Inside BlackRock's Aladdin Copilot build: how the largest asset manager layered generative AI over its research stack

BlackRock's decision to wire OpenAI and Microsoft into the Aladdin platform, disclosed publicly in January 2024, has become the most-cited template on the buy side for putting generative AI into production research work. The case is instructive less for the technology than for the sequencing: BlackRock did not rebuild its research stack, and it did not replace primary-research inputs like earnings transcripts and expert calls. It layered a generative interface over the system of record and treated everything else as retrieval fuel.

That architectural choice, and the way it was staged across a shareholder letter, an engineering hire, and a set of vendor partnerships, is now the reference pattern that adjacent vendors (AlphaSense, Rogo, Hebbia) are designing around rather than against.

Background: the buy-side production-AI gap in 2023

Through 2023, most large asset managers were running generative AI as a cluster of pilots. The gap between a working demo and a tool that a portfolio manager would actually use inside a live workflow was wide, and the reasons were structural rather than technical. The system of record for investment decisions at large firms is not a chat interface. It is a risk-and-analytics platform with decades of accreted workflow, model libraries, compliance controls, and integrations into order management and reporting. Any AI layer that asked users to leave that platform lost before it started.

BlackRock had a structural advantage here that few of its peers could match. It owns Aladdin, the risk-and-portfolio-management platform used across its own $11.5T+ AUM franchise and licensed to external asset managers, insurers, and pensions. If the LLM layer was going to sit somewhere, Aladdin was the obvious substrate. That is the piece of context that makes the case study non-generalizable in one direction (most firms do not own their system of record) and highly generalizable in another (every firm has one, and the architectural principle transfers).

The second piece of background is what BlackRock chose not to build. It did not attempt to replace the primary-research inputs that its analysts already consumed: earnings transcripts, sell-side notes, expert-network calls, and internal analyst commentary. Those remained upstream sources. The generative layer was designed to sit over them, not in place of them.

The approach: staged disclosure, infrastructure partner, engineering hire

BlackRock staged the rollout across three visible moves in the first weeks of 2024.

First, the shareholder communication. Fink's January 2024 chairman's letter placed AI at the center of the firm's productivity narrative, describing the deployment of an internal co-pilot to employees and framing the technology as a way to raise output per Aladdin user rather than as a headcount story. Fink reinforced the framing in interviews with Bloomberg and on CNBC the same week, discussing the OpenAI collaboration in public.

Second, the infrastructure partnership. Reuters reported on January 12, 2024 that BlackRock had tapped Microsoft and OpenAI to build generative AI tools for Aladdin, with Microsoft providing the Azure infrastructure and OpenAI supplying the model layer. Positioning Microsoft as the infrastructure partner was consistent with BlackRock's existing enterprise footprint and gave the deployment a compliance and procurement path that a direct-to-model integration would have lacked.

Third, the engineering hire. BlackRock recruited Lance Braunstein from Goldman Sachs to run Aladdin Engineering, a move that signaled the AI push was going to be executed inside the platform's core engineering organization rather than as a side project owned by a data-science group. That distinction matters. A pilot owned by a data-science team ships demos; a program owned by platform engineering ships production features.

What the co-pilot actually does

BlackRock has described Aladdin Copilot as a generative AI interface layered into Aladdin's existing analytics, designed to let users query the platform in natural language, surface relevant data, and accelerate workflows that previously required navigating multiple screens. Public detail on the exact retrieval architecture is limited, and INFLXD is not going to fill that in with speculation. What is on the record is the framing: an interface layer over the analytics platform, with the underlying models, positions, risk metrics, and research notes as the substrate the LLM operates on.

What happened: reference architecture, not a moonshot

By mid-2024 BlackRock had Aladdin Copilot live for internal users and had extended its AI posture in adjacent ways: taking positions in and partnering with generative-AI infrastructure vendors, and integrating Aladdin with Microsoft 365 Copilot workflows. The Financial Times covered the broader arc of the Aladdin generative-AI program as it developed through 2024.

The part of the outcome that matters for the rest of the buy side is not the feature list. It is the architectural pattern the deployment locked in:

1. The system of record stays put. Aladdin remained the platform of truth for positions, risk, and analytics. The LLM was a layer, not a replacement.
2. Primary-research inputs stayed upstream. Earnings transcripts, expert-network calls, analyst notes, and sell-side commentary continued to flow in as sources. The generative layer read them; it did not attempt to originate them.
3. Infrastructure was outsourced to a hyperscaler. Microsoft and OpenAI provided the model and compute layers under an enterprise contract, avoiding a build-from-scratch model effort.
4. The interface was framed around augmentation. Fink's public language, and BlackRock's product positioning, described the tool as making analysts and PMs more productive rather than automating their judgment.

Each of these choices is defensible in isolation. Together, they define a template.

What it signals for expert networks and research vendors

The BlackRock case reset the competitive geometry for research-AI vendors. Before the Aladdin deployment was public, one plausible read of the market was that a well-funded startup could build a research operating system that displaced the incumbent stack. After it, that read looks harder to defend at the tier-one buy-side level. The system of record at large asset managers is not going to be rebuilt around a chat interface. It is going to be wrapped in one.

That has clarified where independent research-AI vendors can play, and where they cannot.

Where they can play. As retrieval, ingestion, and analysis layers that feed the system of record. AlphaSense's expert transcript and document library, Hebbia's document-analysis workflows, and Rogo's analyst-workflow tooling all sit naturally as inputs or adjacent surfaces to a platform like Aladdin. The pursuit of Model Context Protocol (MCP) and platform-adjacent integrations across these vendors is consistent with a market where the buy side has decided the system of record is not up for grabs.

Where they cannot play, at the tier-one level. Displacing the platform that houses positions, risk, compliance, and reporting. That is the layer BlackRock owns for its own franchise and licenses to a large fraction of the industry, and it is the layer other large asset managers have their own equivalents of.

For expert networks specifically

The interesting implication for expert networks is that the BlackRock model treats transcripts and expert commentary as retrieval fuel, not as a workflow their analysts leave the platform to consume. That elevates the value of clean, structured, machine-readable transcripts and lowers the value of a walled research portal that requires a separate login. Expert networks whose content flows cleanly into a client's LLM-augmented workflow are on the right side of the architecture. Those whose content is trapped behind an interface that assumes the analyst reads it in the vendor's UI are, over time, on the wrong side.

Questions a research analyst at a peer firm should be asking an expert on this topic next:

• How is the retrieval layer in a deployment like Aladdin Copilot handling provenance and audit for transcripts that carry expert-network compliance flags?
• What is the practical latency and cost profile of running generative queries over a full internal research corpus at BlackRock's scale?
• How are portfolio managers actually using the co-pilot in daily workflow, versus how it was demoed?
• Which parts of the Aladdin-Microsoft-OpenAI integration are contractual dependencies, and which could be swapped to another model provider?
• How is BlackRock reconciling the augmentation framing with the productivity gains Fink described publicly, in terms of hiring and headcount planning?