Agentic analysis is Skimle's deepest analysis mode. You state a research question, an objective, or the deliverable you need; a setup agent reads your project and proposes a coding framework you can edit. Skimle then codes every relevant document, builds analytical themes, searches for counter-evidence, and writes a report answering your question, with every claim linked back to a quote.
Every other analysis mode in Skimle answers the question "what is in this corpus?" Agentic analysis starts from the research question to develop an answer based on evidence (and counter-evidence) to support it.
Automatic thematic analysis and predefined categories / inductive analysis map your material into categories, which is what you want when you are exploring. Agentic analysis takes a research question, builds the framework the question implies, codes against it, and then does the analytical work on top: synthesis, contradiction hunting, and a written answer.
This guide covers what it does, how to drive it, and worked examples from both academic and business projects. If you want the reference documentation rather than the walkthrough, see agentic analysis in the docs.
How is agentic analysis different from the other modes?
| Mode | You give it | It produces | Best for |
|---|---|---|---|
| Automatic thematic | Nothing, just documents | A category tree of what is in the corpus | First look at new material |
| Predefined categories | Your own coding framework | Your framework applied consistently, with sub-categories emerging from the data | Repeat studies, established codebooks, open exploration with a specific direction or lens |
| Agentic | A research question, objective, or deliverable | Coded evidence plus analytical themes, counter-evidence and a written report | Answering a specific question with depth |
The trade is time against depth. Agentic runs longer than the other modes, because it makes several passes over your material rather than one. Credits in Skimle are spent when you confirm documents, not when you analyse them, so running an agentic analysis (or re-running it) does not cost you anything beyond the wait. Use the lighter modes to explore, then run agentic when you know what you are asking.
What can you type into the setup box?
This is the part people get wrong on their first run. The prompt is "What do you want to find out, achieve, or produce?", and all three work:
- A research question. "Why do staffing decisions diverge across presidential terms?"
- An objective. "Indicate which customer segments the data suggests we should invest in next year."
- A deliverable. "A prioritised list of onboarding pain points, ranked by severity."
You do not need to convert an objective into a properly formed research question yourself. The setup agent infers the implicit question, states it back to you in one line, and carries it forward. Naming the deliverable is often better than naming a question, because it constrains what the analysis has to produce: a prioritised list forces ranking steps into the framework, and a design recommendation forces constraint and trade-off steps.
What does not work well is "analyse my data" or "what is in here?" With nothing to infer, the agent will stop and ask you questions rather than propose. If that is genuinely where you are, run e.g., the Automatic thematic analysis first and come back once you know what you want to ask.
The 7 stages of an agentic run
Stage 1: The setup chat proposes on the first turn
You type your question into a two-pane setup dialogue. The agent has already read your project before it answers: document count, existing metadata fields, entities relevant to your question, prior analyses on the project, and a semantic search over content matching what you asked for.
It writes two or three sentences stating the question it inferred and the assumptions it had to make, then proposes a framework in the right-hand pane. The default behaviour is to propose rather than interrogate you, on the reasoning that a framework built on a stated assumption is more useful than an empty pane and a list of questions. Its open questions arrive as a short bulleted list at the end of the rationale, framed against what it already chose ("I split by role, switch to region instead?").
Stage 2: You edit and confirm the framework
The proposed framework has three parts:
- An analysis type from the nine described below.
- First-level categories, typically four to seven, created from Skimle's list of proven lenses and from a prior agentic analysis on the same project when one aligns.
- Broader dynamics, two to four higher-level analytical questions that the synthesis stage will investigate. These become the quests for the Skimle investigation agents, so they are worth reading carefully.
You can rename, add or remove anything. This is the highest-leverage five minutes in the whole process, because everything downstream is built on this skeleton.
On confirmation, Skimle creates a top-level insight category named after your question, opens an analysis log memo that records every step, and queues the run.
Stage 3: Document screening and empirical extraction
Skimle first classifies your documents: what role each plays, what type it is, and whether anything is flagged as low quality. Extraction then skips documents that are irrelevant to the question rather than working through them. This means you can safely do the analysis over a large mixed quality corpus as Skimle will automatically drill down to the docs that matter.
For each first-level category, Skimle uses smart pre-selection to find the chunks most likely to contain relevant material, then examines those chunks in depth.
Stage 4: Metadata enrichment and within-case trajectories
In parallel with extraction, Skimle checks whether your question implies a comparison your metadata cannot currently support. A question about change over time needs a period field; a question about staffing needs a role field. When a field would help and does not exist, Skimle proposes it, names the values meaningfully, and codes every document. When your existing metadata already covers the distinction, it records that reasoning and moves on.
Also in this stage, Skimle traces within-case trajectories: how individual cases develop across the material, rather than only what is common across cases. For process and journey questions this is often where the interesting material sits.
Stage 5: Analytical synthesis
For each broader dynamic, an investigation agent searches the empirical insights, the metadata and the temporal patterns over multiple steps, then synthesises two to five analytical themes.
Stage 6: Counter-evidence and revision
For each analytical theme, Skimle deliberately searches for evidence that contradicts it. Where counter-evidence exists, it is recorded against the theme, and the theme prose is then revised to integrate it, for example by making the claim conditional to specific circumstances (e.g., "In the US...") rather than absolute. Themes with strong counter-evidence are flagged.
This stage is the reason we consider agentic output defensible rather than merely fluent. A synthesis that has not looked for its own contradictions is an argument, not an analysis. We have written about why that discipline matters in rigour means being able to return a negative result.
Stage 7: Findings summary, storyline and report
Finally Skimle evaluates the evidence across the whole analysis, orders the findings into a narrative structure (which finding is the answer, which are mechanisms, tensions, levers or context), and writes:
- a findings summary, the cross-cutting answer to your question;
- a storyline, an extended narrative write-up, including a flowchart of how themes relate (one theme causes, enables, precedes, contrasts with or is associated with another);
- an executive summary and research report in the Research reports view.
Every reference in every artefact is re-verified at the end of the run.
What are the different analysis types that Skimle can apply?
Horses for courses - Skimle automatically detects the type of analysis you are implying in the research question and decides on the approach based on that. Some example analysis types are listed below:
| Type | What it is for | Example research question |
|---|---|---|
| Diagnostic | Explain why something happens: conditions, mechanisms, outcomes | "Why do some pilots convert to rollouts and others stall?" |
| Root cause | Trace a specific problem to its roots and what blocks improvement | "Why has our handover process kept failing despite three fixes?" |
| Classification | Map the landscape into types or themes, without asking why yet | "What kinds of concerns do respondents raise about the reform?" |
| Comparative | Contrast groups defined by a metadata field | "How do renewing and churning accounts describe value differently?" |
| Process / journey | Reconstruct how something unfolds over time, stage by stage | "How does a team's understanding change between first trial and routine use?" |
| Evaluation | Judge something against what it set out to achieve | "Did the mentoring programme deliver what it promised?" |
| Prioritisation | Rank problems or actions so the reader knows what to tackle first | "A ranked list of the integration blockers by reach and severity" |
| Prescriptive / design | Move from evidence to a defensible recommendation | "What should we change in onboarding, and what constrains each option?" |
| Opportunities | Surface unmet needs, gaps and latent demand | "Where is demand in this category currently unserved?" |
Academic examples
Building a process description in grounded theory fashion
This is the case agentic analysis handles with ease, and the one that is hardest to do by hand.
Suppose you have 34 interviews with members of research groups who adopted a new laboratory technique at different points over four years, and your interest is in how the adoption unfolds rather than in whether it succeeded. The conventional coding pass will give you a category tree of topics: training, equipment, scepticism, cost. What it will not give you is a sequence. With Skimle you could of course study how the themes evolve over time using the Comparisons-tool under Visualisations, but that would depend on clear time stamps in the interview metadata and not support the people telling a narrative over time. Agentic analysis constructs the process view automatically.
How to state it. "How does a research group's understanding of the technique change between first exposure and routine use?"
What to check in the framework editor. The stages proposed should be your phenomenon's stages, not generic ones. If the agent proposes Entry and Expectations, Early Experience, Turning Points, Current State, Exit or Outcome, and your data suggests the real structure is exposure, sceptical trial, a champion emerging, and institutionalisation, rename them. The template is a starting skeleton, and renaming nodes to the language of your field is usually the single most valuable edit you make.
What to put in the broader dynamics. For process questions the useful dynamics are: establish the typical sequence and where individual accounts diverge from it; locate the transitions where groups stall or abandon; examine how the early experience shapes what happens later. Those are the questions that convert a stage-coded corpus into a process theory.
What you get. A stage-coded tree, within-case trajectories showing how individual groups moved (including the ones that went backwards), analytical themes about the mechanisms driving transitions, and counter-evidence noting the cases where the sequence did not hold. In process research the deviant case is often where the theoretical contribution lives as it teases out boundary conditions and causal mechanisms.
This is compatible with a grounded theory posture rather than a replacement for it. You still do the theoretical work; the machine does the sorting and gives you the material organised by sequence rather than by topic. Our grounded theory practical guide covers the methodological side, and inductive, deductive and abductive coding covers where the Diagnostic type's abductive logic fits.
Explaining divergent outcomes across cases
A second common academic shape: you have a set of cases with different outcomes and you want an explanation rather than a description.
How to state it. "Why do some university spin-outs abandon their original application while others commit to it?" This is a Diagnostic question, likely seeded from the extended causal skeleton (antecedents, catalysts, mechanisms, confounds, outcomes).
The edit that matters. Check whether confounds is present. Explanatory frameworks without an explicit place for confounding processes tend to produce clean stories, and clean stories from qualitative data usually mean something has been filtered out. The broader dynamics should include testing whether the proposed antecedents are necessary, sufficient or merely common, and an explicit search for cases that contradict the emerging explanation.
Why the counter-evidence stage matters here. For a paper you will have to defend the explanation against exactly the objections the counter-evidence search generates. Having them surfaced and cited before a reviewer raises them is worth the run on its own. The rigour framing in Gioia, Corley and Hamilton's notes on the Gioia methodology applies directly: the credibility of an inductive account depends on showing the path from raw data to concept, which is what the citation chain gives you.
Comparing two sites or two groups
How to state it. "How do clinicians at the two sites describe the same protocol differently?" Type: Comparative, which needs a metadata field defining the groups. If your documents are not yet tagged with site, add the field before running, or let the metadata enrichment stage propose it.
What to watch. Comparative runs are the ones where the framework editor question "I split by role, switch to region instead?" can matter a lot. The segmentation choice determines the entire analysis, and it is much cheaper to change it in the editor upfront than after a full run.
If you work in academic research, the academic researchers use-case page covers how this fits a dissertation or paper workflow.
Business examples for agentic analysis
Market research: ask the main questions
For a consumer insights study, the productive pattern is to run two or three agentic analyses on the same corpus, each asking one main question, rather than one analysis asking everything.
Say you have 60 category interviews plus 800 open-text survey responses. Three runs:
- "What drives brand choice in this category?" Classification. Produces the map of decision drivers with their relative weight and the language consumers actually use.
- "Where is demand in this category currently unserved?" Opportunities, seeded from the unmet-needs or jobs-and-outcomes skeleton. The dynamics here look for workarounds people have improvised, needs implied by behaviour rather than stated directly, and demand concentrated in a segment nobody serves.
- "How do heavy and light buyers describe the category differently?" Comparative, split on a usage metadata field.
Each run produces its own report. Together they cover the study. Trying to make one analysis answer all three questions produces a framework with too many first-level nodes and a synthesis that hedges.
A note on sequencing: the second and third runs will be seeded from the first, because Skimle biases new framework proposals towards a prior agentic analysis on the same project when one aligns reasonably. That keeps your category language consistent across the study, which helps if you are going to compare or combine the outputs later. Our guide on combining customer insights across feedback channels explains why a stable structure across a corpus is worth protecting.
For the wider methodology, see qualitative consumer insights research and the customer and market researchers use-case page.
Consulting: ask for a specific angle, or just name the deliverable
Consulting work rarely wants a map of the corpus. It wants a specific cut, aimed at a specific decision, by Thursday. This is where the deliverable framing comes to play.
Instead of a question, type the deliverable. "A prioritised list of the operational pain points blocking the post-merger integration, ranked by how widely and how severely they are felt across the two organisations." That gets classified as Prioritisation and seeded from the severity-and-frequency skeleton (Widespread and Severe, Severe but Rare, Frequent but Minor, Existing Workarounds). The dynamics will include weighing each candidate on reach and severity, estimating the effort attached to each from what the evidence says, identifying dependencies, and flagging where the evidence is too thin to rank confidently.
That last point is worth dwelling on. A ranked list from an analysis that also tells you which rankings it is not confident about is a much safer artefact to put in front of a client than a clean top ten.
For a recommendation rather than a diagnosis, name it that way: "What should we change in the field service handover, and what has to be true for each change to work?" Type: Prescriptive / design, which traces each proposed response back to the need it answers, surfaces constraints, and identifies trade-offs where serving one need costs another.
For a due diligence angle, ask the diligence question directly: "What would have to be true for this company's retention story to hold?" The Diagnostic or Evaluation types both work, depending on whether you want the mechanism or the verdict.
The output goes into a deliverable rather than replacing one. The research reports view gives you the write-up and the flowchart of how themes relate, and every claim is clickable back to the interview it came from, which is what you want when a partner asks where a finding came from in a review. See consulting research synthesis from calls to deliverable for the surrounding workflow and the consultants and investors use-case page.
Customer insights: the churn question
"How do accounts that churned describe value differently from accounts that renewed?" Comparative, split on outcome. Run it over exit interviews, renewal calls and NPS verbatims together.
The useful output is rarely the list of churn reasons, which you probably already know. It is the Common Ground node: the things both groups say identically, which tells you what is not differentiating, and therefore what to stop investing in. Also, in the Visualisation tool you can then look at different cross-tabulations, trends over time and differences between groups to further discover insights.
What does the analysis leave in your project?
When a run finishes, the project contains:
- the category tree under your research question, with coded insights and verified quotes, visible in the categories view, the table view and the document view like any other analysis;
- category summaries for every populated category;
- the analysis log memo, a step-by-step narrative of what Skimle did, why, and what it found, including the setup conversation;
- analytical theme memos, one per broader dynamic, holding the synthesised themes with citations and counter-evidence;
- new metadata fields with coded values, where the analysis proposed any;
- the executive summary, research report and storyline in the Research reports view.
All of it is ordinary Skimle data. The insights behave like any others, so you can tag them, edit them, add notes, and export them. Check the Research reports tab to find them.
How do re-runs work?
You can re-run an agentic analysis at any time, and it is faster than the first run. Skimle preserves empirical extraction that already passed strict coding and only considers chunks the prior run did not reach, so adding twelve new interviews to a project does not re-code the first forty.
The analytical layer is different: synthesis, counter-evidence and the findings summary are redone from scratch, so the themes reflect the current state of the evidence rather than an outdated snapshot. Placeholder themes that a previous run could not synthesise get another attempt.
The analysis log is appended to rather than replaced, so you can read how the analysis changed across runs. For longitudinal work that log is itself a useful artefact.
Some practical tips
Spend your time in the framework editor, not the chat. The setup agent proposes on the first turn precisely so you can start editing rather than negotiating. Read the broader dynamics closely: they are the prompts that produce your analytical themes, and a vague dynamic produces a vague theme.
Keep first-level nodes to max seven. The limit is enforced, and analyses that push against it are usually asking two questions at once. Split them into two runs.
Answer the agent's questions in the rationale. They are specific and consequential, especially the segmentation ones. Answering takes one line of chat and the framework revises in place.
Run the light modes first on new material. An inductive pass tells you what is actually in the corpus and sharpens the question you bring to the agentic run. Going straight to agentic on material you have never looked at usually means the framework you confirm is the model's guess rather than your judgement, and the whole point of the mode is that you own that skeleton. The broader argument for keeping the human in that position is in responsible AI in qualitative market research and two-way transparency.
Frequently asked questions
How long does an agentic analysis take?
Longer than the other modes, because it makes several passes: screening, extraction, metadata enrichment, investigation per broader dynamic, counter-evidence, revision and writing. Elapsed time scales with corpus size and with how many first-level nodes and dynamics you confirmed. You do not need to wait on it; the run is queued and you are notified when it completes.
Can I edit the framework after the analysis has started?
You can change the research question and framework fields afterwards, but doing so will not affect insights already extracted. In practice, if the framework was wrong, the better move is to edit it and re-run, since re-runs preserve the coding that passed and redo the analytical layer.
Does agentic analysis replace inductive or predefined analysis?
No, and using it that way wastes time. Inductive is for exploring an unfamiliar corpus, predefined for applying a codebook you already trust, and agentic for answering a question with analytical depth. Most projects use more than one: an inductive pass to see what is there, then one or more agentic runs on the key research questions.
What happens if my documents do not really answer my question?
The setup agent is instructed to say so in its preamble and propose the framework the data can actually carry, rather than the one you asked for. During the run, the coverage reflection step examines what extraction actually produced and records gaps. Analytical themes that cannot be supported are dropped rather than written up thinly, and thin evidence is flagged in the evidence evaluation.
Can I trust the quotes in the report?
Quotes are verified against the source document at extraction time, and every reference in every artefact is re-checked at the end of the run. Beyond that, every citation in the report is clickable and opens the underlying insight, document or quote, so verification is a click rather than a search. That said, the standing advice applies: read the passages behind your headline findings before you present them.
Which analysis type should I pick if I am not sure?
Let the agent pick and then sanity-check it against the table above. The one distinction worth checking yourself is whether you want a description or a decision. Classification, Comparative and Process / journey describe. Prioritisation, Prescriptive / design and Evaluation decide. Picking a descriptive type when you needed a decision produces a well-organised analysis that does not tell you what to do.
Ready to ask your corpus a real question? Try Skimle for free, upload your documents, and run an agentic analysis. You will get coded evidence, analytical themes with counter-evidence, and a written answer where every claim links back to the quote behind it.
Related reading: See our guides on grounded theory methodology, how to code qualitative data, and presenting qualitative research findings to executives. For how Skimle compares to other tools, see the complete comparison of qualitative data analysis tools.
About the authors
Henri Schildt is a Professor of Strategy at Aalto University School of Business and co-founder of Skimle. He has published over a dozen peer-reviewed articles using qualitative methods, including work in Academy of Management Journal, Organisation Science, and Strategic Management Journal. His research focuses on organisational strategy, innovation, and qualitative methodology. Google Scholar profile
Olli Salo is a former Partner at McKinsey & Company where he spent 18 years helping clients understand the markets and themselves, develop winning strategies and improve their operating models. He has done over 1000 client interviews and published over 10 articles on McKinsey.com and beyond. LinkedIn profile
