Financial data is about as sensitive as data gets: transactions,
positions, client communications, filings that haven't gone public
yet. Most of it is regulated, and almost all of it is a problem the
moment it leaves your compliance boundary. That's exactly why
finance is one of the clearest cases for private AI: models that
are fine-tuned on your own data and served entirely on
infrastructure you control, rather than a public API call that
hands sensitive data to a third party on every request.
Here are six places it shows up in finance today.
Data goes in, a fine-tuned model comes out. Nothing crosses back out to a third party in between.
The six use cases
TRANSACTION MONITORING & AML
Flag suspicious activity and classify transactions against your own historical patterns, instead of a generic fraud model that's never seen your customer base.
DOCUMENT INTELLIGENCE
Extract terms from contracts, parse filings, and summarize statements, on documents that often can't legally leave the firm in the first place.
RESEARCH & ANALYST COPILOTS
Summarize earnings calls and research notes in your firm's own analytical voice, grounded in your internal research, not the open web.
UNDERWRITING & CREDIT RISK
Score risk on models fine-tuned on your own loss history, where the training data is exactly the kind of information you can't hand to a vendor.
TRADING SIGNALS & RISK MODELS
Serve inference in single-digit milliseconds for signal and risk models, where a network hop to a third-party API isn't just a privacy issue, it's a latency one.
CLIENT-FACING ADVISORY TOOLS
Answer client questions against real account and portfolio data without that data ever being sent to a model you don't control.
Transaction monitoring and AML
Anti-money-laundering and fraud detection live and die on
pattern recognition against your own transaction history, not
general knowledge about the world. A public model has never seen
your customers' normal behavior, so it has nothing to compare an
anomaly against. A model fine-tuned on your own historical
transaction data learns what normal looks like for your actual
customer base, which is what makes it useful for flagging what
isn't.
Document intelligence
Contracts, prospectuses, credit agreements, and filings are
dense, structured, and often under an obligation to stay inside
the firm. A model fine-tuned to extract specific terms, covenants,
or risk factors from that exact document type does the job a
general-purpose model does poorly: general models are good at
summarizing prose, not at reliably pulling the same ten fields out
of a thousand slightly different contracts.
Research and analyst copilots
Every desk has its own house style for how research gets
written up: which metrics matter, how conviction gets phrased, what
counts as a red flag. A model fine-tuned on a firm's own research
archive picks up that voice and grounds its summaries in the firm's
own prior work, rather than defaulting to generic financial
commentary pulled from public training data.
Same model, same weights, radically different latency depending on where inference actually runs.
Underwriting and credit risk
Loss history, income data, and credit decisions are some of the
most tightly regulated data a financial institution holds, and
also exactly the data a risk model needs to be trained on to be
any good. Fine-tuning on that data privately means the model
actually learns your institution's real risk patterns instead of
a generic proxy, without that data ever sitting in a third party's
training pipeline.
Trading signals and risk models
For anything near the trading desk, latency isn't a nice-to-have,
it's the whole point. A round trip to a public model API adds a
network hop, a queue, and inference time on shared infrastructure,
easily 300 milliseconds or more. A private inference engine running
on the same network as the trading system can respond in single-digit
milliseconds, because there's no external hop at all. At that speed,
privacy and performance point the same direction.
Client-facing advisory tools
Advisors and clients increasingly expect to ask natural-language
questions against real account and portfolio data. Doing that
through a public model means that account data is now part of a
third-party request. Doing it through a private, fine-tuned model
keeps the exact same experience while the data never leaves
infrastructure the firm controls.
Where this fits at Numerata
Every one of these starts the same way: fine-tune a model on
your own data with P95, on
Lupine compute that scales to zero
when you're not using it, and serve it through
NinetyFive, Numerata's inference
engine, at sub-50ms latency inside your own trust boundary. Private
cloud or fully air-gapped, the data that makes each of these models
useful never has to leave the building to train or run it.
We've broken down the same idea for other parts of the business too:
see how private AI plays out in
trading and
internal tooling.