Stop Asking for AI Agents When You're Not Ready for Them—Here's What You Really Need

AI agents aren’t the default answer to workplace automation. The core message is that AI capability should be matched to the specific task on a spectrum—from simple “ask for advice” chat to fully autonomous systems—so teams don’t overbuild, overspend, or disconnect humans from the work they’re best at.

Instead of asking “Can we use an AI agent for this?”, the better question is “What level does this task actually require?” Most organizations get stuck because they lack vocabulary for the middle ground between chatbots and fully autonomous agents. The result is a binary mindset: either “just chat” or “go all-in on agents.” That framing leads to doom stories and poor implementations, because many workflows don’t need autonomy; they need assistance at the right moment, with the right guardrails.

The spectrum is laid out in six levels. Level 1 is the familiar pattern: an LLM provides advice while the human does the work—typical of common ChatGPT usage and similar tiers. Level 2 is the “co-pilot” stage, where AI suggests actions as a person works. Examples include GitHub Copilot writing code while someone types, or “Cluey” delivering answers during an interview-like flow. This level fits repetitive, well-patterned tasks where the main gain is speed—often described as getting 40–50% faster—while the human remains in control.

Level 3 shifts to a tool-augmented assistant: the chat system can access tools such as web search, calculations, and even asset creation and editing. The emphasis is that this often delivers far more value than an “agentic” setup, because the real multiplier is the number of tools the assistant can use. It’s also positioned as much easier and cheaper to implement than enterprise-grade agent systems. The transcript argues that many teams mistakenly chase “agents” when they actually need tool-enabled workflows—especially for finance, marketing, and product operations.

Level 4 introduces structured workflows with human review. AI performs steps, but people verify correctness—particularly important in high-liability contexts like contract review. A JP Morgan case study is cited as saving hundreds of thousands of hours annually, with the caveat that the savings come from workflow design and scale, not magic.

Level 5 is semi-autonomous: AI handles routine cases independently while humans review exceptions and edge cases. Customer success is offered as a strong fit because customer complaints often map to a manageable distribution of scenarios. Level 6 is fully autonomous, where AI does everything and humans only monitor metrics. The transcript warns that full autonomy is hard, expensive, and brittle—citing examples like self-checkout attempts and the challenge of scaling self-driving cars across cities. The practical takeaway: aim for “almost all the value” without requiring full automation everywhere.

The recommended next step is to evaluate a recurring task using concrete questions—frequency, consistency, error impact, data location, and speed requirements—to determine the appropriate level. If unsure, start with a level that can be tested without stakeholder approval, with a strong suggestion that many real-world needs land at Level 3. The overarching goal is better AI implementations that improve human work rather than replace it.