Analytic Memos in AI Qualitative Analysis

Qualitati Research Team · 2026-08-06 · 8 min read

Short answer: An analytic memo is a dated, written note in which a researcher records what a code, theme, or decision means and why it was made. In AI-assisted qualitative analysis, memos are the layer that makes machine-generated codes defensible: the AI proposes, the memo records what the human accepted, rejected, and reinterpreted — and why.

Last updated: August 6, 2026

When AI takes over first-pass coding, the fastest thing to lose is the thinking. Codes appear, themes assemble, a report writes itself — and nowhere in the output does anyone record the judgment calls that produced it. Analytic memos in qualitative research are the cheapest available fix, and they are more valuable now than they were when a human did the coding by hand.

Key takeaways

Analytic memos are dated written notes capturing the reasoning behind codes, categories, and analytic decisions — not summaries of the data.
AI-assisted workflows shift the researcher's scarce time away from mechanical labeling and toward interpretation, which is exactly what memos capture.
Major QDA tools are moving in this direction: MAXQDA's 26.2 release (March 2026) redesigned its memo interface specifically to sit side-by-side with AI-generated summaries.
Memos are the practical substrate of an audit trail, and audit trails underwrite dependability and confirmability claims in a methods section.
Use the five-memo ledger below so memo writing stays a 10-minute habit rather than an unbounded journal.

What is an analytic memo?

An analytic memo is a short, dated, attributable piece of writing in which the analyst thinks on the page about the data rather than describing it. Johnny Saldaña treats memo writing as a companion practice to coding rather than an optional extra — the memo is where a code stops being a label and becomes an argument (see the review of The Coding Manual for Qualitative Researchers in The Qualitative Report).

The distinction that matters in practice is this:

A code says: this segment is about onboarding friction.
A summary says: six participants described onboarding friction.
A memo says: I am treating onboarding friction as distinct from setup cost because participants describe the first as emotional and the second as logistical; if that distinction does not hold up in the next five transcripts, I will collapse them.

Only the third one is falsifiable, and only the third one tells a reader — or a reviewer — how the analysis was actually built. Andrea Bingham's five-phase model of qualitative data analysis (International Journal of Qualitative Methods, 2023) places memoing across phases rather than at a single step, because the reasoning that needs recording is generated continuously.

Why memos matter more in AI-assisted analysis

Three reasons, in order of how much they should change your workflow.

1. The mechanical work is no longer where the time goes

Historically, the reason memoing got skipped was that coding consumed the budget. When an AI pipeline produces a first-pass codebook across 60 transcripts in an afternoon, that constraint changes. Industry commentary on AI in qualitative research — including Lumivero's overview of the state of AI in qualitative research — consistently frames AI features as assistive to summarization and early coding rather than as replacements for analytic judgment. The freed hours should not simply disappear; they are the memo budget.

2. AI output has no visible provenance unless you create it

A human coder who splits a code leaves a trace in their own memory and usually in their notes. A model that splits a code leaves nothing except the split. If you later need to explain why trust became trust in the vendor and trust in the data, and the only artifact is a codebook diff, you cannot. The memo is the provenance record. This connects directly to building an audit trail for AI qualitative analysis.

3. Reviewers and stakeholders are getting more specific

“We used AI to assist coding” is no longer an adequate methods sentence. Memos let you write the version that survives scrutiny: which passes were model-generated, what proportion of suggested codes were accepted, on what grounds the rejected ones were rejected.

The Qualitati Memo Ledger: five memo types

Unstructured memoing tends to collapse into a diary nobody rereads. This is the minimum viable set for an AI-assisted project. Write each one in under ten minutes; length is not the point.

Memo type	Trigger	What it must contain	Typical frequency
Code definition memo	A code is created, renamed, split, or merged	Definition, inclusion rule, exclusion rule, one anchor quote, one boundary case	Per code change
AI adjudication memo	You accept or reject a batch of model-suggested codes	What was suggested, what you kept, the rule you applied, disagreement rate	Per coding pass
Theme construction memo	Codes are grouped into a candidate theme	Which codes, why they cohere, what the theme excludes, competing grouping considered	Per theme
Reflexivity memo	You notice you expected a finding, or find yourself agreeing with the model too easily	The prior assumption, where it may have steered a decision, what would falsify it	Weekly
Negative case memo	A transcript contradicts an emerging theme	The contradicting evidence, whether the theme is narrowed or the case is bounded	As encountered

The AI adjudication memo is the one with no pre-AI equivalent, and it is the one that most changes what you can claim. If you keep only one, keep that.

A memo template you can paste into any tool

Date / analyst: who wrote this and when
Object: the code, theme, transcript, or pass this concerns
Decision: what changed
Reasoning: why, in two or three sentences
Alternative considered: the option you did not take
Falsifier: what evidence would make you reverse this

The last two fields are what separate a memo from a note. A decision with no stated alternative and no stated falsifier is not yet analysis.

Where the tools are going

Memo infrastructure is being rebuilt around AI output, not around handwriting. MAXQDA's 26.2 update, released in March 2026, redesigned the memo window into a side-by-side view of memos and summaries, explicitly so the layout accommodates AI-generated content; the same release added segment-level AI coding where suggestions are approved or declined individually. NVivo's AI features are similarly positioned around summarization, memoing support, and code refinement rather than autonomous analysis.

The pattern across vendors, as of August 6, 2026, is convergent: the model proposes at the segment level, the human disposes, and the tool tries to capture the disposition. Whether your tool captures it or not, you should.

Where Qualitati fits

Qualitati is an AI user research platform for product, UX, and customer insights teams. Its analysis layer is built around the proposal-and-disposition pattern above rather than around a single opaque output.

ThemeLens runs a map-reduce thematic analysis across up to 100 transcripts at once, mapping codes to your research questions and synthesizing themes with participant-anchored quotes — so the theme you memo about is traceable to the segments that produced it.
QDA Workspace supports AI-assisted inductive and deductive coding, codebook generation, and theme visualization, with the codebook as an editable object rather than a fixed model output.
AI-moderated interviews and focus groups, conversational surveys, and Voice Analytics feed the same corpus, in ten languages.

Qualitati does not write your memos, and it should not. The interpretive claim is the researcher's to make and to sign.

Limitations and trade-offs

Memos are not validation. A well-written memo documents reasoning; it does not prove the reasoning was right. Memoing complements, and does not substitute for, inter-rater checks, negative case analysis, and member checking.
Memoing can rationalize. Writing a justification after accepting a model's suggestion can harden a decision you have not really scrutinized. Writing the “alternative considered” field first is a partial guard.
Volume creates its own problem. At AI scale, undisciplined memoing produces a second corpus nobody reads. The five-type ledger exists to cap this.
Do not memo with the model that coded. Asking an LLM to explain why it produced a code yields a plausible post-hoc narrative, not the actual basis of the output. Treat any such text as a prompt for your thinking, never as the memo itself.
Human review note: methodological claims in a publication — particularly about dependability, confirmability, or reliability — should be reviewed by a qualified methodologist against your target venue's reporting standards.

Who this is for — and when not to bother

Use this if you are producing findings that will be published, audited, used to justify a costly decision, or handed to a team that was not in the room. Skip the full ledger if you are running a one-off exploratory sprint on five interviews where you are both the analyst and the only consumer — keep the code definition memo, drop the rest.

Bottom line

Analytic memos in qualitative research were always the practice most likely to be cut for time. AI-assisted analysis removes that excuse and simultaneously raises the cost of skipping them, because machine-generated codes arrive with no reasoning attached. The five-memo ledger — code definition, AI adjudication, theme construction, reflexivity, negative case — is a defensible minimum that fits inside a working week.

FAQ

What is the difference between a memo and a code in qualitative research?

A code is a label applied to a data segment. A memo is written reasoning about that label — its definition, boundaries, and the decision to use it. Codes organize data; memos explain the organization.

How long should an analytic memo be?

Long enough to state a decision, a reason, an alternative, and a falsifier — usually 100 to 300 words. Length is not a quality signal; a short memo that names the alternative you rejected is worth more than a page of description.

Can AI write analytic memos for me?

It should not write the ones that carry your interpretive claim. An LLM asked to justify its own coding produces a plausible reconstruction, not the actual basis of its output. AI can usefully draft the descriptive scaffolding — what changed between passes, which codes moved — leaving you the reasoning.

Do memos count as an audit trail?

They are the main content of one. A dated, decision-level memo series lets an external reader reconstruct how conclusions were reached, which is what dependability and confirmability claims rest on. Pair them with versioned codebooks and preserved model outputs.

How many memos does an AI-assisted project need?

Fewer than an unstructured approach produces. Expect one memo per code change, one per coding pass, one per candidate theme, one reflexivity memo per week, and one per negative case. On a 40-transcript project that is typically 30–60 memos, not hundreds.

Where should memos be stored?

Wherever they stay attached to the object they concern — the code, the theme, the pass. Detached memos in a separate document reliably go unread. Most QDA tools now support object-level memos; use them.

Start memoing your AI-assisted analysis

Run an AI-moderated interview, focus group, conversational survey, or thematic analysis project and keep the reasoning attached to the output. Start free with 30 credits — no credit card required — or view transparent pricing. If you are evaluating options, compare Qualitati with NVivo, ATLAS.ti, and MAXQDA.