Introduction 

AI is no longer a “nice to have” in market research. It’s increasingly embedded in the tools people use every day, from shaping research materials to handling large volumes of qualitative data, and many research buyers now assume it will be part of the workflow somewhere.  

But the conversations we’re having with clients are not one-dimensional. Some teams are keen to move fast and test new approaches. Others are curious but cautious, and want clear reassurance on quality, governance and data protection. That range is exactly what we described in our first blog on AI in market research: interest is high, but concerns about accuracy, bias, repeatability and loss of nuance are right up there too.  

So this follow-up is a deep dive into AI in qualitative research, what’s changed recently, what the evidence is saying, and where AI is most useful for saving time, improving consistency and supporting better outputs (as well as where it’s risky, premature, or simply the wrong tool). 

In summary, here’s what we think: 

What’s changed in the last 12 months (and why it matters for qual) 

1) The industry has moved from “AI excitement” to “AI governance” 

Over the last year, professional guidance has become much more explicit about what “responsible AI” looks like in research: accountability, transparency, human oversight, data protection and confidentiality.  

In other words: using AI isn’t the differentiator anymore. Using it well is. 

2) The tooling is shifting from “features” to “workflows” 

Qual platforms have moved beyond “we do transcription” into broader workflows: AI search, summarisation, dashboards, and even AI “agents” that push regular updates or alerts. (This is part of the wider trend towards always on insight, not one-off projects.)  

That’s helpful, but it also increases the importance of governance, because AI is now touching more parts of the journey. 

3) The evidence on “quality uplift” is mixed, and that’s a good thing to be honest about 

Here’s the reality: the most robust recent research does not support a blanket claim that “AI improves qualitative analysis quality”. What it supports is a more useful (and more honest) statement:  

AI can improve speed and consistency very reliably. 
It can improve quality in some contexts, but only when it’s methodologically well setup and strongly governed by humans. 

A 2025 PLOS One study tested a closed GenAI system (Microsoft Copilot) for thematic analysis and found minimal overlap with human themes, frequent fabricated quotes, and a tendency to draw themes from early parts of the dataset rather than the whole thing. The authors concluded they could not recommend that version for undertaking thematic analysis.  

On the other hand, a 2025 paper (Ipsos UK + University of Southampton⁵) found two LLMs (Claude and GPTo1) conducted narrative analysis judged credible and thorough, comparable to human analysis, and even contributed additional insights that enhanced the human work.  

And on structured coding tasks, research suggests reliability can be strong when the schema is well defined and prompts are iteratively refined, including a study where GPT4o achieved intercoder reliability comparable to human coders in that context.  

So: AI is not “good” or “bad” for qual. It’s situational, and that’s where experience matters. 

The trade-off in qual isn’t new: speed vs depth 

Qual has always involved trade-offs: 

AI changes the shape of those trade-offs. It can remove a lot of manual labour, and it can help teams stay consistent across big datasets. But it can also flatten nuance, over standardise interpretation, or create a false sense of certainty if outputs aren’t verified (especially if the AI is drafting themes, implications or even quotes).  

Our view is that the best use of AI in qual is very specific: 

AI does the heavy lifting. Humans do the meaning. 

Our Essential Principles 

Before we talk about where AI helps, it’s worth being clear about our guardrails, because this is what most clients are really asking for. 

1) Human-led design, moderation, analysis and sign-off 

2) Consistency and quality controls 

3) Data protection and governance 

This “transparency and trust” approach aligns with the direction of the ICC/ESOMAR Code and MRS guidance, which emphasise human oversight, accountability and protecting privacy/confidentiality.  

4) No “one size fits all” 

Some AI use cases are fairly universal (e.g., transcription and rapid retrieval). But in practice, the best results come from matching the right tool to the right stage. 

We don’t rely on a single platform for everything. On some projects we may use multiple AI tools across the workflow (sometimes up to four), depending on the task, the security requirements and what the client wants to prioritise. We won’t name our AI tools, but we’ll clarify AI’s role. 

Where AI helps most in qual (and where to be careful) 

Here’s the practical table we follow… 

Stage	Where AI tends to help most	What humans must do	Watch-outs
Design (incl. discussion guides)	Drafting modules, probes, versions; stress-testing logic; simplifying language	Ensure fit to decisions, remove leading bias, tailor to sector sensitivities	Generic “best practice” guides that miss context
Fieldwork (depths & groups)	Optional support: structured note templates, probe suggestions	Build rapport, read nuance, follow the unexpected, safeguard participants	Don’t let AI drive the conversation
AI interviewers / bots	Semi-structured, scalable, time-poor audiences, low-stakes exploration	Decide when the trade-off is acceptable; triangulate with human qual	Weaker probing/rapport; not right for sensitive or deliberative work
Transcription	Speed and searchability; faster turnaround to analysis	Spot-check and correct errors; ensure speaker attribution	Accent/domain errors; confidentiality requirements
Analysis	Fast retrieval (“where does this topic come up?”); draft grids; consistency checks	Interpret, surface nuance, protect minority views, keep claims evidence-based	False certainty; “averaging out”
Reporting	Draft neutral summaries; restructure narratives; improve clarity	Validate every headline against evidence; set implications and limitations	Polished nonsense if inputs are weak

To summarise: in most projects, the biggest ROI comes from transcription + navigation + structured analysis support, not from replacing moderation. 

Stage-by-stage: what it looks like in practice 

1. Design: better materials, faster (with the researcher still in charge) 

AI is genuinely useful at the start of a qual programme: 

But the risk is real: AI can make things sound “right” while missing what actually matters, politics, legacy issues, regulatory sensitivities, or the emotional reality of a topic. That’s why we treat AI drafts as a first pass, then reshape them with sector expertise and moderation experience. 

2. Fieldwork: why our default is still human-led (and why deliberative work is different) 

AI-led interviewing is improving, and there are credible reasons it may be attractive in some situations: speed, scale, and standardisation. 

Some vendor research suggests that voice-based AI-moderated interviews can generate longer and more diverse responses than typed answers, and that participants rate different modalities similarly on ease and willingness to open up, while still often preferring typing for privacy and control.  

That’s interesting, and we’re watching the space closely. 

But our preference remains: humans conduct most depths and focus groups, because moderation quality is not an admin layer. It is the method. 

Human moderators: 

This is especially true in deliberative fieldwork (citizen-style panels, co-creation, consensus building). In deliberative work, facilitation is the engine of quality, not something you can automate without fundamentally changing what the method is doing. 

3. AI interviewers: where we’ve used them (and why it’s still early-stage for most clients) 

We have used AI interviewers/bots on some occasions, typically where: 

But even when we discuss it with clients, it hasn’t “taken off” universally yet. It’s still early for many organisations. 

4-6. Transcription and analysis: where AI is most consistently valuable 

This is where AI has made the biggest practical difference to qual delivery. 

Transcription first: 
Faster transcripts mean analysis starts earlier and teams can work more iteratively. But it only works if you treat transcripts as data that needs checking, not as a perfect record. 

Then analysis support: 
AI can be excellent at: 

Done well, this reduces the risk of “we missed that theme because it was buried in interview 17”. 

However, this is also where the watch-outs are sharpest. The PLOS One Copilot study is a useful cautionary tale: fabricated quotes and limited coverage are exactly why we don’t allow AI to generate unreviewed analysis or unreviewed outputs. In fact, when we first set out using AI in our own research processes, we encountered these same pitfalls firsthand. It didn’t take long to realise the importance of implementing robust review steps and clear checks, ensuring that every AI-generated insight is scrutinised by a human before being accepted or shared. By refining our approach early, we were able to avoid these issues and build more reliable, trustworthy workflows. 

From our own direct experience, we’ve found that the most effective use of AI in qualitative research comes when it is paired with robust human oversight and a carefully designed approach. We have seen first-hand that, when prompts and coding frameworks are thoughtfully refined, AI can deliver reliable support for structured coding and even contribute to credible narrative analysis. This aligns well with recent studies, which suggest that LLMs can meaningfully assist analysis, provided the method is sound and the outputs are always reviewed by humans. In our workflows, we deliberately integrate AI to boost coverage and consistency, while relying on human expertise to safeguard nuance, meaning and accountability. The key is not to replace human judgement, but to enhance it, using AI as a tool to increase efficiency and breadth, while ensuring that every insight is traceable and trustworthy. 

At the same time, recent studies show that with the right method and human oversight, LLMs can meaningfully assist analysis, including narrative analysis judged credible and thorough, and structured coding workflows where reliability can approach human levels when prompts and schemas are refined. 

So our practical rule is: 

The watch-outs we actively manage 

If you’re commissioning qual (or running it), these are the risks worth having on your radar: 

A practical checklist: what you should ask any research supplier about AI 

If you want a simple way to sense-check an agency’s AI approach, ESOMAR’s “20 Questions” checklist is a helpful reference point, it focuses on transparency, trust, human oversight and data governance.  

In summary: 

Those questions don’t slow projects down. They make them safer, and usually better. 

Conclusion: the best qual is still human – AI just helps you get there faster 

AI is raising expectations in qual. Clients want turnaround, clarity and consistency, but not at the expense of depth, nuance or trust. 

The most productive way to think about it is this: 

That’s how we’re choosing to use AI in qualitative research at Impact: not as a replacement for people, but as a set of tools that help our researchers deliver clearer, faster, high-quality insight you can trust, with governance built in. 

---

Want to talk about AI in your research?
If you’re exploring how to use AI in your insight work, or how to make what you’re already doing safer and more effective, we’d be happy to chat. We can review your current workflow, show where AI can genuinely help, and where human oversight is critical.