The Funds Furthest Ahead on AI Research Are Building Their Own Platforms

The firms furthest ahead on AI financial research have mostly stopped buying end-to-end platforms.

Across the industry, buying is still the norm. In Mercer's 2026 survey of 131 asset managers, 63% use vendors' off-the-shelf AI tools. The funds that have pushed AI deepest into their research have moved on, and over the last few months we've seen a much wider group of firms start to follow them.

The largest funds got there first, with real budgets behind them. Bridgewater set up AIA Labs in 2023 with six people and has since grown it to more than 75 scientists, engineers and investors. Balyasny stood up a 20-person Applied AI team in late 2022, and about 95% of its investment teams now use what that team built. Balyasny uses frontier models, but its Chief AI Officer, Charlie Flanagan, said in September that "most of the infrastructure is built in-house at BAM, including the execution harness, data access, and review controls".

What's changed is who else is doing it. We keep seeing the same situation at firms with nothing like those budgets: a small team inside the firm, sometimes just one person, building an in-house research platform around how the firm already works. Mercer's numbers point the same way. 57% of the managers it surveyed have between one and five people dedicated to AI in the investment team, and 51% use vendor tools they've customised with their own work on top.

These teams are codifying the fund's research framework into an AI workflow. That's the one part a platform can't sell them, and it's why they've stopped buying one.

Every fund has a research framework

Every fund runs on a framework, whether or not anyone calls it that. It's the set of questions the team asks of every idea, the order it asks them in, what counts as evidence and what kills a thesis.

It looks different at every firm. A credit desk might start with covenant terms, the maturity schedule and what recovery looks like in a downside case. A long/short equity team might start with where its view differs from consensus and what has to be true for that view to be right. A PE deal team works through the market, the unit economics and the management team, and the output has to land in the exact shape its investment committee reads.

For most firms, this has lived in people: the senior analyst who trains the juniors, the IC memo template, the questions a PM asks every Monday morning. Bridgewater is the well-known exception. For decades it has run what it calls an "expert system" approach, where computers apply the causal rules its investors have written down. When it set up AIA Labs, the new unit inherited fifty years of that work.

AI makes it possible for every firm to do some version of this. Each question in the framework becomes a step in a workflow: pull these documents, check this, compare it against that, write it up in this format. The same workflow can run on one name before a meeting, on 300 names every Monday or on every portfolio company each month.

Firms are now hiring specifically for it. AI Street reviewed more than a dozen recent job postings at Point72, D.E. Shaw, PIMCO, BlackRock, Ares and Carlyle, and described the work behind them as "a translation job: turning the judgment of portfolio managers, analysts and deal teams into prompts, retrieval systems, scoring frameworks and agents". Man Group says its AlphaGPT system "follows our investment methodology to the letter". In private equity, Advent has built what co-chair James Brocklebank calls an "IC robot", trained on 13 years of the firm's investment committee papers. It reviews every new deal brought to the committee and flags assumptions the firm's history doesn't support, such as a margin that comparable companies have never reached.

What a platform sells

An end-to-end research platform is four products sold as one. There's the interface, usually a chat window with some document handling. There's a research framework: prompts and agent steps that encode the vendor's research process, deciding how a question gets broken down, which sources get checked, what the output looks like and where it's saved. There's the model, picked and tuned by the vendor. And underneath all of it there's the data layer, meaning everything the model can retrieve from: the open web, company filings, market data, research papers and whatever content the vendor has licensed.

In 2023 and 2024 most funds had no model access of their own and no way to point a model at their own documents, so buying everything in one contract made sense. It got AI onto the desk in weeks. After a year of use, a team knows which of the four it needs from a vendor.

Where a platform's framework fits badly

This isn't a knock on the platforms. Plenty of them are well built. But every platform shows up with opinions about how research should be done, and that process is one of the things a fund really owns. Bridgewater's AIA Labs put the reason plainly: "When every investor has access to the same sources of information, alpha must come from differentiated judgment and taste."

A platform's framework has to work for every customer, so it fits each one roughly. The gap shows up in the workarounds: analysts re-prompting to force the house format, copying output into the house template by hand, keeping a document of prompts that make the tool behave more like the desk.

Breadth makes the gap wider. Most platforms now offer document grids and scheduled runs, but the steps and the output are limited to what the vendor's product exposes. The desks pushing hardest want their own analysis run every week across a list of 300 names, with what changed since last week flagged at the top. We wrote about that workflow in What Happens When the Coverage Universe Stops Scaling With Headcount, and it's very hard to run inside someone else's interface.

Balyasny and D.E. Shaw are both set up so that each team can encode its own framework. Balyasny's central team builds shared components (agent frameworks, toolchains, compliance guardrails), and each investment team then builds its own agents for its asset class. Balyasny's COO, Kevin Byrne, described the result:

"Every one of our investment teams can decide how to apply the latest AI to their process."

D.E. Shaw decided against building one bot for everyone and gives desks a set of building blocks they put together themselves. Both setups keep the framework with the people who use it.

Most of the usage was the data

Then you look at what people inside the firm were actually using the platform for. A lot of the time they were going in for the data sources underneath: a filing, a transcript or last quarter's numbers. The interface and the workflow on top got much less use than the original business case assumed.

Our own traffic shows what builders call for once the stack is theirs. In the 4.2 million financial queries we analysed earlier this year, filings and financial data made up 35% of volume, and most of those calls came from another agent or an internal pipeline that needed a specific number or clause to keep going. Nobody was reading those answers in a chat window. In Mercer's survey, 69% of managers named data constraints as a significant barrier to using more AI in their investment process.

That makes the platform a pretty hard cost to defend at renewal. If the main thing the team gets out of it is access to sources, paying platform pricing for that access looks expensive next to buying the sources directly. Some firms do keep a platform licence for content they can't get anywhere else, such as a broker research feed or an expert-call library, and they treat it as a data purchase. Lock-in is part of the worry too. Boden Anderson, a director at the consultancy Alpha FMC, said in August: "There's a hesitancy to fully tie up to [platform providers] because then they have full pricing power over firms in this rapidly changing world."

The platform's data is also whatever the vendor has set up, and it's often awkward or impossible to add the internal notes, models and past memos the firm already has. Those internal sources matter more now that public data is easy for everyone to reach. Jacob Bowers, a vice president of quantitative research at BlackRock, said in April that public data has been "commoditized" by AI and that "some of the best unstructured data you have is internal".

Why one person can build this now

Three years ago, building your own meant building the data layer too. In the same query analysis we described the typical pattern back then: four to six engineers, four to six months, and ongoing maintenance every time a filing format or a transcript provider's API changed. Point72's CTO has described spending the firm's first six months normalising ticker aliases and vendor IDs before fine-tuning a single model. Only the largest firms could carry that cost.

Today a large part of the data layer can be bought ready to query, and more of it comes ready for agents to call directly: Neudata's 2026 industry survey found that 44% of data providers now offer access through MCP. Joining that data to the firm's own identifiers and internal records is still the team's job, but parsing, indexing and maintaining the public sources no longer has to be. Models are API calls too, and a team with its own test set can check a new one against its questions in a few days before switching.

What's left to build is the part that's specific to the firm: the prompts and agent steps that encode its research framework, the connection to its own documents and the place where the output lands. One to five people can build that, and sometimes one person does. Ken Griffin said in July that "one of our team members built an agentic system" at Citadel that reads an academic finance paper, reproduces it and tests the result out of sample in two to three hours. That work used to take six to eight weeks. At Balyasny, it was the Chief Economist who configured an agent workflow that cuts a recurring central-bank analysis from about two days to about 30 minutes.

What codifying a framework looks like

Take a long/short equity team's earnings review. The framework is simple to say out loud: before the call, know what the thesis needs to see; after it, check whether the numbers beat or missed, whether guidance moved and whether anything management said changes the position. Written as a workflow, it might run like this:

1. When a company in the book reports, pull the earnings release and the call transcript, and the 10-Q once it's filed.
2. Pull the team's last memo on the name and the three or four points the thesis rests on.
3. For each point, find what the new documents say and cite the passage.
4. Compare the reported numbers and guidance with consensus and with the team's own model.
5. Write it up in the team's format: what changed, what it means for each point and what to ask management.
6. Post it to the desk's channel before the morning meeting.

Steps 1 and 4 mostly come down to retrieval: getting the right documents and numbers quickly, with a source for each. Everything else is the framework. A credit desk would swap in covenant and liquidity checks and a different memo, and the retrieval underneath would barely change. A PE deal team would make the same split on a new CIM: the data room and past IC papers are internal, and what it buys covers public comparables and sector data.

What the in-house setups look like

The setups differ from firm to firm, but most of the ones we see have the same layers:

• Models: usually two or three providers, with different tasks routed to different models and switched when a better one passes the team's tests. Balyasny uses models from OpenAI, Google and Anthropic within what Google calls its "multi-model architecture", and Advent describes its system as "agnostic to the LLMs" and using all of them.
• The research framework: the memo format, the checks and the house view on each sector, written as prompts and agent steps. This is the part the firm treats as its own.
• Retrieval: external sources called through APIs, alongside the firm's own documents, models and notes.
• Controls: which teams can see which documents, walls between desks where they're needed, and a log of every prompt and source that compliance can review.
• Delivery: output lands where the desk already works, whether that's a memo template, an internal wiki, Slack or a scheduled run that emails a summary every morning.
• Evaluation: a set of the firm's own questions with known good answers, rerun every time a model or a source changes.

The retrieval layer is where we come in. These teams come to us for data sources and tools they can drop into that stack and set up however they want. Our Search API gives their agents one endpoint for the full text of SEC filings, academic papers and the open web, alongside company fundamentals, market prices and economic data, called from the team's own code or through our MCP server. Filters cover sources, dates and the most a query is allowed to cost. DeepResearch takes the longer multi-step questions and returns a cited report. The team writes the code that calls each one and decides where the output goes.

How builders choose

These are our favourite people to work with, if we're honest. They question everything and they'll tell you when something's rubbish.

They run the same set of questions through two or three providers and keep whichever does best. What they check is concrete: whether the filing is there and how soon after it's published (we wrote about how we handle that here), whether a result points to the exact passage or just the document, and how much control they get over which sources and dates are searched. They check what a query costs at the volume of a weekly run across a few hundred names. And they check whether the provider stores their queries, because a log of what a fund searches for shows what it's looking at. They keep testing after anything goes into production, and when a result is wrong they tell us, usually with the exact query that broke it.

Who chooses the tools now

The decision maker used to be the CIO signing off a platform. Now it's the person building the thing, and they've usually got the room to try whatever gets them the best result.

That changes how tools get chosen. A platform deal was a top-down decision, made once, often off a demo and a feature list, and then lived with for the length of the contract. A builder adds one source and tests it against the firm's own questions, which takes days. Getting a new vendor through due diligence and security review takes longer, so teams test several and onboard the one or two that come out ahead. Each choice gets revisited when something better shows up, and the research framework stays in-house with the people who run it.

When a platform still makes sense

Building your own has a real cost. Someone has to own the prompts, the evaluation set and the model upgrades, and if that person leaves, the firm is left with a system nobody maintains. A firm with no one to do that job, or whose research process is close to a platform's defaults, is better served by buying the whole thing, and that still describes most of the industry. A platform is also a reasonable first year: it shows the team what it uses, which is the information it needs to decide what to build.

The firms that have been through that first year are the ones we see moving. Most of them found that what they used was the data, so they kept buying that and built the rest themselves.

If you're building one of these in-house, the APIs are open at platform.valyu.ai and the docs are at docs.valyu.ai.