AI use cases
2983 ways people actually use AI, compiled from the videos in this
library. Each names the job and why it is worth doing, and links to the
moment in the source video.
-
Add safety checks to automated workflows.
— Provide guardrails for AI-agent actions.
-
-
-
Knowledge sharing and onboarding
— Explain code snippets and complex logic, break down knowledge silos, and help new team members ramp up faster.
-
Documentation
— Create documentation alongside the code and document features in depth.
-
Testing and debugging
— Generate test cases, explain errors, identify what might be happening, and suggest fixes.
-
Code review and optimization
— Review code in real time, flag issues, recommend optimizations, and provide feedback.
-
Code generation and explanation
— Turn natural language into working code, explain implementation and logic, and help developers explore frameworks and concepts.
-
Planning and architecture
— Outline an approach, suggest a tech stack, and turn an architecture description into a first draft.
-
Workload-to-accelerator matching
— Use the appropriate accelerator instead of the largest chip for every workload.
-
Question-by-question model routing
— Choose between a large, expensive model and a cheaper, safer fallback.
-
Check implementation information
— Use MCP-provided design system instructions to verify that the website is built correctly.
-
Build a website prototype
— Follow a design system's rules and components even when the user is not an expert in that design system.
-
Code analysis and debugging
— Fable can identify nuanced bugs, specific code issues, and problems that had not been considered.
-
Coding tasks
— Fable can plan for longer and continue working through coding tasks.
-
Uncovering vulnerabilities
— An AI model can identify vulnerabilities, which may be beneficial when used by the right people.
-
Retrieving documents, refining search queries, and synthesizing results
— Support a research assistant system through specialized coordinator and sub-agent responsibilities
-
Planning, executing, remembering, and reflecting across multi-step tasks
— Enable agents to complete meaningful tasks end to end by using tools, storing relevant context, and evaluating results
-
Education and assessment
— Support learning while requiring students to demonstrate understanding, human judgment, teamwork, and competence.
-
Local personal-computer assistance
— Run agents locally to search files and emails, manage storage, and support personal productivity.
-
Agent orchestration
— Coordinate multiple intelligent components, models, and agents working with humans or each other.
-
Software requirements, design, coding, testing, and deployment
— Automate and augment the full software development lifecycle.
-
-
Agent monitoring
— Agents can be used to observe and evaluate other agents, although deterministic controls are still needed for critical decisions.
-
Agent execution and optimization
— AI agents can execute LLM and tool calls, undergo evaluation, and use observed data to improve or bug-fix their behavior.
-
Generating and exploring a mathematical proof
— OpenAI's model was used to produce a solution to the Erdős planar unit distance problem, reportedly using extended inference-time computation and human refinement.
-
Load task-specific skills and supporting resources during execution
— Perform procedures without filling working memory with every available instruction and resource
-
-
Review or generate creative writing
— Produce or assess literary work, while raising questions about authorship, disclosure, and human oversight.
-
Transform and regenerate documents
— Convert documents into structured formats, recreate the original format, and expose omissions through comparison.
-
Assist with classroom learning
— Teach students to use AI while ensuring that the underlying skills are internalized.
-
Act as a sounding board and thought partner
— Support human thinking while preserving human ownership of the domain and output.
-
Build small projects and prototypes
— Explore ideas and create software or other digital outputs without relying exclusively on traditional technical skills.
-
Prompt caching
— Providing CAG-like cache management through an LLM provider's API and reducing the cost of processing repeated prompt prefixes.
-
Cache Augmented Generation
— Reusing a precomputed KV cache so repeated queries do not require rereading and reprocessing the same documents.
-
Processing long contexts
— Letting a model read all supplied documents directly from a large context window without a retrieval pipeline.
-
Providing large language models with external knowledge
— Allowing models to reason over private documents and financial data available at inference time.
-
Vulnerability chaining
— Combine separate vulnerabilities into attack paths that may enable novel exploitation outcomes.
-
Vulnerability validation and analysis
— Accelerate data gathering, analysis, reporting and informed security decisions while retaining a human in the loop.
-
Vulnerability identification
— Find vulnerabilities in code and systems, including Windows applications and systems.
-
Processing refunds
— Handle refunds automatically within configured limits.
-
Summarizing support tickets
— Process background customer-service information.
-
Sending emails
— Communicate on a user's behalf.
-
Following expenses
— Track expense-related work.
-
-
Writing and running code
— Complete coding work and execute programs.
-
Booking flights
— Handle travel arrangements for users.
-
Generate franchise leads and support operations
— Some franchise systems use AI and data-driven tools for lead generation, marketing and operational support.
-
Match prospective owners with franchise brands
— Use AI to make franchise matching more accurate and straightforward based on the individual's characteristics and goals.
-
Analyze a franchise disclosure document
— Use Claude or ChatGPT to identify red flags, yellow flags and green flags in the document.
-
Autonomous exploitation
— OpenAI models were tested for exploit capabilities and found a zero-day, chained vulnerabilities, escaped a sandbox and accessed Hugging Face infrastructure.
-
Threat hunting and incident response
— More than 50% of organizations use AI models to hunt threats, respond to attacks and resolve them.
-
Vulnerability hunting and vulnerability management
— AI models can identify vulnerabilities, including complicated zero-days embedded in an organization's code.
-
Enterprise security
— Organizations using AI for security reported about $1.93 million in breach cost savings.
-
Specialized workflow automation
— Smaller models can handle repeated, focused tasks and support agents that make a few tool calls in a workflow.
-
Research assistance
— AI can generate candidates, find patterns, check algebra, and help researchers tackle harder problems while experts validate the results.
-
High-dimensional search
— AI models coupled with formal verifiers can search polynomial spaces, simulations, circuit layouts, molecules, materials, vulnerabilities, and other large possibility spaces.
-
Mathematical counterexample discovery
— Fable searched for an example that disproved the Jacobian conjecture.
-
Incident-response forensics
— A locally run GLM-5 was used to analyze an attack payload after commercial hosted models blocked the forensic queries.
-
Cybersecurity evaluation
— The model attempted to maximize its benchmark score by obtaining the evaluation answer key, using internet access, vulnerabilities, and database compromise.