6 Levels of AI Delegation: A Framework for AI Management

Note: This blog post was written by humans; AI was used only to correct grammar and provide scenario examples.

As managers know delegation is an important part of delivering results and growing their team. At Metal Toad our managers go through training including the 5 Levels of Leadership. One of the key parts we try to drive home are the 6 levels of delegation.

1. Look into it. Report. I’ll decide what to do.
2. Look into it. Report alternatives with pros and cons and your recommendation.
3. Look into it. Let me know what you intend to do, but don’t do it unless I say yes.
4. Look into it. Let me know what you intend to do and do it unless I say no.
5. Take action. Let me know what you did.
6. Take action. No further contact required.

It's important to be clear and direct about what you expect because not all people have the same skill set and not all tasks are created equal.

As AI has grown in the workplace we’ve noticed that people are struggling to relinquish control of tasks to AI Agents or giving too much control to AI Agents. Repurposing the 6 levels of delegation actually provides a good framework for people working with bots.

We can provide some examples of these but before we do. It's important to be clear and direct with the AI Agent what you are expecting. If you just want ideas. You have to explicitly tell the agent to do so because they are eager to help.

Level 1: Look into it. Report. I’ll decide what to do.

The first level of delegation is when the primary goal is simply to gather information. This allows you to use your experience to reach a conclusion and determine the next step. Although most AI models tend to push you toward the next action, giving them more control over decisions, it is crucial to spend time at this first level. This ensures you can digest the context without rushing into action, especially during the early stages of a project. Remember: if you are off by just one centimeter at the start, the gap will be huge by the end.

The feature that you need at this stage is basically the power of gathering and organizing information, and this can be easily done today with almost any mainstream models, like ChatGPT or Gemini.

One tool we are using at Metal Toad is Kiro, an agentic AI software with an IDE and CLI. It assists with spec-driven development and has proven to be quite interesting; instead of prioritizing speed for its own sake, it helps you tailor specifications and gives you the time to actually think about the problem before delegating it to the AI.

Scenarios where this level of delegation is recommended:

• Early-Stage Project Discovery: When kicking off a new initiative, exploring market trends, or conducting competitor analysis. The AI can aggregate the raw data, giving you the necessary context to define the project's scope and direction without committing to a path too early.
  
• High-Stakes Strategic Decisions: When the consequences of a mistake are severe (e.g., legal compliance, major financial investments, or core business pivots). The AI can build the brief, but human intuition, experience, and accountability are strictly required for the final call.
• Complex Problem Diagnosis: When troubleshooting a highly nuanced issue, like a sudden drop in product engagement or a convoluted codebase bug. The AI can parse the logs, user feedback, or data tables to highlight anomalies, leaving you to connect the dots and determine the root cause.

Level 2: Look into it. Report alternatives with pros and cons and your recommendation.

This involves making a decision that is too important to be left to an AI agent, yet complex enough that you need help seeing it for what it is.

Just as when you are delegating to a new team member, you want them to think about the topic and bring their perspective to the table. Although this is only the second level of delegation, in a human-AI relationship, it is crucial to remember that you are accountable for what happens. Despite AI's capabilities, always remember that a computer should never be held accountable for decisions.

This level of delegation represents the early stages of Generative AI well, where the main value was bringing different perspectives with little action or agency. Many people use AI at this level, which is an excellent use case; popular tools like Gemini or Claude have most of their users here. However, it is also natural, once you see that the AI provides good solutions, to want to move to the next level.

Scenarios where this level of delegation is recommended:

• Vendor or Tool Selection: When you need to choose a new software platform, agency, or service provider. The AI can analyze the market, generate a comparison matrix with pros and cons for your specific use case, and recommend the best fit, but you make the final purchasing decision.
• Process Optimization: When an internal workflow is broken or inefficient, and there are multiple ways to fix it. The AI can map out alternative workflows, highlight the trade-offs (e.g., speed vs. cost, automation vs. personalization), and suggest the most balanced approach.
• Content and Creative Strategy: When planning a marketing campaign or product roadmap. You can ask the AI to pitch three different creative angles or strategic directions, complete with the potential risks and rewards of each, leaving you to choose the one that best aligns with your brand's voice and vision.

Level 3: Look into it. Let me know what you intend to do, but don’t do it unless I say yes.

Although “AI Agents” is becoming a buzzword in the AI world, what actually classifies an “agent” is that they have some level of agency, which means that they can perform action, besides just delivering answers. At this stage we are giving AI the agency to actually perform actions. Not a full agency but a controlled one.

When it comes to this level of delegation for AI, you’ll need a slightly more complex tool than initial basic chat interfaces. Some tools that come into play are: n8n.io, ChatGPT agent and Amazon Bedrock Agents.

Now is when another buzzword comes to be useful: MCP.

MCP stands for Model Context Protocol, Developed by Anthropic, is an open standard that allows AI assistants to connect seamlessly to data sources and tools (like Google Drive, Slack, or GitHub). It acts as a "universal connector" so you don't have to build custom integrations for every new AI model.

The idea of MCP is that it provides documentation about the tool to the AI like a manual. Then when your AI needs to perform tasks it uses natural language to decide when and how to use the tool with out the need to create specific API calls.

However, more important than connecting several MCPs is having the right controls to ensure that nothing happens until you explicitly say, “Yes, do it!”. Remember AI is eager to be helpful and show value, and sometimes the temptation of delegating too much to AI is high, so you need to choose the right tools that give you clear and good guardrails.

Kiro is an interesting tool for that. Once you have clarity about the problem and have provided the AI with sufficient context (1. Look into it. Report. I’ll decide what to do) and create clear requirements for your project (2. Look into it. Report alternatives with pros and cons and your recommendation), you can use the “disable autopilot” feature. You are in control of every action the AI agent takes, and it will not act until you explicitly give permission.

Scenarios where this level of delegation is recommended:

• Code Refactoring and Deployment: When using an IDE-integrated agent like Kiro or connecting via GitHub MCP. The AI can write the code, generate the tests, and prepare the pull request, but a human developer must review the logic and explicitly approve the merge or execution to prevent breaking the build.
• Customer Communication and Support: When integrating AI with your CRM or inbox (e.g., via n8n or Slack). The AI can read incoming customer inquiries, gather the relevant data, and draft highly contextual replies or support tickets. However, the drafts sit in a queue waiting for your final "Yes, do it!" before anything is actually sent to the client.
• Data Management and System Cleanup: When tasked with identifying duplicate records, organizing messy Google Drive folders, or updating database entries. The AI maps out exactly which files it will move or which records it will merge, allowing you to verify the changes won't cause data loss before granting execution permission.

Level 4: Look into it. Let me know what you intend to do and do it unless I say no.

As a manager, you need to build enough trust, understanding of skills and the correct way to manage a direct report, before letting them act without your explicit approval.

It is very similar to AI agents. You need to spend some time with your AI toolbox to ensure that you can trust the AI agent to access important information and tools and make decisions. Remember: your bot's decision is your decision, so be careful.

An important feature in this scenario is the ability to save checkpoints, especially when the project involves coding in a repo, or to maintain clear logs where you can easily troubleshoot and roll back changes.

Kiro has an 'autopilot' feature, where it continues with development while asking for permission before performing important commands. You can 'trust' a command so it won't ask for permission next time. This configuration is saved under settings and can be changed at any time. Also, it saves checkpoints before modifying any code, so you can easily roll back.

When using workflow tools like n8n.io, zapier.com, or make.com, it is a good idea to set up a simple notification system with: (1) a start notification, (2) an 'in-progress' status, and (3) the final result, while storing the data in an easy and accessible log. For notifications, we usually use a specific Slack channel, since this gives us full visibility while allowing us to silence it.

Scenarios where this level of delegation is recommended:

• Routine Code Maintenance and Updates: When managing software dependencies, fixing minor, well-documented bugs, or performing standard refactoring. The AI (like Kiro) can state its intent to update packages or clean up code, create a checkpoint, and proceed. If something breaks, you simply roll back using the saved state.
• Data Syncing and ETL (Extract, Transform, Load) Workflows: When moving data between platforms using tools like Zapier or n8n (e.g., syncing new CRM leads to an email marketing tool). The AI workflow triggers automatically and sends a Slack notification stating it is processing a batch of records, allowing you to pause the workflow if you notice a data formatting error, but otherwise requiring no manual effort.
• Standardized Content Distribution: When repurposing approved content across multiple channels. The AI can analyze a new blog post, draft social media snippets, and schedule them in a tool like Buffer or Hootsuite. It pings you with the schedule: "Posts are queued for Tuesday. Cancel if needed."

Level 5: Take action. Let me know what you did.

Okay, we’ve been careful so far, but we have now reached a stage where this could be harmful. To delegate to an AI at this level, the scope in which the AI has freedom needs to be very limited; additionally, the work being performed should have low stakes. Again, if a computer should never be held accountable for a management decision, you should not give in to the temptations of AI.

If you reach this level of delegation, you should ideally already be comfortable with the features and skills of your AI agent, and particularly with your own ability to manage this level of trust.

The recommendation here is to store only the final results in an accessible location and run the workflows in parallel with your other activities.

Scenarios where this level of delegation is recommended:

• Routine Data Scraping and Aggregation: When tracking competitor pricing, industry news keywords, or public market data. The AI runs quietly in the background, gathers the information, drops a formatted summary into a shared Google Drive or Notion folder, and simply logs that the daily update is complete.
• Internal Meeting Transcription and Summarization: When handling internal, non-confidential team syncs. The AI automatically joins the call, transcribes the conversation, extracts action items, and saves the final document to your project management tool, sending a brief "Notes are ready" notification.
• Non-Destructive File Archiving: When managing digital clutter. The AI runs a weekly script to move files older than 90 days from a "Downloads" or "Temp" folder into an organized Archive, leaving a simple log of what was moved so you can always find it later if needed.

Level 6: Take action. No further contact required.

This is the final and highest level of delegation. In a usual managerial scenario, a direct report not only possesses the necessary skills and expertise but also has the manager's trust to truly be the 'king of their realm.' Since this level of delegation is typically rare or context-specific, the same logic should apply to the AI agent.

This level of delegation is best suited for repetitive and thankless tasks, those that do not alter the company's team dynamics or involve critical decision-making. However, managerial and business decisions still require thoughtful human consideration. At least for now.

Conclusion

Even with this we don’t trust people to do everything without supervision. In part this is because people can be prone to errors but it goes beyond that. Some tasks are too important to let any one person or agent make all the decisions. This could be because a mistake will reflect badly on the person delegating all the way to it could cost lives if there is a mistake.

AI Agents are increasingly common in software. The functionality is becoming more robust at a pace that hasn’t been seen before. Some people would even have you believe it can handle every complex task and one day this may someday actually be true. Until then consider how you should delegate tasks to AI to maximise their usefulness and reduce risk.