Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

AI in Marketing › AI Applications in Marketing

Using AI for Research Synthesis

How to use AI to synthesise interviews, open survey answers and reviews into themes that trace back to evidence, with consent, quote checks and bias controls.
Guideintermediate12 min readUpdated

What synthesis is and where AI fits

Synthesis is not summarising. A summary shortens what one source says; synthesis compares many sources and finds the patterns that answer a research question, with the evidence for each and the exceptions that qualify it. The usual sequence is the same whether the material is interviews, open text from a survey or a quarter's product reviews: read enough to become familiar with it, build a codeframe that names what to look for, apply the codes, group codes into themes, weigh the evidence behind each theme and write the implication for the decision the research was commissioned to inform.

Most of the time in that sequence goes into applying codes, which is repetitive, and into finding the passage that illustrates a point, which is searching. Those are the jobs a language model does quickly. The jobs at either end, deciding what to look for and deciding what it means, are judgement, and they are where a model's fluency is most misleading, because a confident paragraph about "what customers want" reads the same whether it rests on a dozen consistent excerpts or on one articulate participant.

StageWhat a model can doWhat stays with the researcherWhat to check
FamiliarisationSummarise each source against the discussion guide or question listReading a sample of sources in fullSummaries that drop hedges, contradictions and tone
CodeframePropose candidate codes from a sampleChoosing, defining and bounding each codeCodes that overlap or mirror the prompt's wording
CodingApply the codeframe passage by passage, with excerptsRuling on ambiguous passagesA sample of coded passages against the definitions
ThemesDraft groupings of related codesDeciding which themes answer the questionThemes that rest on one or two voices
EvidencePull candidate supporting and contradicting excerptsChoosing the excerpts that go in the reportEvery quote word for word against the source
ImplicationsDraft wording for a findingThe inference and the recommendationClaims that go beyond what participants said

The AI applications hub judges any use of AI on value, review cost and risk. Synthesis scores well on value and badly on review cost unless the work is structured so that every output can be checked against a source quickly, which is what the rest of this guide sets up. Numerical material, such as dashboards, query results and report tables, has its own checks, which the AI for analysis and reporting guide covers; this guide deals with words. One common destination for those words is a web page, and the tutorial on turning customer language into page copy applies this guide's redaction and coding rules to reviews, tickets and chats before the phrases reach a rewrite.

Before anything goes into a tool, check what participants were told. The professional codes published by ESOMAR and the Market Research Society, as the site's research guides describe them, rest on voluntary and informed participation, on telling people who is collecting their data and why, and on keeping answers confidential unless participants agree otherwise. If the consent wording for an interview study said the recordings would be seen by the research team, sending transcripts to an outside AI service is a use the participants were not told about. The fix for future studies is to say so in the consent notice; for existing material, the decision belongs to whoever owns the study's data protection, not to the analyst with a deadline. The tutorial on participant information and consent lists every use of what participants say, AI tools included, before the first invitation goes out.

Data protection law treats a transcript or a free-text answer as personal data whenever it identifies someone directly or through what it mentions, and passing it to a vendor adds a processor to the chain. The principles that apply are the familiar ones: a lawful basis, a purpose compatible with the original collection, and no more data than the task needs. The disclosure and data privacy guide sets out which tools are appropriate for personal data and what their terms must say; this guide assumes an approved tool.

Redaction comes next. Replace names with participant codes, and remove contact details, account or order numbers, employer names where they identify a person, addresses and anything sensitive that the research question does not need. Keep the key that links codes to people in a separate, access-controlled place, never in the material the model sees. Read a sample of the redacted files by eye, because names turn up in odd places: a participant quoting a colleague, a support ticket signed off by a customer, a review that mentions a member of staff.

Then give the material a consistent structure, because a model codes structured material far more reliably than a heap of pasted text. One file or block per source, each with a header carrying the participant code, the Segmentation fields the analysis will compare (customer type, tenure band, market), the date and the source type. Speaker labels on every turn of a transcript. Survey answers kept with the question they answer. Long transcripts split at the topic boundaries of the discussion guide, so each request covers a manageable section and every excerpt can be located by source and section.

Coding themes with a model

Build the codeframe before the model touches the full set. The survey and questionnaire design guide describes the standard approach for open text: build a codeframe from a sample, then apply it to the rest. The same holds here. A person reads a sample of the material and drafts the codes; a model can propose candidates from the same sample, which is a useful prompt for what the person might have missed, but the person chooses, merges and defines them. Each code gets a short definition, a line on what it excludes and an example excerpt. Where the volume is small enough to code without a model, the tutorial on coding open-text answers by hand applies the same frame with a second coder's check. A jobs-to-be-done frame gives a sound starting structure for purchase research: the trigger that started the search, the alternatives considered, the anxieties that slowed the decision and the outcome the customer was hiring the product for.

Then instruct the model to apply the codeframe and return a table rather than prose. For each passage it codes, it returns the source code and section, the code applied, the excerpt exactly as it appears, and a flag where the passage is ambiguous. Two instructions matter more than the rest. It may only use codes from the frame, with "uncoded" as an allowed answer, so that it does not invent categories on the fly. And it must copy excerpts verbatim, never paraphrase, so that every excerpt can be found again by searching the source.

Run the same instructions over every batch of material and keep them fixed for the whole study. Changing the wording halfway through changes what the model codes, and the resulting counts are no longer comparable. When the uncoded pile grows, that is a signal the codeframe is missing something: a person reads the uncoded passages, adds or redefines codes, and the affected batches are coded again. An inductive pass, asking the model what patterns it sees that the frame does not cover, is worth running once at the end as a check on the frame, with every suggestion verified against the material before it becomes a code.

Tracing every theme to evidence

A theme is only as good as the evidence a reader can inspect, so the output of synthesis should be an evidence table before it is a slide. Each row carries the theme, its definition, the participants it comes from, the supporting excerpts with their source codes, the contradicting excerpts, and the implication. The table is built from the coded output, not from a model's summary of it, and it is the document reviewers check.

The customer interviews tutorial asks for a separation between what participants did, what they said and what the team infers, and that separation matters more when a model is involved, because model-written summaries blend the three into one smooth sentence. "Customers churn because onboarding is confusing" might rest on three participants describing a confusing first week, on one participant saying so outright, or on the model's inference from complaints about something else. The evidence table forces the distinction: behaviour and statements are excerpts; inference is a separate column in the researcher's words.

Two rules keep themes honest. A theme that rests on a single participant is reported as a single voice, which can be valuable, but not as a pattern. And every theme records the counter-evidence the coding found, because a model asked for supporting quotes will find them for almost anything. Themes that pass both rules can update a Buyer Persona or a segment description with confidence; themes that fail them become questions for the next round of research.

Checking quotes and counts

Quotes are where model-assisted synthesis most often goes wrong in public. Models tidy grammar, merge two speakers into one sentence, shorten a long answer into something more quotable, and occasionally produce a perfect quote that nobody said. Each of these reads naturally in a report, and the first two can pass a casual check because the meaning is close. The rule is that no quote leaves the research team until it has been found, word for word, in the source by its participant code and section. Structured, redacted material makes this a quick search rather than a hunt.

Counts need the same discipline. A model asked how many participants mentioned price will give a number, and over long material the number is unreliable: it double-counts a participant who raised the point twice, misses mentions phrased indirectly and sometimes reports a figure that matches nothing. Take counts from the coded table, by counting distinct participant codes per theme in a spreadsheet, never from the model's prose. Then check the coding itself: a person takes a random selection of coded passages, codes them blind against the definitions and compares. Where the two disagree often on one code, the definition is unclear and gets rewritten before the affected material is coded again.

Remember what counts mean in qualitative work. They describe breadth within the people you spoke to, not prevalence in the market. A theme mentioned by most interviewees is worth taking seriously; it is not evidence of how common the view is among customers, which needs a survey with an adequate Sample Size and a representative sample. The accuracy and review workflows guide treats quotes and figures as facts that must be traced to a source before publication, and research reports belong in that tier.

Illustrative example — not a real result

Synthesising cancellation interviews at a meal-kit subscription

Quillfern Kitchen, an invented meal-kit subscription, interviews 18 customers who cancelled in the previous quarter. Transcripts are redacted to participant codes P01 to P18, split into 4 sections following the discussion guide, and tagged with tenure band and household size. The researcher reads 5 transcripts in full and drafts a codeframe of 11 codes; the model's suggestions from the same sample add 2 more, one of which ("delivery window too narrow") survives. The model codes all 72 sections and returns 214 coded excerpts, including 19 marked uncoded. A blind check of 30 excerpts agrees with the model on 24; most disagreements sit on two overlapping codes about price, which are merged and re-run. The model's summary had said "most customers left because of price"; the coded table shows price raised by 7 of 18 participants, and for 5 of them alongside a household change. The quote check finds 3 of the 12 quotes chosen for the report were tidied paraphrases; they are replaced with exact wording. The finding that reaches the team is "price mattered most when the household changed", with the excerpts beside it.

Bias and what the model cannot see

A model brings its own tilt to synthesis, and it compounds whatever bias the research already had. It favours fluent, articulate speech, so the participant who explains themselves well is over-represented in summaries and excerpts. It gravitates towards the majority view and smooths over outliers, which in customer research are often the most useful passages. It drops hedges, irony and hesitation, so "I suppose it was fine, really" can come back as satisfaction. And it echoes the framing of the prompt: ask for reasons customers love the onboarding and it will find them.

It also cannot see what was not transcribed: the pause before an answer, the laugh, the topic the participant steered away from, the moment the interviewer led the witness. Those are often where the insight sits, and they are available only to someone who was in the session or listens to the recording. Nor can a model correct for recruitment. If the study only reached customers who answer emails, synthesis of any quality describes those customers, and comparisons across Segmentation groups inherit every gap in who was recruited.

The controls are simple and they are the researcher's. Ask the model explicitly for contradicting evidence for each theme, as well as supporting evidence. Read a sample of sources in full, including the ones the model quoted least. Have a second person review the evidence table before the findings are written. Report the method and its limits alongside the findings, as the professional codes expect, including the fact that a model assisted with coding and how its output was checked. For ongoing feedback programmes, the voice of the customer guide explains how coded open text feeds back to the teams that can act on it; a model can keep that coding current between reviews, provided the codeframe and the checks above stay fixed.

Common mistakes

  • Pasting raw transcripts into an unapproved tool. Consent, redaction and the tool's terms come before the first prompt, not after the first finding.
  • Letting the model build the codeframe alone. Codes it invents from the whole set tend to mirror the prompt and blur together; a person defines and bounds them.
  • Asking for prose instead of a coded table. A summary cannot be checked; a table of excerpts with source codes can.
  • Publishing quotes without finding them in the source. Tidied paraphrases and merged speakers read as real quotes until a participant sees one.
  • Taking counts from the model's answer. Count distinct participants per theme from the coded table.
  • Reading interview counts as market prevalence. Breadth among interviewees is not a share of customers.
  • Reporting the majority theme and dropping the dissent. Ask for contradicting excerpts and keep them in the evidence table.
  • Changing the instructions halfway through. Fixed instructions keep codes comparable across the whole study.

Frequently asked questions

Can AI replace a researcher in qualitative analysis?

No. A model can apply a codeframe quickly, pull candidate quotes and draft summaries, which removes much of the mechanical effort. Deciding which themes matter for the decision, weighing thin evidence against strong, noticing what participants avoided saying and standing behind the conclusions remain the researcher's job, and the setup in this guide exists to keep those judgements with a person.

Is it safe to upload interview transcripts to an AI tool?

Only to a tool your organisation has approved for personal data, and only after checking that what participants agreed to covers that use. Transcripts usually identify people directly or through the details they mention, so redact names, contact details and other identifiers first and replace them with participant codes. Professional research codes and data protection law both expect participants to know how their answers will be handled.

How do I know the themes are real and not the model's invention?

Make every theme carry its evidence: the excerpts it rests on, each with a participant code you can look up, and a count taken from the coded table rather than from the model's prose. Then read a sample of the source material yourself and check a sample of the coding against the codeframe. A theme that cannot show excerpts from the people it claims to describe is a hypothesis, not a finding.

Terms used on this page

Frameworks

Where this page is used

Sources

  1. ESOMAR — ESOMAR
  2. Market Research Society — MRS
  3. Regulation (EU) 2016/679 (GDPR) — EUR-Lex

Related