AI model
A system trained to turn an input into an output: text, code, classification, summary, image, score, or action suggestion.
Finance useUseful for drafting, search, triage, classification, explanation, and review support.
Use a model as part of a control map.
An AI model is not a decision-maker by default. In finance work, the safer pattern is: source → model output → evidence check → boundary → human review.
01Sourcedocument, row, message, record
02Modelsummary, score, draft, extraction
03Evidencewhat supports the output
04Boundarywhat it cannot decide
05Reviewperson checks before action
Simple definitions
Definitions before opinions.
Start here before comparing vendors, benchmarks, or model rankings.
Large language model
An AI model trained to predict and generate language. It can write, summarize, translate, reason over text, and call tools when connected to them.
Finance useUseful for policies, notes, research, customer-message drafts, and analyst support when sources stay visible.
Context window
The amount of information a model can read at once. A larger context window helps with long documents, but it does not guarantee accuracy.
Finance useUseful for long PDFs, audit trails, contracts, policy manuals, and multi-source reviews.
Open weight
A model where the learned weights are available for people to download or run under a license. It is not automatically open source.
Finance useUseful when privacy, control, cost, or local deployment matters, but the license and security model still need review.
Inference
The moment a trained model is used to produce an output from a new input.
Finance useUseful to separate model training from day-to-day use, pricing, latency, and review controls.
Benchmark
A test used to compare models. Benchmarks are helpful, but they rarely tell the full story for a real workflow.
Finance useUse benchmarks with cost, latency, source handling, privacy, and human-review requirements.
Model families
Different models fit different jobs.
Do not pick a model because it is famous. Pick it because the task, evidence requirement, privacy posture, and review workflow fit.
FamilyKnown forCheck before use
AnthropicClaude
Long-form analysis, writing, coding help, policy reasoning, and careful summaries.
Check model ID, pricing, context limits, source handling, and tool permissions.
OpenAIGPT
General reasoning, tool-using apps, multimodal workflows, product prototypes, and broad task coverage.
Check current model availability, API pricing, data controls, and output-review needs.
GoogleGemini
Multimodal work, long-context research, document review, and Google ecosystem workflows.
Check current model version, context behavior, file handling, and source trace quality.
MetaLlama
Open-weight experimentation, local deployment, controlled environments, and custom workflows.
Check license, hardware needs, quality tradeoffs, and whether local operation is actually required.
Open and commercial model labsQwen / DeepSeek / Mistral
Cost-sensitive automation, coding, local experiments, multilingual work, and model routing tests.
Check license, hosting route, privacy posture, task fit, and current benchmark evidence.
Search-answer platformsPerplexity / answer engines
Source discovery, quick research, current-event orientation, and finding documents to inspect.
Open the cited sources. Treat the answer as a starting point, not proof.
Comparison rule
Compare models like a workflow, not a scoreboard.
What job is the model doing: search, draft, code, classify, extract, summarize, or review?
Can the reader see which source was used, what was inferred, and what remains uncertain?
What does the task cost at real volume, including retries, long context, and review time?
Can the workflow safely handle documents, customer data, logs, or regulated context?
What decision is the model not allowed to make?
Who checks the output before it affects money, customers, compliance, or reputation?
Finance workflows
Where models can help without taking authority.
- Summarize policy documents and flag sections for human review.
- Draft AML or fraud investigation notes from structured evidence.
- Compare public sources before writing a market or regulation brief.
- Classify support or compliance messages into review queues.
- Generate code or tests for internal tools where a developer verifies the result.
- Create evidence packets: claim, source, model output, boundary, and next review step.
Boundary
AI model output can support research, drafting, triage, coding, and review. It should not approve trades, file compliance decisions, move funds, accuse customers, or replace professional judgment.
Guides
Read the existing model guides.
Definition
What Is an AI Model?
The basic idea: learned patterns, inputs, outputs, and the workflow around the model.
DefinitionWhat Does Open Weight Mean in AI?
Model weights, open weight vs open source, and why the license matters before the hype.
ClaudeClaude Fable 5: What It Is, What It Does, and Why Finance Should Watch It
Launch date, specs, pricing, benchmarks, long-running agents, and finance/crypto use cases.
BenchmarksAI Model Benchmark Cost: The Missing Column
How to read model leaderboards through cost, task fit, and finance-style tradeoffs instead of headline scores.