During a 1948 Princeton talk on computers, as recalled by E.T. Jaynes, mathematician and polymath John von Neumann was challenged: “[A] mere machine can’t really think, can it?”
His response: “You insist that there is something a machine cannot do. If you tell me precisely what it is a machine cannot do, then I can always make a machine which will do just that.” Von Neumann’s point was that of specification - the more clearly a process is specified, the more able a computer can be created to replicate that process… even in the realms of logic, reasoning and thought. For him, specificity was the way to solve for thinking machines - specifying the steps and the process.
Similarly the more specifically we describe a problem, the more clearly we’re able to solve it.
Alan Turing’s work on machine learning in the 1950s paved the way for advances in the 1980s and 1990s. The financial sector has ensured it was at the forefront of implementing these breakthrough technologies.
With the public introduction of generative AI (GenAI), we’re in the midst of a paradigm shift.
There’s no changing that the genie is well and truly out of the bottle. There’s no going back. The adoption rate of AI has been meteoric across all sectors of the economy.
In US survey measures, consumer adoption of GenAI has been faster than that of the PC or the internet at the same stage of their product cycles. ChatGPT reached 100 million users in its first two months (UBS called it the fastest-growing consumer application in history).
By early 2026 44% of US adults had used ChatGPT and 49% had used “any chatbot”. Within 3 years, ChatGPT grew to ~800 million weekly active users. For comparison, it took the internet on the order of ~13 years to reach that point. In October 2025, Sam Altman said about one in ten people worldwide had used ChatGPT.
“This AI trade has been a big one. We spend a lot of time looking at where all of the capex investments are going, where energy is going to be affected and other commodities, and I think we're just trying to figure out what some of those knock-on effects are going to look like over the next couple of years and actually how all of this innovation in AI may ultimately affect the economy.” - Melissa Otto, Global Head of Visible Alpha Research, S&P Global
But what about the tangible impact?In McKinsey’s global survey, ‘The state of AI in 2026: On the road to ROI’, 80% of respondents said AI had improved their individual productivity, and 50% said it helped them make better decisions.
Microsoft’s Agentic AI for Financial Service Leader, Martin Moeller in comments on LinkedIn about OpenAI’s Codex reports AI agents are now working longer hours on harder tasks, shifting the boundary of human-AI collaboration. The fastest-growing users being knowledge workers across various functions, including finance. He says those who adopt agents are using those agents extensive.
This is great for productivity gains and decreasing the timescales for work to be completed in but what about the impact on the bottom line?
We are attempting, like everyone else, to make headway in the AI world, which is again challenging because our information has to be top notch, it has to be vetted. [...] Not only accuracy, of course, but also to make sure you're still maximizing the involvement of your human intellectual capital to make sure what you're providing is timely and relevant.” - Jeff Dunsavage, Head of Research Publications and Insights, Insurance Information Institute
Where is the ROI?
McKinsey’s survey is revealing: Only 37% of businesses attributed even some EBIT impact to AI (unchanged from 2025) and only about 6% qualified as what McKinsey terms ‘high performers’ with a significant impact of at least 5% of EBIT thanks to AI.
“If you have the experience and you know how to use the tools, people's teams' productivity has increased very dramatically. That's something very exciting. But at the same time we also want to watch closely how that will transform everything.” - Yingjin Gan, Head of Index Research, Bloomberg
Of the respondents 89% use AI in at least one business function with 44% reporting enterprise-wide scaling (up from 38%). 1940% of large organisations (>$1bn revenue) report scaling AI agents, up from 27%.
“How are [other teams] adopting AI into the research process today? [W]e have our own approach, but there doesn't seem to be consensus” - Sam Cravo, Head of Research, Proprietary Trading Firm
Sam Altman, CEO of OpenAI, expressed surprise with the lack of significant disruption, “I thought [...] there was going to be much more disruption [...] than it turned out to be. The economy just has so much inertia. People keep doing the same things, buying from the same company, wanting to use their tools the same way.” - Sam Altman
That may be true for the economy and society en masse. But there are changes and they are forcing business models to evolve.
Buy-side firms are running GenAI firm-wide. Ahead of transcripts and market data, they’re hungriest for high-quality broker research as their primary machine-readable input into those systems.
A July 2026 survey of 35 of the world’s largest asset managers found 77% of them were running organisation-wide GenAI platforms. Also, 77% named broker research as the machine-readable feed they most wanted. Earnings transcripts came second at 57%, then market data at 42%.
Of these 35 asset managers, 69% highlighted broker and data licensing restrictions as their main barrier, ahead of compliance and entitlements (54%)
Now every client is asking for API feeds for machine-readable research. So rather than going through all of the research in the morning, they're getting Claude to curate a morning note for them, or when they have a question on Nvidia, they don't go through their email. They ask their AI tool a question and then that pulls in the answer with attribution to the firm and the analyst and a link. So we're sort of in the early part of this journey and firms are just getting set up for that.” - Bill Featherston, Head of US Research, Mizuho Securities
The buy-side demand positions sell-side research as a core intellectual asset. Which it is.
The catch — how to make this work for the sell-side?
The legacy model, emails and PDFs, clearly doesn’t support the rapidly evolving AI-age we’re finding ourselves in.
Commenting on the findings, Gavin Skinner, COO of Aiera, says: ‘Buy-side firms overwhelmingly want broker research inside their AI workflows, but today’s licensing, entitlement, and compliance frameworks weren’t designed for machine-readable, AI-driven environments.
Technology has rapidly advanced, client demands have shifted, the regulatory burden is constant and evolving, whilst internal constraints (static or declining headcount among others) make for a challenging environment to navigate.
Then there’s the desire for more and more detailed analytics with the need to prove the research is adding to the bottom line. If that’s not all you need to contend with there’s also the question of getting your research in front of the right eyes. How exactly do you stand out and get noticed amongst all the noise, the glut of data and content produced and spat out at breakneck speed.
“Our current challenge is to offer actionable market analysis and commentary to our brokers, clients, and prospects that is differentiated from the “AI swamp”, which I will define as voluminous, low-value information embellished by LLMs.” - Dags Chen, Head of IPA Research, Marcus & Millichap
The glut of publicly available data is expanding exponentially. The internet is awash with content generated by AI creating, what Dags Chen calls, the “AI swamp”.
The Global DataSphere — the annual data created, captured, replicated and consumed — according to IDC, was estimated at roughly 180 zettabytes in 2025 and is forecast to reach 393. 8 ZB by 2028 and to surpass 700 ZB by 2030. Only one-tenth of each year’s total is newly created; the rest is copied. A much smaller share is ever stored.
Increasingly, it’s created at the edge, by machines, and in the service of AI workloads that generate new data as a byproduct. For context, 1 ZB is equivalent to 1 billion TB.
If the relentless influx of low-quality, easily reproduced content online wasn’t enough, it’s resulted in an ever increasing challenge.
Attention is fragmented more than ever and the path to the right audience - your audience - is beset with challenges. The glut of information clogging up the internet, social media and your audience’s mental bandwidth isn’t helping.
“AI has lowered the bar for a lot of research, it also makes it more difficult to get your clients' attention [...] there is an overwhelming supply of research” - Rodolfo Ramos, Head of Mexico Research and Strategist, Bradesco BBI
The highly regulated finance industry is already tasked with being a standard-bearer of reliable information. What may not be so clear is that remaining visible to the next generation will require greater differentiation, originality and more nuanced insight than ever. After all, the next generation can get the same information (and misinformation) at the swipe of a screen.
The generational divide - the chasm of difference in the way Boomers, GenX, Millennials and GenZ consume, respond to and act on information.
“We are clearly in the new landscape of attention. Everyone's fighting for attention. There is so much content and information out there. How do you reach the right audience? What is the right format? Long-form content seems to have had its time” - Steve Hou, Head of Research, Silicon Data
The Boomer patriarch is more comfortable with reading a physical copy of the Financial Times. His grandchild, well, she’s happy to get her information from bite-sized chunks of shortform video on TikTok and Instagram, along with quick-snack memes and flash-in-the-pan ragebait trends mixed in with an unhealthy dose of misinformation.
Sitting between the two, GenX and Millennials dive into a range of different information sources including podcasts, online reports, video and to some extent, printed material.
How do you engage your right audience in the right way? Ad targeting has figured out targeting on Instagram to get you to buy a dress. But for research that remains a little bit unsolved, some people have done it better than others. Torsten at Apollo writes these very nice short pieces, data updates with charts, and engages with a big audience. Banks have their own natural distribution, but I wonder if they also perhaps see some of these challenges about getting the right audience to pay attention.” - Vincent Mak, Head of Global Research Strategic Initiatives and Data Services, Bank of America
Getting attention, standing out - and remaining relevant - will require great agility. This may mean using some of the tools the attention economy creators use.
After all, there is a rise in the popularity of online finfluencers (financial influencers), demonstrating a mainstream appetite for financial advice. Whether the advice is sound is debatable. Young investors use finfluencers “because content is accessible, in plain language, and visual”.
Kakhbod, et al, discovered “unskilled accounts are more engaging, post optimistic content that precedes reversals, and attract larger followings than skilled ones. Social platforms prioritise engagement over skill, so bad advice spreads.”
It’s more “digital financial entertainment” than a source of alpha. This is certainly not an argument to whip out your phone and begin livestreaming. But there is an accessibility gap here which is worthy of investigation.
Where institutional research has borrowed the format of the influencer distribution physics — short, visual, shareable — we look to Apollo’s Torsten Sløk and The Daily Spark.
The Daily Spark is usually one sentence and one chart, with sources and a disclaimer, emailed most mornings. The charts then do the rounds on social media. Behind that sit the longer decks and white papers on Apollo Academy.
“In order for content providers to take full advantage of these new platforms, they will need to, first and foremost, embrace one simple philosophy: COPE (Create Once, Publish Everywhere).” - Daniel Jacobson, then Director of Application Development, NPR (2009)
Your analyst creates an in-depth, quality piece of research. It’s well structured, sourced and meets all compliance standards. This content is then published in many formats to different platforms. Your AI agents can draft channel-native derivatives: an email, LinkedIn thought piece, blog post, an infographic, a script, or a podcast. The claims your analyst makes remain locked to the source.
The pieces of that architecture now exist: a human creates, an agent derives, compliance (or an AI pre-checker with human oversight) clears and then you publish everywhere. Multiple pieces of content on multiple channels resulting in more visibility in front of your desired audience.
“We are dealing with third-party aggregators and kind of holders of our content that are seemingly wanting to make decisions about our content before we are. I think we want to make decisions about how we monetise things, how to best coerce and encourage engagement by our clients, but also make sure that we get paid for it.” - Tripp Frank, Managing Director | Equity Product Manager - Research, Truist Securities.
Coming back to our ever present monetisation priority: just how do we go about generating financial return on the research we produce? And how do we demonstrate this internally?
“[M]y time is spent on the actual monetisation of content and I think it is a challenge because the world's evolving. In large part because of AI and the way in which clients consume research, research is more valuable than it ever has been. But at the same time it takes a lot of effort to make sure that you're monetising it and you're structured” - David Siffringer, Head of Equity Research, Cantor Fitzgerald
Of course, if you’re not able to measure how your insights are being used, you’re not able to prove the value of your research. You cannot prove value if you cannot see it. Which is the point of research-spend analytics.
There’s been an evolution in the research distribution models. It’s gone from push to pull to prompt. Previously market intelligence was pushed out via PDFs and bulk emails. There was low engagement. Emails remained unopened and PDFs remained unread. You had little actionable data, besides open rates, on what resonated.
This evolved, becoming a pull model, relying on active, client-side engagement. Your client would log into an interactive web portal and mobile dashboard to filter data themselves. This was interactive and allowed for more insights via your portal analytics.
But of course, those valuable insights are no longer there anymore with the current era. The prompt era. Clients are now actively running queries on the data they do have access to and are frustrated by restrictions on what they can and can’t feed into their AI assistants.
However, reading of your research, by client-side AI is likely taking place right now. A machine is the primary reader of your research. This is not being monitored or monetised.
To be clear, this isn’t adding an AI layer or chatbot on top of existing underlying, legacy technical architecture or research processes.
Not at all. Rather, to survive the evolution currently impacting the industry, a new approach to research operations may be required. An approach that prioritises the very structure and governance of that data.
Enter what is called componentised structured authoring.
This approach to ‘assembling’ data treats a report as a collection of modular nodes, like an executive summary, a specific chart or a legal disclaimer. Each of these nodes are machine-readable data.
The advantages to this decoupling allows co-authors, editors and compliance officers to work on different parts simultaneously. There’s no more waiting for a master document to be passed around.
This modularity eliminates transcription errors whilst allowing a live data link so that, for example, a legal disclaimer node can be updated when new rulings are released and this auto updates across all iterations of the report. When underlying data changes, the report is updated. If source information updates, so do the corresponding charts or graphs.
The administrative grunt work is minimised since an analyst no longer needs to manually format reports. The formatting is taken care of when, in a well-built system, the nodes self-assemble into a well-formatted, well researched report. Essentially allowing your analyst to focus on their strengths and produce more value for the business.
Instead of compliance remaining a final bottleneck delaying time-sensitive release of market insights it gets embedded into the architecture of the workflows. Regulatory checkpoints and permission gating gets integrated into the production pipeline.
Embedded control turns compliance from a manual hurdle into an automated safeguard. This secure environment ensures AI tools are safe for use, ensuring summaries or charts are derived strictly from verified internal archives.
“We've gone from a world where people have wanted delivery in very distinct, like API-type forms, to now thinking more about MCPs and also thinking about raw data. And so I think that's the stuff that's changing and we're very focused on.” - Brad Rogoff, Global Head of Research, Barclays
Once your research is structured and compliant, these data nodes are then delivered into client AI workflows via entitled API or Model Context Protocol (MCP) endpoints — with usage logged back into your system.
MCP is a buzzword for a reason. Anthropic open-sourced this technology in November 2024 as a “universal, open standard for connecting AI systems with data sources, replacing fragmented integrations with a single protocol.” They describe it as a USB-C for Large Language Models (LLMs) and has thus been adopted by frontier models including ChatGPT, Claude, Gemini, Copilot and others.
MCP standardises the hook-up between your data and your client’s AI model of preference.
Your internal architecture (and the contract or licence that governs your client relationship) controls who may query, how much and whether content may be used for training the client’s model.Between the client-side AI model and the treasure trove of your data exists three important architectural filters to support and control delivery.
“From a technology platform perspective, MCP technologies — that's the thing we're really exploring to see how that might be a more efficient and preferred mechanism for opening up content to AI tools from our clients.” - Vincent Mak, Head of Global Research Strategic Initiatives & Data, Bank of America
A real-time entitlement engine gates access, enforcing specific subscription policies for each client. Real-time controls protect your archive and reports from mass scraping or over-consumption of restricted content. And that third filter: this is your detailed query logger. It tracks which institution retrieved specific research components. This allows syndication to a third party network without sacrificing intellectual property whilst giving valuable insights into the queries run against your research.
Tony McManus, Head of Bloomberg’s Enterprise Data and Index Businesses has this to say about enterprise MCP: “It means semantic metadata so the AI understands what each field means and how to use it reliably. It means delivery infrastructure built for production: live data, historical time series, and a governance layer that enforces entitlements and maintains audit trails.”
He adds: “Without rich semantic metadata, AI systems can’t reason reliably over financial data. They pattern-match against it, which produces unreliable results in a domain where precision is non-negotiable.”
“(T)his is a time where you almost have to adapt a culture to really have buy-in to change — being flexible, dynamic. I think they're joined at the hip. Evolving a culture to support innovation and change is certainly relevant.” - Bryan Polovina, Head of Global Equity Manager Research, BNY Investments
The point von Neumann made in 1948 holds true.
Whatever you can clearly specify, a machine can be built to do that. The limitation is not knowing exactly what the steps are that need to be specified.
Sell-side research — and the processes and thinking involved in conducting, sourcing, verifying, compiling and passing compliance — is specified enough.
Buy-side clients want your research as a direct feed into their AI models. The demand is there. What stops them is a licence, entitlement and a feed you can meter.
Individuals are reaping productivity gains from AI. EBIT lags behind, unless there’s a redesign of workflows. Simply bolting a chatbot on top of current tech architecture, processes and distribution paths is not going to cut it. This is the path that leaves EBIT unchanged.
On the other hand, a firm implementing a COPE layer gets a cleared note in front of more than one audience. Componentised structured authoring streamlines production and compliance. Entitled MCP controls who may query and on what terms, whilst sending usage back to the desk.
The desk that owns that loop still has a business. The desk that only ships PDFs does not.
Machine-readable broker research they can feed directly into their AI tools. In a July 2026 survey of 35 major asset managers, 77% named broker research as their most wanted feed. Earnings transcripts (57%) and market data (42%) came next.
Licensing restrictions were the main barrier (69%), followed by compliance and entitlements (54%). Current frameworks were built for emails and PDFs, not AI-driven workflows.
It's a term for the flood of voluminous, low-value, LLM-embellished content. Standing out takes originality and more accessible formats.
COPE stands for "Create Once, Publish Everywhere." An analyst writes one compliant piece, and AI agents draft versions for each channel, such as email, LinkedIn, infographics or podcasts. Compliance clears them, and every claim stays tied to the original source.
Research delivery has moved from pushed PDFs and emails, to client portals, to clients' AI tools querying the research directly. The paper argues that machines are now the primary readers, and that this reading is neither tracked nor monetised.
It treats a report as a set of modular, machine-readable parts, such as a summary, a chart or a disclaimer. Teams can work on different parts in parallel, updates carry through automatically, and compliance is built into the workflow instead of being a final bottleneck.
Model Context Protocol, an open standard Anthropic released in 2024, connects a firm's research to clients' AI tools. Delivery is controlled through entitlement checks, protections against scraping, and query logging that shows who used what.
