Is Research and Insights being too cautious about AI?

At the Esomar Congress in September, we will host a red-teaming debate on how AI is being adopted by the research and insights industry.

The format will be simple, but I hope powerful. The blue team will argue that the industry is broadly heading in the right direction, and that our main problem is overcaution: too many blanket restrictions, too much waiting for certainty, and too much delay dressed up as diligence. The red team will argue for a change of course: that we should slow down in some areas, shift our focus in others, and ask harder questions before rushing ahead.

Crucially, both teams will be speaking from conviction. This will be a real argument, with people arguing the position they hold. We need to hear from people who believe we are moving too slowly, from those who believe we are being too careless, and from those who think the question itself needs reframing.

I have a professional and a personal interest in this argument. As Chair of Esomar's Professional Standards Committee, part of my job is to help the industry find the right balance of guidance, caution and experimentation. The standards we write, the guidance we publish and the questions we encourage buyers and practitioners to ask all depend on getting that balance right. Guidance that is too restrictive will push AI use into the shadows, where it is unmanaged and undisclosed. Guidance that is too permissive will erode the trust that makes research worth buying. Hearing the strongest versions of both arguments, made by people who believe them, is one of the best ways I know to test where the balance should sit.

As part of my preparation for the debate, I reached out to a wide cross-section of people across the industry. I contacted people from agencies, clients, platforms, technology companies, professional bodies and consultancies. I also started a discussion on LinkedIn.

The question I asked by email was this:

"At present, the Research and Insights industry is too cautious about the implementation of AI, and this is to the disadvantage of the industry and our clients."

I asked people to answer first with one of three options: Support, Oppose, or Something else. I then asked them to add a sentence or two explaining why.

On LinkedIn, I asked a related question:

"Do you think market research is rushing into AI too fast? If so, what would be the two or three things we should pay more attention to?"

This was an exercise in taking soundings before a debate, rather than a survey. It was informal, self-selecting and qualitative, so I am going to resist any temptation to say "X% of the industry thinks this" or "Y% thinks that". But the soundings were revealing.

The case for moving faster

There are plenty of people who think the research and insights industry is being too cautious about AI. Their argument, in most cases, is that we are in danger of using caution as a respectable word for delay.

One of the strongest arguments is that AI is already here. Clients are using it. Marketing teams are using it. Insight teams are using it, whether formally or informally. The wider business world is beginning to build AI into its processes, systems and expectations.

If research and insights wait until the dust has settled, it may find that the dust has settled somewhere else.

There is also the question of reach. Traditional research has always been relatively expensive, relatively slow and relatively specialised. This means that insight is often reserved for the larger decisions. AI creates the possibility of putting insight into many more decisions, including the long tail of smaller decisions where researchers are currently absent.

For some, this is the greatest prize. Doing the old work faster matters, but making evidence and customer understanding more available across the organisation matters more.

There is also a productivity argument. AI can already help with desk research, proposal writing, questionnaire development, interview summarisation, coding, reporting, knowledge management and presentation building. It can help juniors learn faster and seniors spend less time on the mechanical parts of the job. It can make existing knowledge easier to find and reuse.

From this viewpoint, the industry needs to be braver. We need to experiment more, learn faster and accept that some things will need to be improved as we go along. Waiting for 100% certainty has never been a good strategy in a fast-moving world.

The case for slowing down, or changing course

The other side of the argument is strong too. Indeed, it is essential.

Research and insights are paid for to help clients make better, evidence-based decisions. If AI produces something that sounds right but is wrong, the danger goes beyond an error in a document. The danger is a weakening of the link between evidence and decision-making.

Several people made the point that market researchers are hired to provide facts, not hallucinations. AI can be persuasive when it is wrong. It can be plausible when it is ungrounded. It can be fluent when it is thin.

This matters especially when AI starts to move from being an assistant to the researcher to becoming part of the evidential chain itself. There is a big difference between using AI to tidy notes and using AI to simulate respondents. There is a big difference between using AI to help draft a discussion guide and using AI to make claims about what people think, feel or will do.

The more AI affects data, evidence, inference or recommendations, the higher the bar must be.

There are also concerns about data quality, transparency, confidentiality, intellectual property, governance and the status of synthetic data. Some of the people I spoke to are worried that the industry is being dazzled by the speed and confidence of AI outputs without doing enough validation. Others worry that clients and buyers may not yet know the right questions to ask.

This is why the red team matters. Their job is to make sure the industry keeps innovation and carelessness clearly separated, and to argue that, in some areas, the right move is to slow down, and in others, to redirect our energy entirely.

The "something else" position

For me, the most interesting responses were often from people who refused the framing of the question.

Their point was that we are cautious in the wrong places.

We may be too slow to use AI where the risks are low and the benefits are clear. For example, we may be too slow to improve internal workflows, make existing knowledge searchable, reduce repetitive tasks, and support greater productivity.

At the same time, we may be too trusting, where the risks are much higher. For example, we may be asking too few questions about synthetic data, model outputs, automated interpretation, representativeness, bias, provenance and validation.

In other words, the real issue may be the quality of caution, rather than the amount.

Good caution asks: What is the use case? What is the risk? What evidence do we have? How will we validate the output? What human judgement is required? What should be disclosed to the client? What happens if the output is wrong?

Bad caution says either "no AI" or "AI everywhere".

That, I think, is one of the most important distinctions to come out of the exercise, and it maps neatly onto the debate itself. The blue team will draw on the first half of this position, the areas where we are too slow. The red team will draw on the second half, the areas where we are too trusting.

Some of the themes

A number of themes came through repeatedly.

Human judgement remains central. AI can draft, summarise, search, classify and suggest. But the value of research is still rooted in asking the right questions, recognising weak evidence, understanding context and knowing when something does not smell right.

Verification is becoming a key skill. We have spent years teaching researchers how to collect and analyse data. We now also need to teach them how to check AI outputs. Prompting is a starting point; the important skill is knowing whether the answer should be trusted.

Synthetic data is one of the fault lines. Some see it as one of the great opportunities. Others see it as an area where claims are outpacing the evidence.

The industry is uneven. Some agencies are moving very fast. Some clients are cautious. Some technology providers are making ambitious claims. Some procurement teams are applying blanket restrictions. There is no single industry position.

And perhaps most importantly, AI will require us to rethink the workflow, rather than just speed up the old one. Much of current AI use is about faster transcription, coding, reporting, and synthesis. That is useful. But the larger question is how organisations listen, learn, remember and act in an AI-enabled world.

What I learned from asking the question

When I started this exercise, I had hoped the proposition would divide the house. It did, but not cleanly enough.

There are people who think we are too cautious. There are people who think we are too careless. There are people who think we are moving too fast in some areas and too slowly in others. There are people who think the bigger issue is competence, not speed. And there are people who think the real question is whether research is shaping how AI is used, rather than whether AI is entering research.

In my own sense, based on the replies and discussion, the largest group is probably those who think research and insights should adopt AI faster, but not foolishly. They want more experimentation, more capability and more practical use. But they also want stronger validation, better governance and a clearer understanding of where human judgement remains essential.

That is a demand for a more mature conversation, rather than a simple "pro-AI" position. And it is exactly the kind of finding that shapes the PSC's work: it tells me the industry wants standards and guidance that enable good practice, rather than rules that simply say yes or no.

The motion matters

My conclusion from this exercise is that we need a better question for the Congress debate. The first question was useful because it flushed out the range of views. But for the debate itself, we need something sharper: a motion that gives the blue team something genuine to defend and the red team something genuine to attack, and that forces the audience to choose.

So watch out for the final motion, which we will publish ahead of Congress.

The session promises to be one of the most attended at Congress, and I think that is a good sign. The industry needs a debate where people who genuinely disagree can test each other's assumptions, rather than another bland panel about AI.

Watch out for news about the final wording of the debate.