“What are the 4 types of AI? And why do some guides list 7?” Different models group AI by capability (what it can do) or architecture (how it works). This plain-English explainer maps both frameworks, shows real business examples, and gives you a quick chooser to decide which type you actually need. For a broader primer, see What Is AI for Business? and ready-to-use prompts in Prompt Playbooks for Teams.
Quick Summary
Use the 4-type model to explain capability levels to non-technical teams. Use the 7-type model when you need more nuance for training or governance. Most SMEs today deploy Narrow AI systems (classification, generation, retrieval-augmented chat) with guardrails and approvals.
- 4 types: Narrow (Weak), General (AGI), Superintelligent (theoretical), plus common cognitive buckets (Reactive, Limited Memory).
- 7 types: Reactive, Limited Memory, Theory of Mind, Self-Aware, plus scope categories like Narrow/General/Superintelligent.
- Practical takeaway: Choose by job-to-be-done; pair with privacy, citations, and approvals for safe rollout.
Manager Mode – the 4-Type model (capability levels)
Use this table when you need a simple, executive-friendly story. It frames AI by how broadly it can reason and act.
The left column keeps it simple for non-technical audiences. The right column anchors expectations with everyday examples you can pilot in weeks.
| Type | Plain-English definition | Examples you can ship |
|---|---|---|
| Narrow AI (Weak) | Excellent at a specific task with clear bounds | Email drafting with JSON schema; support macros via private RAG; invoice field extraction |
| General AI (AGI) | Human-level capability across many tasks (not here yet) | N/A for SMEs—plan governance and evaluations instead of waiting for AGI |
| Superintelligent | Surpasses human capabilities broadly (theoretical) | Not an implementation target; keep risk policies current |
| Common cognitive buckets* | How systems reason/remember (Reactive, Limited Memory) | Chat assistants with short/long context; agents with tools + approvals |
*We expand these in the 7-type model below.
Builder Mode – the 7-Type model (how systems behave)
This model splits by how an AI reasons and remembers. It’s useful for training, prompts, and guardrails.
Map features to rollout risks. Anything that writes to systems or customers should use approvals, citations, and logs.
- Reactive – responds without memory (classic rule-based or small models). Use for: simple routing, keyword triggers.
- Limited Memory – uses short-term context/history. Use for: chats with recent messages, summarization of uploaded docs.
- Theory of Mind* – models user goals/intent. Use for: assistants that adapt tone/level; keep transparency notices.
- Self-Aware* – a thought experiment, not a practical category for SMEs.
- Narrow – specialized task models (classify, extract, generate). Use for: most business automations today.
- General (AGI)* – not deployable today; keep governance plans up to date.
- Superintelligent* – theoretical; include in risk conversations, not roadmaps.
*Conceptual categories for education and policy discussions; not deployable products.
Quick Chooser – which “type” do you actually need?
Pick the smallest, safest option that meets your acceptance criteria, then add controls.
Use this to avoid over-engineering. Start narrow; layer evidence and approvals; then scale.
| Job-to-be-done | Recommended “type” | Controls |
|---|---|---|
| Draft emails, rewrite text | Narrow + Limited Memory (chat) | JSON draft schema; editor sign-off |
| Answer policy questions | Narrow + RAG (Limited Memory) | Source-only prompts; citations required |
| Multi-step actions across tools | Agentic (Limited Memory + tool use) | Approvals on risky verbs; audit logs |
| OCR → extract fields → validate | Narrow extractor + validators | Strict schema; idempotency; DLQ |
Prompts & Frames You Can Copy
These keep outputs predictable and safe across both models. Adapt to your stack.
1) JSON drafting frame (Narrow, Limited Memory)
System: Return ONLY valid JSON matching the schema.
User: Draft a {{doc_type}} for {{audience}} about {{topic}}. Tone={{tone}}, length≈{{words}}.
Schema: {"title":"string","intro":"string","sections":[{"heading":"string","bullets":["string","string"]}],"cta":{"text":"string","url":"string|null"},"notes":["string"]}
Return ONLY JSON.
2) Source-only answer frame (policy/support)
System: Answer ONLY from provided sources. If insufficient, return {"answer":null,"reason":"no_sources"}.
User: Question: {{question}}
Sources (title,uri/url,excerpt,date): {{sources}}
Schema: {"answer":"string|null","citations":[{"title":"string","uri_or_url":"string","date":"YYYY-MM-DD"}],"confidence":"low|medium|high"}
Return ONLY JSON.
3) Agentic action frame (propose → approve → execute)
System: Propose actions as JSON; risky verbs require approval.
User: Task: {{task_text}} • Tools: ["crm.update","email.send","rpa.run_job"]
Schema: {"goal":"string","steps":[{"tool":"string","inputs":{},"reason":"string"}],"proposed_action":{"summary":"string","risk":"low|medium|high","needs_approval":true}}
Return ONLY JSON.
FAQ — 4 types vs 7 types of AI
What are the 4 types of AI?
Most business guides use: Narrow (today’s practical systems), General (AGI), Superintelligent (theoretical), and cognitive buckets like Reactive/Limited Memory for how systems behave.
What are the 7 types of AI?
A more granular teaching model: Reactive, Limited Memory, Theory of Mind, Self-Aware, plus scope categories Narrow, General, and Superintelligent.
Which model should my team use?
Use 4-type for exec briefings and roadmaps. Use 7-type for training, risks, and prompt/workflow design.
Do we have AGI today?
No. Plan governance, evaluations, and approvals around Narrow systems that solve specific jobs safely.
Final thoughts
Both models describe the same landscape at different zoom levels. Start narrow, pair with citations and approvals, and measure outcomes. When everyone shares the same simple vocabulary, projects move faster and stay safer. For more context, see What Is AI for Business? and ready-made templates in Prompt Playbooks for Teams.
AI Tools Business is independent. We test tools hands-on and publish results with citations or screenshots where relevant.
Editorial safeguards
- Claims verified by a second reviewer before publication.
- Changes and price updates are date-stamped and appended.
- We may use affiliate links - rankings are never paid.
