Home / Learn / Data & AI / Offline AI / Lesson 18.3
All lessons

Module 18: Who uses this: sectors and careers · Lesson 18.3 · about 20 min

Finance and banking

Analysts, banks and insurers use local AI to read reports, check documents and write code against data that may not leave the building. What works, what regulators expect, and a sensible setup.

Key points, meanings and formulas on one page: choose "Save as PDF" in the print box.

Module 18 · about 20 minutes · read first: 17.7. You’ll be able to plan local AI for an analyst team or a finance firm.

Finance firms can run local models on data that may not leave the building, and because a local model file never changes under them, they can test, record and govern it like any other model.

1. Why this sector

Banks, insurers and investment firms hold customer data and market-sensitive information under strict rules, often with policies that bar sending it to outside AI services. They also already know how to govern models. Local AI fits both facts.

2. Top uses

  1. Reading reports: summarise annual reports, filings and research notes; pull out key figures with page references.
  2. Analyst coding help: draft SQL and Python against internal data (Lesson 13.3), checked before running.
  3. Document checks: know-your-customer files, loan documents and claims checked for missing items.
  4. Search across policies and procedures with citations (Module 12).

3. What regulators expect

  • Canada: OSFI’s Guideline E-23 on model risk (effective 2027) covers AI models in federally regulated banks and insurers.
  • USA: the long-standing SR 11-7 model risk guidance, plus fair-lending and consumer rules.
  • UK: the FCA and PRA apply existing rules, including the Consumer Duty and model risk principles.
  • India: the RBI’s FREE-AI framework (2025) sets out responsible AI expectations for regulated entities.

In practice: an inventory of models, testing before use, documented limits, monitoring, and named owners. This is model risk management. A local model file with a fixed version is easy to put on that inventory.

4. A sensible setup

A server with a strong card (Lesson 5.5) inside the firm’s network, a serving engine (Lesson 15.2), access through existing logins, logs kept per policy, and a model record: file name, hash, licence, tests, owner.

A checksum (or hash) is a short fingerprint worked out from every byte of a file. Change one byte and the fingerprint changes completely, so it proves you are running exactly the file you tested.

Pin the file. The team records the exact model file and its SHA-256 checksum, for example from sha256sum model.gguf. Same checksum next year means same model.

Pin the settings. Quantisation, context length, temperature and the full system prompt go in the record. Change any one and the behaviour can change.

Test it. 50 past filings with known answers: the model pulls 25 fields from each, and the record notes how often it is right and which fields it misses.

Name an owner and a date. One person is accountable, and a review date is set. A new file, quant, setting or prompt means testing again before use.

Maths and principle

Principle: you can only validate what you can pin down.

Equation: a model record = file + checksum + settings + test results + owner + review date.

Analogy: a laboratory instrument with a calibration certificate.

Try it: what changes would make you re-test? (A new model file, a new quant, new settings or a new system prompt.)

file + checksum
exactly which model, proved by its fingerprint
settings
quantisation, context, temperature and system prompt
test results
how often it was right on real past cases, and where it failed
owner + review date
the accountable person, and when it must be checked again
Physical meaning
Like a calibrated weighing scale in a lab: the certificate is only worth something while the scale is the same scale. A cloud model can change behind the same name; a local file with a checksum cannot, which is what lets a finance firm sign off on it.

Page references and flags keep the human checking quick:

“It runs inside our network, so it does not need validating.” Where a model runs and whether it works are separate questions. Regulators expect any model used in business decisions to be tested, documented, monitored and owned. Running locally just makes that easier, because the model cannot change under you.

5. Worked scenario

Made-up team. A credit research team of 10 at a Mumbai bank reads 300 company filings a quarter. A local model on the bank’s server extracts 25 fields from each filing with page references; analysts check the flagged ones. Reading time per filing falls from 50 to 20 minutes, and the model record satisfies the bank’s model risk team.

6. Practice

  1. Write a model record for the model you use.
  2. Why is a page reference important in finance extraction?
  3. Name one finance task to keep away from AI decisions.
Answers
  1. File name, checksum (sha256sum or certutil -hashfile), quant, settings, prompt, tests, owner, review date.
  2. Analysts can check each figure quickly against the source.
  3. Credit, claims or fraud decisions about a customer, unless built and governed as proper models.

New words: model risk management, model inventory, checksum, validation.

  • Good uses: reading filings with page references, analyst coding help, document checks and policy search.
  • Regulators (OSFI E-23, SR 11-7, the FCA and PRA, the RBI’s FREE-AI) expect models to be inventoried, tested, monitored and owned.
  • A model record = file + checksum + settings + test results + owner + review date; re-test after any change.
  • Decisions about customers (credit, claims, fraud) need properly governed models and people, not a general chat model.

Check yourself

1What do finance regulators in all four countries expect for AI models used in decisions?
2Why do banks like that a local model "stays the same"?
3Which task should never rest on a general language model alone?

Sources

Your ticks stay in this browser. Save them as a file on the course home.

Updated 2026-11-22. Prices and model names change fast: check the official source before you buy.