Back in early 2025, there were thought leadership reports aplenty on the topic of AI, but almost none actually featured AI as part of its research approach. Nearly all studies published at that time were based on a survey. 

There have been a few notable exceptions since then, such as the Guinndex and its AI calling agent, as well as some others trying out some novel research methods. These were stand-out examples, but were essentially outliers: the majority of studies remained survey-based. 

It’s too soon to say that this norm has definitively shifted, but the tide does seem to be turning. An updated review of recent thought leadership publications reveals a wider range of data inputs, and novel uses of AI to help analyse the data. 

Prompt research 

The most common approach is the use of large language models as classifiers: reading or labelling large volumes of text or behavioural data that would be impractical for human researchers to work through. A range of recent studies highlighted different things that can be done with this kind of approach:

• The Swiss Re Institute and the LSE used an LLM to read the 10-K filings of 91 Fortune 100 companies and mapped the causal links each company draws between its own risks. The result is a risk network that can be compared between 2019 and 2026. The LLM classified each risk disclosure into a category and then identified causal links to other risk categories (as mentioned in companies’ risk disclosures themselves). The risk that grew the most over that time period? AI and new technology risk. 
• Revelio Labs' AI Labor Market Tracker draws on people’s profiles, job postings, reviews and layoff notices to provide a monthly update. Its methodology is somewhat complex, but an AI-based classifier flags firms as AI adopters based on their hiring, and an LLM scores work activities for AI exposure, with those scores checked against established academic measures. In other words, it aims to infer adoption from what firms actually do, rather than what they say in a survey. 
• GitHub's Octoverse 2026 is perhaps the most technically ambitious. It is based on analysis of the company’s in-house data over the past year. It’s particularly fascinating, given how central GitHub is to much of the AI-based code creation work going on. Some of its headlines highlight how rapidly things are changing: for example, developers pushed code changes to Github more in the past year than they did in the prior two years combined. In this study, AI does three main jobs: an LLM extracts contributor locations (reweighted against IP data), an AI model sorts public repositories into industry sectors, and a set of forecasting models aim to predict developer numbers out to 2031. There’s plenty more, but the point is that the sheer volume of data being analysed requires an AI-led approach in order to be feasible. 
• Anthropic's Economic Index opens the window on when, where and what people are using Claude for, based on a sample of conversations with its Claude models (see below). These conversations are labelled by AI classifiers, and maps them to occupations and wage data. To its credit, Anthropic also shares a lot of detail on its methodology here (including specific prompts, which would be of interest for anyone working in the field). The company also links around 9,700 survey respondents to their own usage, so stated views sit alongside observed behaviour for the same people. 
• Finally, while the methodology doesn’t explain how the analysis was done, inferring skills from job titles across a billion profiles is almost certainly done with machine learning or an LLM. This is from LinkedIn’s Green Skills Report, now the sixth edition of this study, which shows that employer demand for green talent continues to grow faster than workforce supply.
  
  

Not AI, but novel nonetheless   

Not every example relies on AI. Three other recent reports stand out from the crowd as a result of turning data that a company already holds into a research asset that can provide real insight. While AI wasn’t necessarily used within the research, they’re nonetheless great examples of companies making use of non-survey data to deliver fresh insights: 

• The Allianz Center for Technology (AZT) published Swipe. Tap. Crash., an impactful new report highlighting the degree to which digital distraction is leading to increased dangers when driving. It uses data from over 34,000 crashes, in car eye-tracking, and other inputs to estimate that distraction may contribute to one in three road crashes today.  
• JPMorganChase Institute used de-identified transaction data from more than 20 million customers, covering 2015 to 2026, to track how often households are drawing on investment wealth to fund their spending. It's a behavioural measure that a survey would struggle to capture reliably, and the resulting insights are accordingly strong. 

• Consultancy HKA's CRUX Insight report, now in its ninth year, draws on the firm's own claims and dispute investigations across 2,410 projects in 122 countries. It is an excellent example for any professional services firm sitting on years of engagement data: casework can become a long-running proprietary dataset in its own right.

It would be wrong to suggest the survey is on its way out. Another recent report from the past month, AXA's Future Risks Report, is now in its 13th edition, polling 4,345 experts in 62 countries alongside 24,000 members of the public in 19 countries. Its value comes from scale, consistency over time, and the ability to compare expert and public views. (Fittingly, AI has now overtaken climate change as the experts' top-ranked risk.) 

More generally, surveys remain the norm for most corporate thought leadership research today. But it is intriguing to finally start to see a more meaningful batch of studies coming to market that are using AI in new ways. 

Embedding your own insights 

For thought leadership teams considering these methods, the examples highlighted above suggest three main pointers:

1. Start with the data you already hold. Transaction records, platform usage and casework are often more distinctive than anything a survey can produce, and are by definition proprietary. 
2. Use AI to read what humans can't. Filings, job postings, reviews and transcripts are rich sources that were previously too large to analyse in full with a human team.
3. Validate, and say how. The more credible studies check their AI-generated outputs against a benchmark and document the process. That transparency is also what makes the findings defensible with journalists and sceptical readers.

Two years ago, the absence of AI-based research in thought leadership was more notable than its presence. That is no longer quite the case. It remains a minority approach, but it's a growing one, and the range of methods is widening.