Conversational AI
Conversational AI enables software to understand and reply to people in natural, human-like dialogue.*
Maturity Levels
|
Level |
Name |
Description |
Technology |
Example tools |
|
0 |
Absent |
No bot – Users have no conversational interface and must rely entirely on human-operated channels for assistance. |
None |
|
|
1 |
Scripted FAQ |
Menu-driven – users follow fixed button trees or type keywords and receive single-turn answers; the assistant cannot remember context or deviate from its script. Answers are generally correct for the narrow, pre-written set of questions, but anything outside scope is either refused or wrong. |
Rule/intent engine, predefined responses |
No-code/Low-code Intent engines of Google Dialogflow, RASA, IBM Watson Assistant, Kore.ai |
|
2 |
Context-aware chat |
Multi-turn assistant – Users can ask follow-up questions in the same session; the assistant remembers recent context, clarifies ambiguities, and supports a natural chat flow. Response correctness rises for in-scope topics, but occasional hallucinations still occur. |
Off-the-shelf LLM, retrieval-augmented grounding (simple RAG), tone and persona defined, basic guardrails |
|
|
3 |
Adaptive & personalised |
The assistant offers users proactive suggestions or next steps tailored to their role or history, and handles text, voice or document inputs within the same conversation (multimodal). Personalised replies are accurate for each user role. |
Personalisation, advanced RAG, multimodal, ASR/TTS, advanced guardrails |
ChatGPT (Study), Perplexity AI Custom-built chatbots |
|
4 |
Full autonomy |
The assistant offers a fluid, channel-agnostic experience that moves seamlessly across devices; the assistant improves its responses from live feedback. There’s almost no distinction with a human conversation in terms of tone, flow, empathy and contextual awareness. The system maintains near-perfect factual correctness on its covered domain. |
*A lot of vendors are adding to the hype by practicing “agent washing,” where they market older tools—like AI assistants, robotic process automation (RPA) systems, and chatbots—as if they were advanced agents, even though they don’t actually have significant agent-like abilities.
| AI Maturity Level: Indicates the level the technology vendors claim to have reached in deploying AI solutions that actually work in real-world applications |
Agentic AI
Agentic AI enables software agents to autonomously plan, reason, and use tools to achieve high-level goals with minimal human guidance.*
Maturity Levels
|
Level |
Name |
Description |
Technology |
Example Tools |
|
0 |
Absent |
Manual tasks – The system performs no tool use or automated reasoning; each action is manually executed by a human, with zero automated planning, memory, or task decomposition. |
None |
|
|
1 |
Single-Step Tools |
Macro executor – When a user issues an explicit command, the agent invokes exactly one whitelisted tool, applying minimal rule-based reasoning without further planning or decomposition; the human stays in the loop for every invocation, tasks are simple and atomic, and no memory is retained beyond the call. |
Simple function-calling, deterministic planning, static allow-list |
Cloud platforms:
Open-source frameworks:
|
|
2 |
Limited Multi-Step |
Task assistant – For moderately complex requests, the agent plans and reasons through a short sequence of 2-3 tool calls, breaks the goal into small subtasks, has short-term session memory, and prompts the human for confirmation at branch points; autonomy is still bounded by a pre-defined workflow and the agent operates solo. |
ReAct-style chain-of-thought, vector store for memory, MCP (Model Context Protocol) |
Cloud platforms:
Open-source frameworks to build AI agents:
MCP servers:
MCP clients (applications that support MCP servers) :
Build MCP server (SDK):
|
|
3 |
Goal-Oriented Autonomy |
Co-worker agent – Given a high-level goal, the agent decomposes it into many subtasks, forms a full execution plan, selects appropriate tools, and reasons about fall-backs, using persistent project or user memory; it escalates to humans only on policy violations or ambiguous objectives, executes largely autonomously, and coordinates specialised peer agents or services when needed. |
Planner-executor split, graph-based memory, policy engine, Agent2Agent protocol |
|
|
4 |
Full autonomy |
Digital employee – The agent learns new tools on the fly, spawns and collaborates with sub-agents, and continuously replans to meet open-ended, multi-stakeholder objectives; it applies meta-reasoning to pick strategies, maintains lifelong hierarchical memory, and optimises workflows without human prompting, leaving people to handle only exceptional cases. |
Autonomous tool-schema induction, hierarchical agents, reinforcement learning from real-world outcomes |
No knowledge of tools that operate at this level, although some claims are being made about services to hire fully autonomous “digital workers” |
*A lot of vendors are adding to the hype by practicing “agent washing,” where they market older tools—like AI assistants, robotic process automation (RPA) systems, and chatbots—as if they were advanced agents, even though they don’t actually have significant agent-like abilities.
| AI Maturity Level: Indicates the level the technology vendors claim to have reached in deploying AI solutions that actually work in real-world applications |