Why do we need research when we have got AI?

Jane Frost, chief executive, MRS

About eight years ago, non-researchers were predicting the end of the research function. I vividly remember this quote by the chief strategy officer of a major London agency. He said to me, “We don’t use research, we just use Google”. Then he went quiet and after a moment reflected, “Maybe that’s why our insight isn’t very original”.

Big data appeared to be the only thing people cared about. No one asked whether there was any virtue in the adjective “big”, but people did think the more they had the better, the more they would understand their customer, the more problem solutions they would find.

Here’s the thing; big data is no use to anyone if it isn’t good data, contemporary data, data that suits the investigation or the problem you want to solve, data above all that has been analysed and converted into usable answers to our problems. The chasing after more and more data, without quality or purpose wasted a lot of time, consumed a lot of energy, and led us down blind allies that took investment that could be better applied elsewhere.

Think about it, the very famous saying from Avon; “In the factories we make cosmetics, to the public we sell hope” or Unilever’s “Domestos isn’t about bleach, it’s about protection” or Persil’s “It isn’t about clothes, it’s about Mother’s care”, which all imply we need to collect data that reflects emotions and psychological needs and not factory units. It was mostly the physical that got counted and turned into data.

So why is “artificial” intelligence a good thing? Why do we accept it as something everyone needs to sell, that we all need to have in our portfolio or we risk being dinosaurs? Why is it easier to get the CFO to invest in anything with the acronym, AI, than trained professionals? Why do we have to defend the real against the synthetic? Why the tendency to assume artificial is better than human?

Maybe it is a very human tendency to be awed by technology, to be fascinated by process, because it is easier to measure outputs than outcomes. Think about it, procurement departments are naturally more comfortable with units of ball bearings than with buying non-physical services.

The thing we should do is relegate AI to what it is, a process. A process of great potential in all of its formats when applied right. A process that can be good or poor quality. A process we need to run to deliver our objectives, rather than let it run us. A process we need to apply intentionally. A process we need to understand and professionally manage. We need to understand the quality of the data input, the way the LLM works needs to be transparent and the output needs to be verifiable.

Perhaps we need to avoid the acronym and call it what it is: “Artificial Intelligence”. AI somehow lets us get away with not considering that artificial anything has as many drawbacks as benefits. AI could be your best friend, so you don’t think too carefully about its weaknesses.

AI can be marvellous – scratch that, good AI can be really, really useful as part of our armoury. It can be faster, go wider and deeper. It may be cheaper at times (although I have yet seen a total cost to serve model for quality AI) but sometimes it’s no cheaper than just seeing people, watching their expressions, seeing them interact. No senior leader I know thinks we can substitute that and the thing is that can be cheaper and just as fast as you need it to be.

AI can’t speak better truth to power or give a voice to the marginalised as it stands because it is made up of our prejudices, our data weaknesses and importantly its output is defined by the way we frame our questions and our problems (sorry, “engineer the prompts”).

So let’s start with our purpose, be crystal clear about our outcomes and then apply the relevant methodology, whether that be an AI engine, use of synthetic data or an application of behavioural science, ethnographic study or focus group. Let’s select our methodology based on our need and embrace what each of them is good for.

That means we need to increase our skills to embrace AI as we do other research practices, we need to understand our cost models frameworks with confidence. We definitely need to write better briefs so we understand our required outcomes more clearly and get colleagues to share the vision and we need to be in charge.

As the Harvard mathematician and satirist Tom Lehrer once said: “what you get out of it depends on what you put into it”.

AI can run the process, but the thinking must remain ours

Marie Robelin, growth strategy & transformation CMI director, Unilever

If AI is, as Jane argues above, a process to be applied intentionally, the real question is how it is now reshaping the craft of research and whether we are at risk of letting the process run us.

AI is already rewriting how we work: generating briefs, surveys and discussion guides, moderating conversations, synthesising outputs, even producing synthetic respondents. It can “go wider and deeper”, as highlighted earlier. But its outputs are compelling not just because of what they say, but because of how they say it: fluent, confident, effortless.

And that effortlessness is precisely the problem. AI taps into cognitive ease, our brain’s preference for not having to think, amplified by authority and automation bias. In a discipline already tempted to prioritise outputs over outcomes, the risk is not just inefficiency, but intellectual passivity, researchers quietly ‘falling asleep at the wheel'.

This is why research matters more, not less. As Jane reminds us, AI is only as strong as the questions we ask and the data we feed it. It cannot define the problem, nor observe the nuance of what people do and feel beyond words.

AI surfaces patterns, but often the most frequent ones, and that is how you end up with safe, average, “beigified” insight. Real insight lives in tension, contradiction and minority voices.

• The full Delphi report, ‘Who owns understanding? How AI is reshaping value, expertise and accountability in research’ is accessible via the MRS website.