Capabilities and Limitations

Knowing what LLMs can and cannot do helps you set realistic expectations and design better workflows.

What LLMs Do Well

Text Generation and Transformation

Writing drafts, summaries, and translations
Reformatting data between structures
Generating variations of content

Analysis and Extraction

Identifying patterns in text
Extracting specific information from documents
Classifying content into categories

Reasoning (With Caveats)

Following multi-step logical arguments
Solving math problems (especially with chain-of-thought prompting)
Making comparisons and evaluations

Code

Writing, explaining, and debugging code
Translating between programming languages
Generating tests and documentation

What LLMs Cannot Do Reliably

Factual Accuracy
Models can state incorrect information with total confidence. They do not have a "knowledge database" — they predict likely text. Always verify facts, especially numbers, dates, names, and citations.

Real-Time Information
Models have a knowledge cutoff. They do not know about events after their training data ends. They cannot access the internet unless specifically connected to search tools.

Perfect Consistency
The same prompt can produce different outputs each time. This is by design (controlled by temperature settings), but it means you cannot expect identical results from identical inputs.

Long-Form Coherence
Models can lose track of requirements, instructions, or earlier content in very long outputs. Breaking complex tasks into steps improves consistency.

Counting and Precise Math
Models often make errors with exact counting, complex arithmetic, and precise numerical reasoning. Use code execution tools for anything requiring exact math.

The 80/20 Rule of LLMs

For most business tasks, an LLM can get you 80% of the way to a good result very quickly. The remaining 20% — verification, refinement, and edge cases — requires human judgment. Design your workflows around this reality rather than expecting perfection.

Practical Guidelines

Task Type	LLM Reliability	Best Practice
Drafting text	High	Use as a starting point, edit for accuracy
Data extraction	Medium-High	Verify extracted data against source
Factual questions	Medium	Cross-reference with authoritative sources
Complex reasoning	Medium	Use chain-of-thought prompting
Precise calculations	Low	Use code execution instead
Real-time information	None	Connect to search or database tools