Most large financial institutions now have AI agents somewhere in the building. They summarize research, draft memos and answer questions about internal policy, and they do it well. Far fewer have agents working inside the processes that move money, confirm trades, onboard clients or test covenants, even though those are the processes where the economics of automation are most compelling. The gap is usually explained as caution, and caution is part of it, but the more practical explanation is that an agent can only act on a record it can trust, and in much of finance that record is still a document.
Finance has automated before, and the pattern is consistent
The industry's own history is the best guide to how this plays out. In the mid-2000s, credit derivatives volumes grew faster than the back offices supporting them, and unconfirmed trades piled up to the point that regulators called the major dealers in to fix it. The answer was not more people reading faxes. It was electronic confirmation on shared platforms, which worked because the industry first agreed on what a trade record contained.
The same sequence has repeated since. Central clearing of standardized swaps after the financial crisis depended on trades being expressed as data that a clearing house could accept. Each wave of regulatory change, from margin rules for uncleared derivatives to the retirement of LIBOR, sent firms back into tens of thousands of bilateral agreements to find out what they actually said, because the terms that mattered were sitting in negotiated legal text and not in any system. Institutions that had already turned those agreements into structured data moved quickly, and the rest staffed up.
The lesson that people who lived through those programs tend to draw is that straight-through processing arrives wherever the underlying record has been standardized, and stalls wherever it has not. Automation follows the data, and where the data is locked in documents, the work stays manual regardless of how capable the tooling becomes.
Why agents make this more pressing
An agent differs from earlier automation in that it carries out a sequence of steps and makes decisions along the way. A script that receives a bad input typically fails, which is annoying but visible. An agent that receives a plausible but wrong input proceeds, and the error travels with it into the next step, whether that is a payment calculation, a position update or a message to a counterparty.
This is why the conversation in operations has shifted from keeping a human in the loop on every action toward keeping a human on the loop, supervising a process that runs by default and stepping in on exceptions. That model is far more efficient, and it only holds up under scrutiny from risk, audit and regulators if every action the agent takes can be traced to the evidence it relied on. The control framework for agentic workflows is, in large part, a question about the quality and provenance of their inputs.
Where the documents still are
Decades of investment have made listed markets close to fully electronic, so the remaining manual effort is concentrated in the parts of finance that are bilateral, bespoke or private. These are also the parts that have grown fastest.
- Lending and private credit. A loan is defined by its credit agreement and amendments, and serviced through a stream of agent notices, each of which someone reads and keys into a system.
- Private markets portfolio monitoring. Portfolio companies and fund managers report through board decks, financial statements and capital account statements in whatever format they prefer, and investors rebuild the numbers by hand every quarter.
- Client onboarding and regulatory documentation. Opening an account or a trading relationship means collecting and validating formation documents, tax forms, authorized signatory lists and negotiated agreements, then repeating much of it whenever a rule changes.
- Credit analysis and underwriting. Analysts spread borrower financials from PDFs before any judgment can be applied, and the definitions they need are specific to each deal.
- Managed services. Much of the above is outsourced to providers whose cost base is the people doing the reading, which makes extraction accuracy a direct driver of margin.
In each of these areas an agent could plausibly take on a large share of the workflow, and in each of them the first thing the agent would need to do is read a document that was never designed to be read by a machine.
What an agent-ready record looks like
General-purpose language models can read these documents after a fashion, and for a research summary that is often enough. For a workflow that settles a trade or tests a covenant, the bar is different. The structure of a table has to survive extraction, because a pricing grid or a schedule of investments means nothing once its rows and columns are scrambled. Every value needs a pointer to the place on the page it came from, so that a supervisor reviewing an exception can verify it in seconds. The same document has to produce the same output each time it is processed, since controls cannot be built around a result that varies from run to run. And the whole thing generally has to run inside the institution's own environment, given how sensitive the underlying material is.
None of this is glamorous, and it resembles the reference data and messaging standards that made earlier generations of market automation possible. It is infrastructure, and it tends to be noticed only when it is missing.
What comes first
This is the layer Pulse works on. Its proprietary vision-language models are built for the layouts that financial documents actually use, reconstructing tables, multi-page schedules and scanned pages into structured output that can be mapped to the schema a given workflow expects, with each value tied back to its source. Institutions run Pulse as a hosted service or inside their own cloud, and use it to feed the systems and agents that sit downstream, from credit models and servicing platforms to portfolio monitoring tools.
The order of work matters more than the choice of model. Institutions that first settle what their documents say, in a form that is structured and auditable, can then let agents act on it under controls that risk and audit teams will accept. Those that skip the step are likely to keep running pilots that never leave the sandbox, and to conclude that the technology is not ready when the gap was in the record underneath it.



