What Type of AI Is Copilot?

The Confusion Around Copilot's AI Identity

"Copilot" appears in every Microsoft product announcement, every IT budget conversation, and every AI adoption meeting—yet ask ten practitioners what kind of AI it actually is, and you'll get ten different answers. One practitioner on a Gartner community thread put it bluntly: Microsoft has "an interesting marketing naming scheme where everything is named Copilot something."

The confusion starts with the product family itself. "Copilot" isn't one product but a portfolio spanning consumer Copilot, Microsoft 365 Copilot, Windows Copilot, Security Copilot, and Dynamics 365 Copilot, each with different capabilities, pricing, and underlying technology. Microsoft 365 Copilot runs roughly $30 per user per month, while Security Copilot is bundled with E5/E7 licenses at 400 Security Compute Units per month per 1,000 users. When the name covers everything, it tells you nothing.

Then there's the model opacity problem. Microsoft's official pages say Copilot "brings together the latest AI models" but never name them. Third-party analysis suggests OpenAI-based models plus a multi-model approach involving Anthropic's Claude starting with Wave 3 in March 2026—but that comes from a Copilot consulting firm, not Microsoft. Meanwhile, one unverified, dated practitioner claim circulating in the Gartner peer community asserts Copilot "is still running on a Microsoft instance of GPT 3.5 Turbo." Microsoft has never confirmed it.

The taxonomy dispute compounds things. Salesforce argues copilots are "primarily a form of generative AI" that lack autonomous multi-step execution, while Microsoft markets Copilot as combining "agentic capabilities with Work IQ." Both can be true—Copilot is generative AI at its core with agentic features layered on—but the conflicting vendor narratives leave buyers guessing.

Why does this matter for adoption decisions?

• Budget accuracy: at $30/user/month for Microsoft 365 Copilot, knowing whether you're buying a chat assistant or an autonomous agent changes the ROI math.
• Real-world results vary widely: one VP of Operations surveying ~100 users reported only a ~10% efficiency gain, while Lumen Technologies credits Copilot with "$50M in revenue" from four hours saved weekly.
• Security posture differs radically between a reactive text generator and an agent that "takes action across Microsoft 365."
• Vendor claims need scrutiny—Microsoft's Copilot vs. ChatGPT Enterprise comparisons use explicitly simulated demonstrations.

This identity question isn't academic. If Copilot is becoming the "interface" for what Microsoft calls "the apps of the AI era"—agents—then understanding what's underneath determines what you can safely build on top. And as agent ecosystems grow, so does supply-chain risk: security researchers have demonstrated a malicious skill that exfiltrated an entire codebase with only four user interactions. SkillzBench is building a free answer to that trust gap—verifying, security-scanning, and benchmarking agent skills before you deploy them—because "what type of AI is this?" is exactly the question you should never have to guess at.

The Answer: Generative AI at Its Core, With Agentic Layers

So what kind of AI is Microsoft Copilot, really? The short answer: it's a multi-model, generative AI-powered assistant — with agentic capabilities layered on top rather than baked in from the start.

At its foundation, Copilot runs on large language models. According to technical analysis of Copilot's architecture, it uses OpenAI-based models plus a multi-model approach involving Anthropic's Claude technology, beginning with Copilot Wave 3 in March 2026. That makes Copilot less like a single model and more like an orchestration system.

Microsoft's own Copilot product pages describe the "Auto advantage": the system weighs accuracy, speed, and cost for each request to determine the right model and effort level for the task. In other words, Copilot doesn't commit to one model — it picks per request.

Why doesn't Microsoft just name the models? Its official pages don't disclose specific underlying model names, likely for competitive and contractual reasons. You may see practitioner claims that Copilot runs on GPT-3.5 Turbo, but per Gartner peer community discussion, those claims are unverified and dated. Treat any specific model attribution with skepticism.

The taxonomy matters here. Salesforce's framing of copilots holds that they are "primarily a form of generative AI" that generally lack autonomous, multi-step planning and execution. Copilot fits that core definition — but Microsoft is actively adding agentic layers:

• Copilot Autopilot — an always-on personal agent with its own identity, memory, and tool access (currently in Private Preview)
• Agent Builder — create work-grounded agents in minutes using natural language, extendable via Copilot Studio
• Work IQ grounding — Copilot reasons over your company's context and data within existing Microsoft 365 permissions

Microsoft's own analogy ties it together: agents are "the apps of the AI era, with the copilot as the interface." That framing matters if you're building agent workflows — because the apps (skills) an agent loads determine what it can actually do, and published research has found roughly 26% of agent skills carry at least one vulnerability. SkillzBench is building a free answer to that: a directory and MCP connector that lets you discover, verify, and benchmark agent skills before you equip them. Skills, proven — before you wire them into anything as powerful as Copilot.

Grounding and Growing Autonomy: Work IQ, Graph, and Agents

What turns a chatbot into a colleague is context. Copilot answers differently than ChatGPT because it doesn't just generate text — it reasons over your company's actual work data, and increasingly, it acts on it.

At its core, Copilot uses a retrieval-augmented approach. Microsoft's Work IQ layer "grounds Copilot in your company, your team's roles, and the context behind your projects," orchestrating large language models with Microsoft Graph data — emails, documents, chats, and calendars — filtered through each user's existing permissions. Crucially, Copilot operates within the Microsoft 365 boundary, honoring sensitivity labels and data loss prevention policies, and Microsoft states it doesn't train on your data. That grounding-in-place is what separates it from general-purpose chatbots.

The agentic layer is where things get interesting. Copilot Autopilot is described as an "always-on personal agent with its own identity, memory, and access to tools," cloud-hosted and currently in private preview. Agent Builder lets users create personalized, work-grounded agents in minutes using natural language, extendable through Copilot Studio. Microsoft's framing is bold: "Think of agents as the apps of the AI era, with the copilot as the interface."

Not everyone agrees on the taxonomy. Salesforce argues copilots are "primarily a form of generative AI" that lack autonomous, multi-step planning — a conveniently narrow definition from a company selling its own agent platform. The honest synthesis: Copilot is generative AI at its core, with agentic capabilities being layered on top, wave by wave.

Does the investment pay off? The results data is promising, if mostly vendor-sourced:

• Lumen Technologies' CRO claimed that "four hours back each week is worth $50M in revenue over a 12-month period" from Copilot use.
• A VP of Operations surveying roughly 100 Copilot users reported about a 10% average efficiency gain — roughly 45–50 minutes per person per day.
• Pricing runs from $9.99/month for individuals to $23.50 per user monthly for Business and $30 per user monthly for Enterprise (paid yearly, including Copilot Studio).

Practitioner experiences vary — some Gartner community respondents call Copilot "a one trick pony" best suited to meeting summaries, so pilot before you commit.

Microsoft's "agents as apps" analogy opens a broader ecosystem of agent skills, and that ecosystem carries real supply-chain risk: published research found roughly 26% of agent skills carry at least one vulnerability, and security researchers have demonstrated a malicious skill silently exfiltrating an entire codebase. That's why SkillzBench is building a free directory that verifies, security-scans, and benchmarks agent skills — so the "apps of the AI era" you install are the ones you can trust.

Extending Copilot-Style Agents Safely With Agent Skills

Microsoft calls agents "the apps of the AI era," with the copilot as the interface — and like apps, agents get extended with third-party components. Those components are increasingly agent skills: instruction packages (typically SKILL.md files) that teach tools like Claude Code, Codex, Cursor, and Gemini CLI new capabilities. The convenience is real, but so is the risk.

The threat is not hypothetical. OWASP's Agentic Skills Top 10 catalogs the supply-chain dangers of installing community-built skills into agents that hold your credentials and code. Worse, security researchers at Mitiga demonstrated a malicious skill that silently exfiltrated an entire codebase after just 4 user interactions. A helpful-looking skill can be a Trojan horse with almost no friction.

This matters because the agent era Copilot is entering runs on trust. Microsoft's own framing positions agents that "reason over work context" and "take action across Microsoft 365" within enterprise controls — but community skills bypass that governance entirely. Published research cited in the security community has found roughly 26% of skills carry at least one vulnerability, and nobody installing one today can easily tell which.

That verification gap is what SkillzBench is being built to close. The free, independent directory will continuously crawl and dedupe skills from GitHub, then put each one through a verification pipeline:

• Pin every version to a commit SHA, validate it, and check its license
• Security-scan and quality-judge each skill before it earns a listing
• Benchmark every skill in a sandbox with the same model, with and without it, publishing the Lift in percentage points
• Track upstream changes, re-scan, re-benchmark, and notify when a skill you rely on shifts

Every skill will carry a 0–100 trust score built from security, quality, license, maintenance, and community signals. Crucially, Lift is always shown separately from ratings, and money never touches scores — the platform stays free, funded by advertising around its articles rather than pay-to-play listings.

Equipping agents safely is the final piece. A free MCP connector will let you toggle skills and tools per agent, so each agent sees only what you deliberately chose — no accidental attack surface. If you're extending a Copilot-style agent with skills, that combination of verified skills and scoped access is the practical path forward. Skills, proven.