Where are we on AI and sell-side research?

I don’t have the answer to this question, but AI is now washing over my industry like a tsunami, forcing incumbents either to integrate it into their products and processes or, where they’re not doing so, to figure out how to bolster and optimise existing products against what will invariably be a rising tide of AI-generated and automated content of increasingly high quality.

AI is a unique technology in the sense that it comes with powerful non-linearities. Used correctly, it can increase your productivity by orders of magnitude and improve your product. Used poorly, it can sloppify an otherwise unique and valuable brand and product.

I should be honest about my bias here. On a spectrum between those who see AI as a technology that overpromises and will eventually underdeliver, and those—like Marc Andreessen and DHH—who see it as a transformative technology with enormous promise, I am firmly in the latter camp. All that really means, however, is that I am choosing to lean into the technology for two reasons. First, because I genuinely think it is cool. Second, because whether I like it or not, it will transform the industry in which I work. I can’t afford not to understand and use it.

As a metaphor, I think of the world before AI as me driving down the motorway in a VW Golf. As AI comes along, I am busy upgrading my little Golf with new tyres, a bigger turbo and whatever else I can bolt onto it. That feels about right for an incumbent in an industry being forced to respond to the emergence of a new technology. But I am also looking over my shoulder for the start-up in my space that was built around AI from day one, and which I fear will soon zoom past me in a Ferrari.

So, how do I get a Ferrari?

Before we get to that, let’s go through the areas of my industry and work-flow where AI is currently encroaching itself.

Copy-editing/writing — This is the lowest-hanging fruit for AI users, but it is uniquely relevant—and scary—for people in my industry because, let’s be honest, we write PDF documents for a living. AI has been good at standard editing for a while, and it is getting better still. This applies to basic spelling and grammar as well as more contextual editing of flow, presentation of arguments and overall structure. As you move towards the latter, however, you have to be careful that the AI doesn’t blunt the edges of your argument, voice and ultimate message.

That said, the newest models are now so good that I would bet they could improve just about any research report sitting unedited after a first write-up by a human. The elephant in the room, then, is why we wouldn’t simply get the AI to write the whole thing from scratch.

We could. But we also operate in an industry where context—specific data, models, past views, current positions and accumulated domain knowledge—matters enormously. Asking an AI to write market research without that context will almost certainly produce the kind of “slop” you don’t want to send to a client. To be clear, AI can now produce reams of perfectly respectable-looking market research from a simple prompt. At first glance, much of it will look good. But look closely enough and it will often let you down.

If you want the AI to generate worthwhile research, you therefore have to constrain it with context: the data it should use, the relevant timeframe, your previous work, the parameters of the output and, ultimately, what you actually want to say. At that point, however, we are moving beyond AI as a writing tool and further up the taxonomy of AI usage.

Structured and unstructured research — Before AI, 95% of all research started with a Google search—or, in my industry, a search on a platform such as Bloomberg—followed by the manual collection and parsing of information into a coherent body of research. It could take a while.

AI can now do much of this for you in minutes, and you can have a conversation with it about the topic while it does so. Anyone who is not leveraging this capability in my industry is being left behind. This is true both for unstructured research, which I define as research that starts with a relatively simple or exploratory prompt, and for structured research, where a domain expert knows precisely what she is looking for and can direct the AI accordingly.

The most common objection is that AI will hallucinate, producing data, claims or sources that either don’t exist or are wrong. This is a real risk, but it needs to be weighed against the enormous reduction in the cost and time involved in discovering and synthesising information. The answer is not to trust the AI blindly; it is to retain your critical judgement and verify the claims and sources that matter. Don’t outsource your critical sense to the machine.

“Deep Research” is probably the clearest example of how AI has already changed exploratory research on the internet. If you remain unconvinced, try it. Pick a topic you know well, run a detailed Deep Research prompt, inspect the result critically and then ask yourself how long it would have taken you to assemble the same body of information in a pre-AI world. That comparison, more than almost anything else, illustrates the productivity shock now hitting knowledge work.

Coding — AI translates natural-language prompts into executable code, which is revolutionary. For sell-side research, however, it is a power that needs to be used with some caution.

Some market research already relies heavily on code for its output and product offering. Here, AI offers enormous productivity gains in writing, maintaining, debugging and improving existing code. For other research providers, however, the allure of being able to vibe-code almost anything in a day will inevitably lead to people creating suboptimal code to solve problems that never needed coding in the first place.

For the domain expert whose creativity was previously constrained by a lack of coding skills, however, the opportunity is much bigger. AI can now generate code for everything from simple analytical tasks and data visualisations to entire front- and back-end systems. It dramatically reduces the technical barrier between having an idea and being able to build something that tests or implements it.

We are all coders now, even if some of us still aren’t very good ones.

Data/content analysis — AI has long been able to interact with user-generated data and content: analysing it, refining it and creating new outputs from the underlying information. For market and economic research, the most relevant capability is its ability to work directly with quantitative data, either by connecting to a data API or, more simply, by reading files supplied by the user.

AI can now ingest large datasets, run quantitative analysis and produce outputs across almost every market and economic research domain. It is quick, but perhaps more importantly, the ability to explore data through a natural-language conversation creates the prospect of an AI research assistant that is always available, never gets tired and can move through many analytical tasks faster than most humans.

Here, however, the hallucination problem becomes more serious. If I spend an hour analysing a dataset with AI, but then have to reproduce the entire analysis independently to establish that the results are correct, much of the productivity gain disappears. Trust therefore becomes an important constraint on adoption.

For my part, I think the newest models are getting very good at quantitative analysis, particularly when the underlying data and methodology are clearly specified. But I concede that this remains something of a black box. For researchers producing a large volume of quantitative work, the question is not simply whether AI can generate an answer. It is whether the workflow can be structured so that the answer is sufficiently transparent, reproducible and verifiable that we can actually use it.

Rise of the agents — Agentic AI is the most recent iteration in the AI stack and introduces autonomous AI applications that can act independently across your IT systems and software according to the instructions you give them.

This capability is currently presented to users in a number of overlapping and sometimes contradictory ways, but the basic structure of an AI agent is relatively simple. First, you give the AI permission to act—either across an entire computer or within specific applications and environments. On ChatGPT, for example, I can now give the AI access to a range of applications and instruct it to carry out tasks within them. Second, you define the task.

That sounds trivial, but as the models become more capable, it is anything but. You can give an agent extremely narrow instructions—go to this folder, read this file, run this code, save the output here—or define a broader objective and give it considerable freedom over how to achieve it. Both approaches can be effective. The latter, unsurprisingly, also creates considerably more scope for costly accidents.

On ChatGPT, I currently have the ability to schedule tasks which, combined with selected application integrations, can already operate as a relatively simple form of agent. On my work subscription I have access to a more sophisticated agent environment, where I can provide detailed instructions, upload documents, add skills and give the agent access to specific applications and data. Beyond that are systems such as OpenClaw, where entire computers can effectively be operated as agents.

The implementations differ, but the direction of travel is clear: we are moving from asking AI to help us perform individual tasks towards giving it an objective and allowing it to execute the workflow required to achieve it.

For market research, agentic AI creates the possibility of automating specific tasks within the research process, but also, increasingly, of automating the entire process itself. We are already seeing a surge in A-to-Z research products in which an agent collects the data, performs the analysis, writes the research, produces the charts and ships the final output.

The success of these products, in my view, depends crucially on the scope and instructions given to the agent, the quality and uniqueness of the information available to it, and how the workflow and output are maintained over time. My own admittedly limited experience is that getting the back-end right matters enormously. You need to specify what the agent can access, what it should do, what the output should look like and where that output should ultimately go.

But the bigger implication is economic. Agentic AI will inevitably commodify some existing research products. If a research product can be defined as a sufficiently predictable sequence of inputs, analytical steps and outputs, there is a good chance that much of its production can eventually be automated. Research firms will then have to respond either by shifting their offering further up the value chain or by producing much more at a lower marginal cost.

It is a brave new world, but an old rule of business still applies: if you can easily create something with an agent, others probably can too. The interesting question, then, is where the durable value sits. How much advantage can be captured by having a better agent than your competitor, with better instructions, proprietary data, stronger domain knowledge and a more streamlined workflow? And how quickly will those advantages themselves be competed away?

About that Ferrari

I don’t think the answer is simply to generate more research with AI. If anything, the explosion in AI-generated content will make generic research cheaper and less valuable. The Ferrari, I suspect, is a research operation built around the combination of human domain expertise and AI-native workflows: proprietary or difficult-to-replicate data, accumulated knowledge and views, good models, clear instructions, judgement about what matters, and automated systems that allow all of those things to be deployed at a scale that previously required a much larger team.

This will require significant change for incumbents in my industry.

I am therefore excited about, and terrified of, AI at the same time. The technology will make my industry look almost unrecognisable over the next few years, creating real innovation and value for those who can leverage it, while those who can’t risk drowning in a sea of AI slop or simply being left behind as they refuse to adapt.

I intend to do this work for some time still, so I’d better figure out how to drive the Ferrari.