Applying Research Thinking to AI Adoption

When shoe company Allbirds announced it was pivoting to AI, it looked to many like a business chasing a trend rather than solving a problem. But it’s unlikely to be the last ‘AI-pivot' we'll see. That’s because boards are increasingly asking businesses to become "AI-first”. And while AI will undoubtedly transform how businesses operate, there’s a risk of AI projects ending in failure by treating the technology as the starting point instead of asking what problems need solving.

Market research isn't immune from the pressure to adopt AI, but it’s also one of the industries best equipped to avoid the mistakes by applying the same disciplined thinking it already brings to research design.

Building Basics

You wouldn't start a survey before agreeing on the methodology, sampling approach, and questionnaire. AI deserves the same rigor, starting with the data it uses.

Large language models are often trained on broad public information, not your company's history, customers, or ways of working. Without that context, AI can only produce generic outputs. Closing that gap goes beyond pasting documents and files into an AI chatbot. It requires building a harness between an AI model and your governed data, giving it a defined map of what that information means, so it can understand your business the way you do. That map, often called a semantic layer, can transform AI into a tool that feels custom-made for your business. The more you add trusted internal knowledge and structured data to the mix, the better the result will be.

A 2025 Gartner survey of infrastructure & operations (I&O) leaders found that 38% attributed AI project failure to poor data quality, and Gartner found that returns depend less on the sophistication of the model than how well it is integrated and governed. Experimenting with general off-the-shelf LLMs is a good place to start, but real value will come from engineering the structure around the models so it can leverage your organization's institutional knowledge safely and consistently.

Set Objectives

Your research project will have a list of KPIs or objectives attached, and so should any AI initiative. But instead of asking “where can we use AI?”, businesses should start with the business problems they are trying to solve. Look for the time-intensive work that could be automated: summarizing interview transcripts, identifying emerging themes, speeding up desk research, or surfacing anomalies.

Our partnership with market research and brand intelligence firm Basis Global is a good example of that in practice. The industry has long treated the tradeoff between depth, speed, and practicality as unavoidable. Brand tracking data had grown too large for any team to explore by hand. Emerging AI solutions in the space focus on getting clients faster reports without necessarily getting better answers, leaving them to make decisions from a partial view of the data. Basis Global saw an opportunity to move beyond automating the existing process, leaning into AI’s ability to test hundreds of hypotheses across markets, audiences, and time periods, surfacing patterns a manual review could never reach. Every insight AI surfaces is checked against the underlying data before it goes near a client, and researchers stay in charge of shaping the plan and interpreting the findings. What clients get is a more complete picture of their brand, backed by evidence and delivered as guidance they can act on.

Measure the Right Metrics 

Measurement is arguably what market research professionals do best. But when it comes to AI, token usage is often used as a proxy for measuring AI adoption, even though it tells you nothing about the business impact.

So what are the most useful metrics? They’re the ones businesses already care about: increased quality and consistency, improved client delivery, and better client decisions from faster or deeper insights. These outcomes are harder to measure than token counts, but they're what determine whether AI is creating value.

That focus on outcomes is supported by Gartner, which found that by the end of 2025, at least half of generative AI projects had been abandoned after proof of concept, with poor data quality, inadequate risk controls, escalating costs, or unclear business value among the causes. Tracking business outcomes from the outset makes it much easier to demonstrate whether AI is genuinely improving research rather than simply increasing usage.

Validate Before You Trust

One lesson from my early days of IBM Watson was how much effort went into ensuring the system only answered questions when it was sufficiently confident it had the right answer. In many cases, it was better not to answer than to risk confidently giving the wrong one.

Today's large language models are designed differently. They're built to be helpful, which means they'll often generate an answer even when the underlying evidence is incomplete or uncertain.

There are a couple of steps that can be taken to mitigate this. If AI is coding open-ended responses, compare a sample against human coding. If it's summarizing interviews, check the outputs against the transcripts. If it's identifying patterns, go back to the underlying data. Researchers have always validated new methods before trusting the results, and AI deserves that same discipline.

Refine and Revisit

Adapting to market changes means revisiting the research process and AI requires the same mindset. Models evolve rapidly and behavior  can change with new releases. It’s not uncommon to see models not performing as they once did, when a newer version is released. Organizations that remain model-agnostic and continually validate outputs are far less likely to find themselves using a system that is generating bad results. That’s why treating AI as a one-off implementation rather than an ongoing program is one of the fastest ways for performance to deteriorate, and for projects once heralded as successes to turn into failures.

Old Playbook vs New Technology 

The market research industry doesn't need to invent a new playbook for AI. It already has one. Researchers know how to define objectives, design robust methodologies, validate results, and measure outcomes. Bringing that same process to AI workflows is the difference between being an ‘AI-first’ business on paper and a truly AI-enabled business in practice.