The Mistake Everyone Makes at the Start
When people first get serious about AI, they go one of two ways.
The first group automates everything they can touch. They build workflows for newsletter drafts, social captions, email responses, meeting summaries, weekly reports. Six months later, half the automations are broken, the outputs are mediocre, and they've spent more time maintaining the system than the system has saved them.
The second group stays cautious. They use AI to help with individual prompts but never build anything systematic. They save time here and there but never cross the threshold where AI feels like a real force multiplier.
Both groups are missing the same thing: a clear framework for deciding what kind of AI involvement each task actually needs.
That's what the delegation stack is.
The Three Levels
Not all tasks are created equal. The question isn't "can AI do this?" — it can do almost anything, at some level of quality. The question is "what's the right role for AI in this specific task?"
There are three levels:
Level 1: Fully automate. AI runs the task start to finish. You review the output occasionally, but it doesn't require your active involvement each time.
Level 2: AI assists, you decide. AI does the heavy lifting — research, drafting, structuring — and you apply judgment, edit, and approve. You're still in the loop, but the time is dramatically compressed.
Level 3: Keep it human. AI might inform the work, but the execution stays with you. The judgment, the relationship, the creative direction — these don't get delegated.
Most people intuitively understand level 1 and level 3. The mistake is in the middle. Level 2 is where most of your high-value time is, and most people either underuse AI there (staying manual when they don't have to) or overuse it (removing their judgment when it actually matters).
Level 1: What You Should Fully Automate
Fully automatable tasks share a few characteristics:
- They follow a consistent template or structure
- The inputs are predictable and sourced from the same places each time
- The output doesn't require your personal voice, relationship context, or fine judgment
- A mistake would be inconvenient but not damaging
Examples:
Weekly summary reports. If you're pulling data from the same sources, running the same calculations, and producing the same format every week — this is automation territory. Set up an n8n or Zapier workflow that pulls your data, feeds it to Claude with a fixed prompt, and delivers the report to wherever it needs to go. You review it once a week for anything unusual.
Social post formatting. Not ideation — formatting. Taking a blog post or email and producing the LinkedIn version, the Twitter thread structure, the Instagram caption. Same content, different format. This is a mechanical transformation. Automate it.
Meeting prep summaries. Before a client call, you want a brief on who they are, what you last talked about, and any open items. If that information lives in a CRM or note system, an AI can pull and format it automatically. You read it on the way into the meeting.
Data categorisation. Tagging, sorting, labelling — anything where you're applying the same classification logic repeatedly. AI is very good at this and it's almost always fully automatable.
The common thread: structure. If the task has clear inputs, a predictable process, and consistent outputs, automate it.
Level 2: Where AI Assists but You Decide
This is the largest and most valuable category. Level 2 is where knowledge workers actually spend most of their time, and where AI provides the biggest leverage without the risk of going on autopilot when you shouldn't.
The pattern is: AI does the laborious parts, you apply the judgment.
Client proposals and emails. You know the context — the client's situation, what matters to them, the relationship history. AI can take those inputs and produce a well-structured draft. You edit, adjust the tone, add the thing only you would know to include, and send. What used to take 45 minutes now takes 15.
Research and analysis. Pulling information together, summarising documents, identifying patterns across data. AI is fast at this. But the interpretation — what this means for your decision, what the risks are, what action to take — that stays human. Use AI to build the brief. You make the call.
Content ideation and drafting. AI can produce a solid first draft from your notes and a clear prompt. But your voice, your point of view, your specific examples — those come from you in the editing pass. The draft collapses the blank page problem. You provide the substance.
SOPs and documentation. Describing a process you know well to an AI, having it structure it into a clear document, then cleaning it up yourself. Faster than writing it from scratch. Better format than you'd naturally produce. Still requires your knowledge of how the process actually works.
Level 2 tasks are where people often make the mistake of going too far. They let AI handle too much of the thinking, and the output suffers — it lacks the specific context, judgment, or voice that made it valuable in the first place.
The fix is being clear about your role in the workflow. You're not reviewing AI output for typos. You're editing as the expert who knows what's actually right.
Level 3: What Should Stay Human
Some things shouldn't be delegated to AI, even partially.
Not because AI can't produce something. It can produce a passable version of almost anything. But because the value of the thing comes specifically from you doing it.
Client and stakeholder relationships. The people you work with can tell the difference between a thoughtful, human response and a polished AI-generated one. Not always in the words. In the feeling. In whether they sense you actually thought about them. Automate the admin around relationships. Don't automate the relationships.
Strategic decisions. AI is excellent at gathering information and presenting options. It's poor at making calls under genuine uncertainty, with real stakes, where the right answer depends on things no model has access to — your risk tolerance, your values, your read on people, your gut. Use AI to prepare for strategic decisions. Make them yourself.
Anything where being wrong is expensive. Legal, financial, medical, compliance. AI can assist with research and drafting in these areas. But the judgment and accountability stay with a human professional. The risk of an AI-assisted error in these domains is too high to treat outputs as final.
Creative work with a point of view. If your value is your specific take on the world — your writing voice, your design aesthetic, your creative direction — automating that erodes the thing people are actually buying. Use AI to handle the mechanical work around your creative output. Don't use it to replace the output itself.
How to Audit Your Work Against This Framework
Take 20 minutes and list the 10 tasks you spend the most time on in a typical week. For each one, ask:
- Does this follow a repeatable structure? (Yes → consider level 1)
- Does this require my specific judgment, context, or voice? (Yes → level 2 or 3)
- What's the cost of AI getting this wrong? (High → level 2 or 3; low → level 1)
- Is the value here in the process or the outcome? (Process/relationship → level 3)
Most people find that around 20–30% of their work is genuinely level 1 (fully automatable), 50–60% is level 2 (AI-assisted), and 15–25% is level 3 (stays human).
The goal isn't to push everything to level 1. It's to be accurate about which level each task belongs to — and then build your AI systems accordingly.
The Stack You End Up With
Once you've done the audit, you have a map.
Level 1 tasks become automated workflows — set them up once, let them run, review occasionally. Level 2 tasks get AI-assisted processes — clear prompts, consistent inputs, a defined editing pass. Level 3 tasks get protected time — you stop wasting mental energy on the automatable stuff so you have more to give here.
The result isn't a fully automated business. It's a business where automation handles what it should, AI assists with the middle ground, and you're spending your actual hours on the things that only you can do.
That's the stack worth building.