# AI Affinity Mapping: Grouping Notes Without Losing the Team > Affinity mapping groups observations into themes. AI can propose groups, but if the team does not discuss them, the map loses value. A practical guide. *Tags: ai, research, affinity-mapping, collaboration* --- > [!info] Quick Definition > **Affinity mapping** is a visual method for grouping research observations, such as quotes, behaviors, and responses, into themes that repeat. Its value lies in the map and in the conversation of the team that builds it. AI can propose initial groups, but it should not replace that conversation. ## How does the method work? According to a guide from Capicua, the method has four steps: 1. **Capture**: write each observation on its own note. 2. **Cluster**: bring related notes together into themes. It helps to let themes emerge from the notes instead of fixing them in advance. 3. **Label and prioritize**: give each group a clear name in plain language. If the team cannot agree on what matters most, the guide recommends dot voting to rank the groups. 4. **Convert into actions**: assign each priority theme a next step and an owner. Also record the source of each note (which participant said it), so you can go back to the original quote. The method descends from the KJ method, which the Japanese anthropologist Jiro Kawakita developed in the 1960s to organize unstructured field notes. The underlying idea is that the group reaches a shared understanding, rather than someone handing over a classification. ## What does AI add, and what does it not? AI adds speed to the grouping step. With hundreds of notes, a model can propose in seconds a first structure that would take hours by hand. That is useful. What it does not add is agreement. A short paper on qualitative coding with language models (Kirsten and colleagues) concludes that adopting models calls for task-by-task evaluation. The authors found that large codebooks and extensive contexts make the work harder for both people and models. It is an initial result, not a general rule, but it works as a warning: do not accept a grouping just because it looks tidy. Another study measured how consistent several models were when analyzing themes in a single interview transcript, with six runs per model, and found high agreement between runs (Cohen's kappa above 0.80 for all three). That shows the model repeats its results, not that the themes are useful for your product. An AI-generated group can be coherent and still say nothing to your team, because nobody on the team built it. That is why the AI map should be treated as a draft. The biggest risk is that the team accepts it without discussion and turns it into the official finding. ## How do you combine them without losing the team? A flow that keeps the discussion alive: 1. **Silence first**: each person groups notes quietly for a few minutes, without seeing the AI proposal. 2. **Show the AI proposal**: display the model's grouping, with the notes it assigned to each group. 3. **Compare**: mark the groups that match, the ones that split, and the notes that moved. 4. **Discuss the differences**: the differences are the most valuable part. Ask why a note landed somewhere else. 5. **Name and prioritize together**: the team decides the names, not the model. This way the AI saves the mechanical step, and the team keeps the understanding. ## When should you NOT use it? - When the team will read only the summary and not the notes. - When the study has very few interviews: with ten notes, the manual method is faster. - When the grouping will be used as evidence for a client without reviewing the quotes. - When the topic is emotional or cultural and the nuance gets lost inside a category. ## Example Illustrative example: a team has 240 notes from twelve interviews about onboarding in a finance app. Before showing the AI proposal, each designer groups quietly. When they compare, they find that the model put "I don't understand the terms" together with "I don't trust the app," two different problems. They discuss, split the groups, and name trust as the priority theme. The final map has four themes, and each has an owner. What they learned: the disagreement with the AI was the most useful part of the session. ## How many notes and participants does a map need? The method sets no number. What does matter is coverage: the map should include observations from every participant in the study, not only from the most articulate ones. If one person contributes twenty notes and another contributes two, the groups that emerge may lean toward the first. Before you start, count notes per participant. If the imbalance is very large, decide whether to correct it, for example by limiting notes per person, or to record it as a limitation of the map. ## How do you name groups so they are useful? A group's name decides whether the team understands it and uses it. A weak name describes a topic ("Trust"); a useful name asserts something that can be discussed ("They do not trust that only the agreed amount is charged"). Three rules help: 1. **Use the participants' words** when you can. If three people say "the charge scares me," the group is called that, not "Perception of costs." 2. **Write the name as a sentence.** A claim can be disproved with data; a label only gets filed away. 3. **Review groups with a single note.** If a group has one observation, ask whether it is a theme or an anecdote. If it is an anecdote, do not turn it into a priority. After naming, do a final pass: does each group have someone responsible for follow-up? Are there groups that overlap and should be merged? A map with fifteen groups often says less than one with four well-named groups. ## Common mistakes - Showing the AI grouping before the team has grouped on its own. - Accepting groups without reading the notes that form them. - Naming groups with the model's words instead of the team's. - Prioritizing by group size without considering how serious the problem is. - Leaving the map in a file with no next step or owner. ## Sources - [Affinity diagram UX: guide by Capicua](https://www.capicua.com/blog/affinity-diagram-ux) - [Decoding Complexity: Exploring Human-AI Concordance in Qualitative Coding (arXiv 2403.06607)](https://arxiv.org/abs/2403.06607) - [Multi-LLM Thematic Analysis with Dual Reliability Metrics (arXiv 2512.20352)](https://arxiv.org/abs/2512.20352) --- Source: https://www.fernandoux.com/en/wiki/ai/ai-affinity-mapping/