A Field Guide to AI and Agents
- Increment
- 45 minutes ago
- 3 min read
The internet is full of AI maturity models, positioning statements and frameworks, most using terms like "agent" and "agentic" loosely and interchangeably. Even within Microsoft's own ecosystem, the definition shifts depending on the product: in SharePoint, an "agent" might just be a glorified search box; elsewhere, it's a system that plans, uses tools and pursues a goal with real autonomy.
That confusion is the real starting point for AI governance. Before any business can secure or manage its agents, it needs a shared, practical way to describe what it's actually looking at. That's why Increment built the Field Guide to AI and Agents, a grounding framework that formed the centrepiece of one of our recent webinars, Demystifying Agent 365, and the subject of our new video explainer.
Six generations, one spectrum
Rather than a single definition, the Field Guide maps AI capability as a spectrum of six generations, each defined by how much planning, tool use, autonomy and persistence is involved:
Gen 0 – Conversational AI: a question in, an answer out. No tools, no planning, no memory beyond the chat. Think a standard Copilot chat.
Gen 1 – Grounded Assistant: given a task ("find the latest leave policy"), it uses foundational tools to complete it, with minimal planning and session-bound persistence. Microsoft 365 Copilot Chat, grounded via Microsoft Graph, sits here.
Gen 2 – Action Agent: delivers a specific, bounded outcome, for example, pulling every open opportunity for a customer from a CRM. Some planning, some tool use, persistence tied to the workflow. The Dynamics 365 Sales agent is a good example.
Gen 3 – Agentic AI: you set a goal, not a task, "investigate why onboarding completion is dropping, graph it, and recommend fixes", and the agent decomposes the work itself, with high planning and high autonomy. This is where tools like Copilot Work mark a genuine step-change from everything before it.
Gen 4 – Autonomous Agent: operates continuously against a standing objective ("maintain inbox SLA compliance"), not triggered by a prompt or event, with high planning, high autonomy and a wide toolset.
Gen 5 – Multi-Agent / Agentic Workforce: a strategic outcome is handed to a coordinated team of specialised agents that decompose and manage the work between themselves. No mainstream Microsoft product delivers this out of the box yet, but organisations are already building toward it with Copilot Studio and Foundry.
Why the generations matter more than the definitions
The Field Guide isn't an academic exercise, it's there to focus attention on where the real risk sits. From Gen 3 onward, agents are goal-oriented and operating with meaningful independence, which is exactly where the familiar governance questions get harder to answer: What is this agent doing? What data can it reach? Who's accountable when something goes wrong?
Those questions map to four consistent risk areas: visibility (do we even know how many agents exist and who's using them?), accountability (is ownership clear?), lifecycle (do we have a strategy to manage agents from creation to retirement?), and security (can we catch unintentional or malicious data loss before it happens?). The webinar walked through real examples of well-meaning agent builders who exposed sensitive data simply because nobody had applied these disciplines yet, proof that this isn't a hypothetical problem waiting in the future.
If you want to go further, our Getting Agent Ready with Agent 365 engagement is a structured two-week program, four facilitated workshops assessing your current maturity, demonstrating Agent 365 in action, and delivering a prioritised roadmap tailored to your business. Reach out to get started.
