AI Research Tools: Repositories Like Dovetail and Condens
Research repositories store notes, quotes, and themes. Dovetail and Condens add AI. What to check when choosing and how to test accuracy. A practical guide.
What problem do they solve?
A product team runs dozens of interviews a year. Without a repository, findings live in loose documents, chat threads, and the memory of whoever ran the session. When someone asks “have we talked to users about this?”, nobody can answer.
A repository centralizes the material. AI adds the layer that costs the most time: reading everything, finding patterns, and returning quotes. The decision still belongs to the team.
What do Dovetail and Condens offer?
Both are mature repository tools. Their plans and prices change often; this is what their official pages showed on 6 November 2026:
| Aspect | Dovetail | Condens |
|---|---|---|
| Free trial or plan | Its pricing page lists no free plan | 15-day trial, no card required |
| Plans | Bronze (annual platform fee plus credits per data point processed), Silver and Gold on quote | Lite from €15/month (€165/year), Business from €6,000/year, Enterprise custom |
| Featured AI | Agents, chat, theme clustering; digital twins from Silver | AI transcription, suggested tags, summaries, questions over your data; ChatGPT and Claude connection from Business |
| Privacy | AI training opt-out from Silver; AI redaction in Gold | Anonymization that grows by plan; EU or US hosting, GDPR and DPA |
Condens says its AI answers from the project data, with safeguards, and that people have the final say on every output. Dovetail is described as a repository with AI analysis on top; that description comes from an article by Great Question, a competing platform, and neither that text nor the pricing pages measure the accuracy of automatic tagging. That is why the test in the next section matters more than any comparison.
How is it used in a project?
A reasonable flow:
- Define a theme frame before uploading anything: which questions the study answers.
- Upload the transcripts and let the AI propose groupings.
- Review each grouping and merge, split, or rename themes.
- Link each finding to its quotes and to the participant who said it, without personal data.
- Leave the repository searchable for the next team.
This flow connects with AI affinity mapping and interview synthesis. If the transcripts are poor, the repository only stores well-organized noise: review the transcription step first.
How do you test accuracy before buying?
Run a short test with your own material:
- Take three real interviews, anonymized.
- Tag about 50 quotes by hand with your own themes.
- Ask the tool to group the same interviews.
- Count how many quotes match your tag, how many land in a wrong theme, and how many do not appear.
- Decide on those numbers, not on the vendor demo.
This test takes an afternoon and prevents buying a tool that misreads your kind of interview.
When should you NOT use it?
- If your team runs one or two interviews a quarter: a spreadsheet is enough.
- If the material holds sensitive data and the tool does not offer the processing terms you need.
- If nobody will maintain the repository: without an owner, it fills with inconsistent tags.
- If you want recruiting or running studies. Those are other categories of tools.
Example
Illustrative example: a product team has two years of interviews stored in loose documents. It picks a repository tool, runs the test with fifty hand-tagged quotes, and finds the AI groups billing complaints well but confuses support issues with onboarding ones. The team decides to use the AI grouping as a draft and keep manual review for the support themes.
What it learned: the tool was useful, but the real number came from its own test, not from the brochure.
How do you structure a finding in the repository?
A useful repository does not store only quotes. Each finding should have the same structure, so it can be searched and compared six months later:
| Field | What to store | Example |
|---|---|---|
| Claim | One sentence with a subject and a behavior | “People who pay in installments abandon if they do not see the final total” |
| Evidence | Two or three quotes with their timestamp | Quote from P03, minute 12:40 |
| Participants | How many, and from which segment, support the finding | 5 of 8, frequent travelers segment |
| Study and date | Which research produced it and when | Payment interviews, Sept. 2026 |
| Status | Whether it is validated or a hypothesis | Hypothesis, pending test |
The status field is the one most often neglected. Without it, a hypothesis from a year ago looks like a confirmed fact.
How do you keep it alive?
An abandoned repository is worse than none, because it creates a false sense of history. Two practices help: assign one person to review the tags each quarter, and merge nearly identical tags before they multiply. Also check which findings were used in real decisions; those nobody consults can be archived.
Common mistakes
- Choosing from a demo prepared by the vendor.
- Uploading material without anonymizing it or reviewing the data terms.
- Accepting AI groupings without reviewing them.
- Creating dozens of nearly identical tags that nobody will reuse.
- Believing the repository replaces the judgment of the researcher.