BlackRock groups its internal AI applications into document extraction, workflow automation, question answering, and agentic systems.
2
A Kubernetes-native framework gives domain experts a sandbox for refining extraction templates before an app factory packages their work into deployable applications.
3
In regulated financial workflows, BlackRock keeps people in the review loop and weighs the cost of an AI application against an off-the-shelf product.
Summary
Infant Vasanth and Vaibhav Page describe how BlackRock helps investment operations teams build document extraction applications. Their example is new issue operations, where staff ingest prospectuses or term sheets, extract security details, validate them, transform the results, and send them to downstream systems. Building each application used to take between three and eight months. BlackRock built a modular framework with a sandbox for domain experts and an app factory that turns the resulting configuration into a cloud-native application. The sandbox supports prompts, extraction fields, validations, dependencies, document management, extraction review, and low-code transformation workflows. The speakers also explain why scaling requires prompt versioning, evaluations, different model strategies, deployment choices, access controls, and cost controls. They are cautious about full automation. In a regulated setting, human review remains part of the design, and teams must assess whether building an AI app costs more than buying an existing product.
Investment operations needs internal tools that move from documents to trades
Investment operations teams acquire data, execute trades, run compliance checks, and handle post-trading work. Their tools support the daily activities of portfolio managers and analysts, so each domain can require a fairly complex internal application. BlackRock therefore wants to build and release these applications quickly. The speakers frame the work around the flow from incoming information to an action in the firm's internal systems, rather than around a general-purpose chatbot.
BlackRock classifies its AI work into four application types
The speakers group their applications into document extraction, complex workflows and automation, question-and-answer systems, and agentic systems. Large language models can augment existing systems in each group. The classification helps teams choose the type of application they are building before deciding how its prompts, models, integrations, and operational controls should work.
New issue operations turns prospectuses into configured securities
When a company goes public or a stock split occurs, the new issue operations team sets up the security in BlackRock's internal systems before traders or portfolio managers can act on it. A simplified workflow ingests a prospectus or term sheet, sends it through an extraction pipeline, involves equity or ETF domain experts, produces structured output, and applies transformation logic before integrating with downstream applications. The domain knowledge needed for complex instruments remains difficult to encode fully in an agent.
Scaling requires managing prompts and model strategies together
Prompts that begin as a few sentences can grow into several paragraphs describing a financial instrument. Domain experts need to iterate, version, compare, and evaluate those prompts against datasets. The model strategy also changes by document and instrument. A small, simple document may work with in-context prompting, while a document thousands of pages long may require another approach because of context limits. Teams may combine several strategies with their prompts.
Deployment adds infrastructure and cost decisions after prototyping
Once an application works, BlackRock still has to handle distribution, access control, user federation, cluster selection, and cost controls. An equity team processing hundreds of research reports overnight may need a GPU inference cluster. The new issue workflow may use a burstable cluster instead. These deployment choices need to be defined so application delivery can approach a CI/CD pipeline.
A sandbox lets domain experts own more of the iteration cycle
BlackRock's framework separates a sandbox from an app factory. The sandbox is a playground where operators build extraction templates, run them on document sets, and compare results. It supports fields, prompts, data types, extracted or derived values, required fields, validations, and inter-field dependencies. For example, a callable bond can require call date and call price fields. This puts configuration and testing closer to the people who understand the instruments.
The app factory packages tested components into end-to-end applications
The app factory is a cloud-native operator that takes the definitions built in the sandbox and spins out an application. The framework also includes document ingestion and management, extraction runs, review, and low-code or no-code transformation and execution workflows. The resulting user-facing application hides template configuration and downstream integration. Users upload documents, run extraction, and receive the configured pipeline rather than assembling its parts themselves.
The speakers advise designing for a human in the loop first in regulated environments. Compliance and regulation require a four-eyes check, even when teams are tempted to make the workflow fully agentic. They also say teams should compare the return on investment of building an AI application with using an off-the-shelf product. Security controls apply across infrastructure, platform, application, and user layers, while model providers and extraction strategies vary by use case.
"You have to really evaluate what your ROI is and is it going to be more expensive actually spinning up an AI app versus just having an off-the-shelf product that does it quicker and faster."13:33
Who should watch
You are building document extraction or workflow applications for investment operations and need domain experts to refine them without waiting for an engineering release.
Your team is deciding how to support prompts, evaluations, model strategies, cluster provisioning, access control, and downstream transformations in one delivery path.
You work in a regulated environment and need a practical approach to human review, security boundaries, and the economics of custom AI applications.
