Torres tested ChatGPT on her interviews, about 30% of the quotes it produced were wrong or not in the source at all. Her advice is to always go back to the source.

Feedbaq is built so your agent can, and so can you:

• The full response, not a summary. The agent reads every answer and follow-up in full, with the exact wording of the question each person saw.
• A link to every answer. Each answer comes with its own link, and every word of a spoken answer is timed, so the agent can send you to the exact second a quote was said.
• The recording itself. Click through and hear the person say it, in their own voice and with their own emphasis.

That last point matters beyond accuracy. Nielsen Norman Group warns that a team that hands research to AI gets a report, but not the learning. Listen to a few of the recordings your agent points you to. Hearing a customer describe their problem changes how you build in a way a bullet point cannot.

Built for startups, priced for startups

Letting an agent run studies with real people has meant an enterprise contract. Feedbaq includes it in its one plan, for $49 a month. Research platforms that let agents launch studies sell that on enterprise plans, the kind you negotiate with a sales team. Feedbaq is built for the teams that have never had a researcher: founders, product managers, designers and marketers who need answers this week. Every feature is included: voice answers, screen recordings, AI follow-ups and the MCP connection.

Connect your agent

1. Add Feedbaq to your AI agent. It works with Claude, ChatGPT (in developer mode), Codex, Cursor, Grok and any MCP client that supports HTTP with OAuth.
2. Sign in and pick a workspace. The agent can only do what you can do, and only in that workspace.
3. Ask for what you want to learn. “Find out why trials go quiet.” “Test these two headlines with our waitlist.” Or start from the customer discovery template and ask your agent to adapt it to your market.
4. Send the link, then ask what people said.

Building a study is free. You subscribe when you publish: $49 a month, with 100 responses and every feature included.

Questions people ask

Sources

1. Forrester, “The state of customer obsession” 2025. Growth figures are reported by executives.
2. Y Combinator, “YC’s essential startup advice” Geoff Ralston, 2017.
3. Nielsen Norman Group, “UX, design and development ratios” Kate Kaplan, 2020.
4. User Interviews, State of user research 2025.
5. Great Question, MCP integration Offered on its enterprise plan, checked 30 September 2026.
6. John Cutler, “12 signs you’re working in a feature factory” 2016.
7. Lyssna, Research synthesis report 2025.
8. Maze, The future of user research 2026.
9. Teresa Torres, “Can AI analyze customer interviews?” Product Talk, 2025.
10. Condens, AI in user research analysis 2026.
11. Nielsen Norman Group, “Human-led research still matters” 2026.
12. Model Context Protocol, “MCP joins the Agentic AI Foundation” December 2025.