Everyone's racing to build "AI agents," but most companies are thinking about this wrong. They're jumping from basic chatbots (Level 1) straight to talking about fully autonomous agents (Level 6), missing the most practical and valuable levels in between.
The problem isn't AI capabilities. It's the binary thinking: basic chat vs. fully autonomous agents. Nate Jones (video below) suggests six distinct levels of AI autonomy you can consider, and the business value of each depends entirely on matching the right level to your specific problem.
The companies getting the best results aren't chasing the highest level of autonomy. They're asking: "What level of autonomy does this specific business challenge actually need?"
Think of AI implementation as a spectrum of autonomy. Each level represents how much decision-making and execution you delegate to AI versus retain with humans. The business impact varies based on your use case, not the level itself. Here's the framework:
Level 1: The Adviser | Zero autonomy
What it is: AI provides advice and answers. Humans do all the work.
Autonomy level: AI has no agency. It only responds when asked. You make every decision, execute every action.
Real-world application: Using ChatGPT to brainstorm email copy, debug code, or research market trends. The quality depends heavily on your prompting skill.
Business value: This is where many people starts. It's useful for individual productivity and learning AI capabilities, but the impact is limited to how well humans can translate advice into action. The ROI depends entirely on who's using it and how skilled they are at prompting.
Why it matters: If your entire "AI strategy" stops here, you're leaving massive value on the table. Not because Level 1 is "bad," but because you're not leveraging AI's ability to take action.
Level 2: The co-pilot | Predictive assistance
What it is: AI suggests next steps as you work in real-time.
Autonomy level: AI anticipates what you're trying to do and offers to complete it. You're still driving - you start each action, frame each task. AI just helps you find "an extra gear" by predicting patterns.
Real-world application:
Developers using GitHub Copilot for code completion
Writers using AI-powered autocomplete
Analysts getting suggested formulas in spreadsheets
Business value: This is the first level where you see measurable productivity gains without changing workflows. Users stay in their tools, AI accelerates what they're already doing. The value comes from reducing repetitive typing and reducing errors, not from AI making decisions.
Why it matters: Great for high-repetition tasks with known patterns. But you're still the one thinking through the problem. AI just speeds up execution.
Level 3: Tool-augmented assistant | Task completion
What it is: AI accesses tools, APIs, and data sources to complete multi-step tasks.
Autonomy level: AI can execute entire tasks, not just help you execute them. You set the goal, AI figures out which tools to use and in what sequence. This is about capability, not just speed - AI can now do things you might not have had time to do manually.
Real-world application:
AI that can read customer data, generate reports, and draft responses
AI that searches internal docs, synthesizes findings, and creates presentations
AI that pulls data from multiple systems, analyzes patterns, and recommends actions
Business value: This is what Nate calls "one of the most significant jumps in value." Instead of just advising, AI actually completes tasks end-to-end. The value isn't from predetermined ROI percentages - it's from unlocking work that wasn't economically viable before. You can now afford to analyze every customer conversation, research every prospect, or draft personalized responses at scale.
A critical insight: Most organizations jump from Level 1-2 directly to talking about "agents" (Level 6), completely missing this level. But this is where you get transformative capability gains without the complexity and risk of full autonomy. And according to Nate, it's "10 times, 100 times, maybe a thousand times easier than an enterprise agentic system to install." This is also the approach I believe we should be taking - at least in the near future, in combining AI and human capabilities.
Level 4: Structured workflow automation | Choreographed collaboration
What it is: AI executes predefined workflows. Humans review at key decision points.
Autonomy level: AI handles multi-step processes, but humans are deliberately placed in the workflow at critical junctures. This is choreographed work: AI does a step, human reviews, AI continues. The structure matters - you're designing exactly where human judgment is needed.
Real-world application:
Customer support: AI handles intake, research, and draft responses; humans approve before sending
Document processing: AI extracts data, categorizes, routes for approval; humans review exceptions
Sales ops: AI qualifies leads, researches accounts, drafts outreach; reps review and send
Business value: This is about operational transformation, not just productivity gains. You're redesigning how work flows through your organization. The ROI can be enormous (JP Morgan saved 300,000+ hours annually on contract review), but that's a function of scale and process design, not the AI itself.
Why it matters: Nate emphasizes that "people sleep on this piece" because they think full autonomy is the goal. But Level 4 is often the RIGHT answer for high-stakes processes. You get massive efficiency gains while keeping your best humans engaged at the points where their judgment matters most.
The key philosophy: The goal isn't to remove humans from the work. It's to let "your best humans touch the work more, not less" - just at the moments where they add the most value.
Level 5: Semi-autonomous | Exception-based operations
What it is: AI handles routine cases independently. Humans only review exceptions.
Autonomy level: AI operates on its own for the majority of cases. Humans are no longer in every loop - they only get pulled in when something falls outside normal patterns. This requires AI to know what it doesn't know and escalate appropriately.
Real-world application:
Customer service: AI resolves 80-90% of inquiries; only escalates complex cases
Compliance or fraud detection review: AI approves standard transactions; flags unusual patterns for human review
Content moderation: AI handles clear-cut cases; humans review borderline content
Business value: This is where you transition from "AI helps humans work faster" to "humans handle exceptions while AI runs operations." The business value can be transformative for high-volume, low-variability processes. But the ROI depends on getting the exception criteria right - if AI escalates too much, you lose efficiency; too little, you lose quality.
Why it matters: This level fundamentally changes capacity economics. You're no longer constrained by headcount for routine work. But it requires high confidence in AI accuracy and clear criteria for exceptions.
The requirement: You need robust monitoring and the discipline to keep humans engaged with the work even when they're only seeing edge cases. Your best people shouldn't feel disconnected from operations.
Level 6: Fully autonomous | Lights-out operations
What it is: AI handles everything. Humans monitor metrics and intervene only if metrics degrade.
Autonomy level: Complete delegation. AI runs the entire operation. Humans aren't in the workflow at all - they watch dashboards and step in only when something breaks. This is the "completely autonomous Borg-like thing" people imagine when they hear "AI agents."
Real-world application:
Algorithmic trading: AI executes trades based on market conditions
Dynamic pricing: AI adjusts prices in real-time based on demand
Infrastructure management: AI provisions, scales, and maintains cloud resources
Business value: The value is highly variable and use-case dependent. For some applications (high-frequency trading, infrastructure scaling), full autonomy is essential because milliseconds matter and human intervention isn't practical. For others, the complexity and risk outweigh the marginal gains over Level 4 or 5.
Why it matters (and why it's not always the goal): This is "what people think AI agents are," but Nate's key insight is that it's often not necessary or even desirable. Full autonomy only makes sense when:
Volume is extremely high and variability is very low
Speed requirements make human review impractical
The cost of errors is low relative to the cost of human review
You have sophisticated monitoring to catch failures before they compound
The risk: As we discussed in the leadership section, you're delegating judgment, not augmenting it. You need clear criteria for when AI should escalate, and you need humans who stay engaged enough to recognize when the autonomous system is optimizing for the wrong thing.
The future isn't "AI agents" vs. "no AI agents." It's a spectrum of autonomy, with different levels solving different business challenges.
Stop asking "should we build AI agents?". Start asking:
What level of autonomy does this specific problem need?
Where should humans stay in the loop?
What are we trying to enable our best people to do more of?
A company might use Level 3 for creative work requiring judgment, Level 4 for high-volume operational processes, Level 2 for repetitive coding tasks, and Level 6 for specialized technical automation. Different problems need different levels of autonomy.
What level does your specific business challenge actually need?
Watch the video below to see Nate Jones' overview of the framework.