You use AI every day, but you still start every step, move every file, and remind it what to do next. The five levels in Video 4 help you decide which part of that work you could hand over.
This is a practical framework for the video, not a formal certification or a ranking of people. You can use level 1 for an email and level 4 for a weekly report. Choose the amount of delegation that fits the task.
The five levels at a glance
Follow one example through the guide: keeping your team informed about five competitors.
| Level | What it looks like |
|---|---|
| 1 · You + a chatbot | You paste an announcement and ask for a summary. |
| 2 · You + an agent | You ask for a sourced competitor report and review the draft. |
| 3 · You + a team of agents | Separate agents research different competitors; a lead combines their findings. |
| 4 · Recurring work | The report runs every Monday and arrives ready for review. |
| 5 · Agents as teammates | An agent handles incoming competitor questions and flags relevant changes within an agreed remit. |
The useful question is: which decisions and handoffs still need you? Modern chat products can use tools too, so the levels describe how you work, rather than five separate apps.
Official reference: How agents and workflows work
Which tool do I open?
| Your next step | Where to start |
|---|---|
| Level 1 · Ask and refine | Claude chat |
| Level 2 · Delegate a complete task | Still Claude chat, in the merged experience |
| Level 3 · Split work between agents | Claude Code · walkthrough video coming |
| Level 4 · Repeat work automatically | Claude Code + Routines · walkthrough video coming |
| Level 5 · Work alongside a team | Slack with Claude Tag, or Teams with a configured agent |
Checked 11 October 2026. These are practical starting points, not exclusive routes: a simple recurring task can also be scheduled directly in Claude. The sections below show the first action for each level.
Official reference: Claude chat and task capabilities
Official reference: Claude Code Routines
1. You ask, AI answers
How to try it · Claude chat
Open Claude, start a conversation, and paste your question or text. Ask for a summary, rewrite, or explanation, then refine the answer with follow-up messages.
Official reference: One Claude conversation for questions and tasks
You guide the work one exchange at a time. This is useful when you want to think together, refine wording, or understand something before deciding what to do.
Try it with: a difficult email, a meeting agenda, or a confusing document. For our competitor report, paste one announcement and ask what changed.
Summarize this competitor announcement in five bullets. Separate the facts stated in the text from possible implications for our business. Flag anything we cannot tell from this announcement alone. [Paste announcement]
Your part: provide the material, choose the next question, and decide which points matter. Stay here when the conversation itself helps you think.
Official reference: Working with Claude
2. Hand over a complete task
How to try it · Still Claude chat
In the same Claude interface, describe the finished result you want, attach the files, and specify the format. For example: “Compare these supplier quotes and return a spreadsheet with prices, delivery times, and missing information.” Connect any needed apps through Customize → Connectors, then review the result.
Chat and Cowork are merging into one conversation. The rollout starts with Pro and Max; if your account still shows both modes, use Cowork for this kind of delegated task until the new experience arrives.
Official reference: The merged Claude experience and rollout
You define the result. An agent uses available tools, checks what it finds, and takes further steps toward that result. You review the output and resolve questions that need your judgment.
Try it with: a supplier shortlist, a presentation built from your files, or our competitor report. Give it the sources and access it needs before starting.
Create a two-page briefing on these five competitors: [names and official websites]. Cover announcements from the past seven days. Use official sources where possible and link every factual finding with its publication date. Group the briefing by competitor, then add three implications for [our business]. Label those implications as your analysis. If you find no update, say so. Save a draft for my review and list any gaps; do not send it to anyone.
Check the result: open the source links, verify the dates, and confirm all five companies are covered. A polished report can still contain mistakes.
Official reference: Delegate work in Claude
3. Coordinate a team of agents
How to try it · Claude Code
Start a Claude Code session with your task files available. Use the team prompt below to request separate researchers for customers, competitors, and distribution, then ask the lead to combine their findings. You can also ask Claude Code to create reusable subagents for jobs you repeat.
Video coming: I’ll walk you through setting up your first team of agents, step by step.
Official reference: Delegate work to subagents in Claude Code
A lead agent divides a job into separate pieces and combines the results. This helps when the pieces can be researched independently. The lead still needs to remove duplicates, resolve disagreements, and check the final report.
For our report: give each competitor to a separate researcher. Another useful split is market size, customer needs, suppliers, and competitors for a new product launch.
Research [market and geography] for [business decision]. If your environment supports subagents, delegate independent research on customers, competitors, and distribution to three researchers. Give each the same date range: [range]. Require source links, dates, findings, and uncertainties. Then combine their work into one report, check conflicting claims against the sources, and remove duplicates. If subagents are unavailable, tell me and propose a single-agent approach.
Asking a regular chat to “act as five experts” does not by itself create five separate agents. In Claude Code, check the task transcript for actual subagent activity.
Your part: choose the research question and review the synthesis. Compare the result with a single-agent run before making teams your default; coordination adds work and usage.
Official reference: Create and use subagents
4. Put recurring work on a schedule
How to try it · Claude Code + Routines
Open Claude Code Routines → New routine. In the desktop Code tab, look for Routines → New routine → Cloud.
- Name the job and paste the complete, tested instructions, including the report format.
- Select the repositories, environment, and connectors it needs. Choose a Schedule trigger and check the time zone.
- Create the routine, use Run now, and review its output before relying on the schedule.
Cloud routines run with your laptop closed. Local scheduled tasks run on your computer. Routines are in research preview on paid plans; organization owners can disable them. Included connectors can act without asking during a run, so keep access limited to the job.
Video coming: I’ll show you how to automate your workflow with a routine.
Official reference: Create and run a Claude Code routine
Once a task works well, you can stop starting it manually every time. A schedule supplies the trigger; the task still needs fresh inputs, a clear output, and someone to review it.
Try it with: a Monday competitor briefing, a Friday project update, or a quarterly draft report. A scheduled task can use one agent; you do not need a team first.
Set up a competitor briefing every Monday at 9:00 AM Europe/Paris, covering the previous seven days. Track [competitor names and official websites]. Use the report format we tested. Include source links and publication dates, and distinguish new findings from repeated news. If a source is unavailable, mark the gap. Deliver the draft here for my review. Show me the schedule and instructions before activating it.
- Run it once.
Check the report against your sources and adjust the instructions.
- Confirm the schedule.
Verify the time zone, inputs, and delivery destination in the scheduling interface.
- Check the first automatic run.
Confirm it actually ran and used fresh information. Review run history when an expected report is missing.
Claude’s scheduling options depend on your account and setup. Remote tasks need accessible files and connected tools; tasks that rely on local files or apps have local execution requirements. See the current documentation before relying on an unattended run.
Official reference: Create and manage scheduled tasks
5. Work with agents as teammates
How to try it · Slack or Microsoft Teams
Slack: use Claude Tag, available in beta on Claude Team and Enterprise. A Primary Owner or Owner sets up its identity, tools, and allowed channels. Once ready, tag @Claude in a channel with a specific task, then follow the work in the thread.
Microsoft Teams: one documented route is a Copilot Studio agent. Create and test it, publish it, connect the Teams channel, and share it with the intended users. Your organization must allow the app; work with its admin for availability and access.
These are two different product paths. For level 5, define the agent’s ongoing responsibility, the events it should react to, and when it hands work back to a person. A chatbot placed in a channel still needs that workflow.
Official reference: Set up and work with Claude Tag in Slack
Official reference: Connect a Copilot Studio agent to Teams
Here you give an agent an ongoing responsibility in a team workflow. It responds to relevant events and chooses actions within the scope you set. This usually needs integrations, rules, and a team owner; adding a bot to a channel is only one part of the setup.
Illustrative example: your competitive-intelligence agent checks new questions in a designated team channel, finds relevant evidence, drafts a response, and flags unanswered questions for the owner. It maintains a shared record so the same question does not trigger duplicate work.
Another example: a client-support agent classifies incoming requests, drafts replies from approved information, and routes unusual cases to a person.
Help me design an agent responsible for making sure every client request in [channel] is acknowledged within one business day. Define the events it should respond to, the information it needs, actions it may take, and cases it must hand to a person. Start with draft replies for review. Specify how it records completed work, avoids duplicate responses, and alerts the owner when it cannot proceed. Return a proposed workflow; do not activate it yet.
Your part: define success, review exceptions, and adjust the workflow when real cases expose gaps. The example above is a design brief, not a feature that a prompt automatically enables.
Official reference: Agent patterns and customer-support examples
Choose your next step
Pick one task you would like to take off your plate. Write down its inputs, the result you want, how often it happens, and what you need to review.
| What feels repetitive | What to try |
|---|---|
| You keep prompting every small step | Try level 2: describe the complete deliverable. |
| One large research task has independent parts | Try level 3: split the research and define a shared output format. |
| You repeat a reliable task every week | Try level 4: test a schedule and check its first run. |
| Your team needs ongoing responses to events | Explore level 5: define the remit and integrations with the team owner. |
Next in the video series
I’ll show you how to move to levels 3 and 4: building a team of agents and automating recurring work. Start by choosing the task you want to use.
Official reference: Choose the simplest approach that works
Sources to keep handy
Explore Boris’s framework and the documentation behind the examples. Product availability and menus can change.
Boris Cherny’s five levels of AI adoptionOriginal framework · Explore Boris’s post on X.Claude Code subagentsOfficial documentation · Separate agents, tool access, and delegation.Scheduled tasksOfficial documentation · Setup, run history, and scheduling requirements.Building effective agentsOfficial documentation · When to use agents and how common patterns differ.The merged Claude experienceOfficial documentation · Questions and delegated tasks in one conversation.Claude Code RoutinesOfficial documentation · Cloud routines, setup, and run history.Claude Tag in SlackOfficial documentation · Team setup and channel-based work.Agents in Microsoft TeamsOfficial documentation · Connect and share a Copilot Studio agent.