A structured reference module for understanding AI, its categories, tools, token economics, and practical use — prepared for finance and audit professionals.
Designed for one specific task — image recognition, spam filtering, recommendation engines. Most AI tools in use today are narrow AI.
Creates new content — text, images, code, audio. Powered by models like GPT-4, Claude, and Gemini. The fastest-growing category.
Hypothetical AI with human-level reasoning across all domains. Does not yet exist but is the long-term research goal of labs like Anthropic and DeepMind.
How modern AI works: Large Language Models (LLMs) are trained by predicting the next token in a sequence across billions of documents. They learn grammar, facts, reasoning patterns, and context. When you prompt them, they generate a statistically likely continuation — which is why they are fluent but can sometimes hallucinate facts.
Understands and generates human language. Powers chatbots, writing assistants, and question-answering systems. Uses transformer-based LLMs.
Generates or edits images from text descriptions using diffusion models. Used in design, marketing, and visual prototyping.
Writes, reviews, debugs and explains code. Dramatically accelerates software development and is increasingly used in data analytics workflows.
Combines web search with AI reasoning to answer questions with cited sources. Replaces traditional keyword-only search for many queries.
Transcribes speech to text, converts text to voice, and generates music or audio. Key for accessibility, documentation, and media production.
Analyses structured and unstructured data to surface insights, forecast trends, detect anomalies, and automate reporting.
| Tool | Category | Primary Purpose | Best For | Model |
|---|---|---|---|---|
ChatGPTOpenAI |
Conversational | Writing, analysis, coding, Q&A, document summarisation | General professionals, students | GPT-4o / o1 |
ClaudeAnthropic |
Conversational | Long-document analysis, audit reports, finance writing, reasoning | Finance, legal, compliance | Claude 3.5 / Sonnet |
GeminiGoogle |
Conversational | Multimodal tasks — text + images + Google Workspace integration | Google users, researchers | Gemini 1.5 Pro |
CopilotMicrosoft |
Productivity | Embedded AI in Word, Excel, Outlook, Teams — drafts, summaries | Microsoft 365 users | GPT-4 / Azure OpenAI |
PerplexityPerplexity AI |
Research | Real-time web search with cited answers; fact-checking | Researchers, due diligence | Multiple models |
GitHub CopilotMicrosoft / GitHub |
Code | Autocomplete and generate code in IDE; code review, test writing | Developers, data analysts | GPT-4 / Codex |
MidjourneyMidjourney Inc. |
Image | High-quality image generation from text prompts | Designers, creative teams | MJ v6 |
Otter.aiOtter AI |
Audio | Meeting transcription, real-time captions, AI meeting notes | Board meetings, interviews | Whisper-based |
Julius AIJulius |
Data | Upload CSV/Excel and ask questions; auto-charts, statistics | Finance, audit analytics | GPT-4 backbone |
Canva AICanva |
Design | AI-generated presentations, social graphics, brand materials | Business communication | Multiple models |
Notion AINotion |
Productivity | Summarise pages, draft content, extract action items within Notion | Project management, notes | Claude + GPT-4 |
GrokxAI (Elon Musk) |
Conversational | Real-time X/Twitter data + conversational AI; less filtered responses | Social media intelligence | Grok-2 |
Example: "Audit the financial statements for FY2024"
= 8 tokens (roughly 6–7 words per token on average)
Remove pleasantries ("Could you please kindly...") and redundant instructions. A clean, directive prompt uses 30–40% fewer input tokens and often gets a better response.
Instead of pasting a 60-page annual report, extract only the Income Statement and Notes. You get the same quality analysis at 10% of the token cost.
Tell the AI exactly what format you want — "Return a 5-row table with columns: Issue, Risk Level, Recommendation." Structured outputs are shorter and avoid verbose explanatory prose.
Each message in a conversation thread carries the full prior history as input tokens. Start a fresh conversation for unrelated tasks to avoid paying for irrelevant context.
For repetitive tasks (audit checklist review, VAT return analysis), write one precise system prompt once. Cache it if using the API. Reuse across all similar tasks.
Not every task needs GPT-4o or Claude Sonnet. Summarising a meeting transcript is a low-complexity task — use GPT-4o-mini or Claude Haiku (10–20x cheaper) for those.
All major AI providers offer a free tier with limitations and a paid subscription. Enterprise and API pricing is separate and usually more cost-effective for business use above ~50 hours/month. Prices are in USD and reflect 2025 rates.
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Context window | Ideal use |
|---|---|---|---|---|
GPT-4oOpenAI |
$2.50 | $10.00 | 128K | Complex reasoning, vision, balanced cost |
GPT-4o miniOpenAI |
$0.15 | $0.60 | 128K | High-volume simple tasks, classification |
Claude 3.5 SonnetAnthropic |
$3.00 | $15.00 | 200K | Document analysis, audit, finance tasks |
Claude 3 HaikuAnthropic |
$0.25 | $1.25 | 200K | Fast, cheap summarisation, extraction |
Gemini 1.5 ProGoogle |
$3.50 | $10.50 | 1M | Very long documents, video analysis |
Gemini 1.5 FlashGoogle |
$0.075 | $0.30 | 1M | Cheapest per token for bulk tasks |