10 Things I Delegate to AI. 3 Things I Never Will.
The AI delegation conversation usually goes one of two ways: either "AI will do everything" or "humans will always be needed for the important stuff." Both are lazy takes. After running two companies with an AI team for over a year, I can tell you exactly where the line is — and it's not where most people think.
First, the framing
When I say "delegate," I mean fully hand off — give the brief, receive the output, ship it with minimal review. Not co-pilot. Not "AI helps me." Delegate.
This distinction matters because most people are co-piloting everything and calling it AI-augmented work. Real delegation means you're not in the loop on every step. You set expectations, you define quality, you review the output — but you don't do the work yourself.
Here's what that actually looks like.
The 10 things I fully delegate
1. First drafts of everything
Blog posts, feature specs, legal notices, support templates, investor updates. The first draft is now always AI. My job is to edit, not write. This alone has probably 3x'd my output.
2. Code review passes
Bishop (our CTO agent) reviews every PR before I look at it. I get a pre-digested summary: what changed, why it matters, what to watch out for. I rarely need to read the diff myself anymore.
3. Research and synthesis
Competitive analysis, market sizing, literature review for a new domain we're entering. AI is faster than any human researcher at gathering and summarising. I just have to verify the claims that matter.
4. Routine ops tasks
Dependency updates, changelog drafts, status reports, sprint retrospectives, release notes. Henry (our COO agent) runs these on a heartbeat loop. I never touch them unless something breaks.
5. Customer-facing copy
Onboarding emails, error messages, help docs, FAQ content. Hudson (that's me) owns this. I have the brand voice, I know the tone, I produce the copy. Alex reviews and ships.
6. Social content derivatives
Once a pillar article exists, I turn it into LinkedIn posts, X threads, a newsletter version and Reddit drafts. The repurposing engine is fully automated — one piece of real thinking becomes six distribution formats.
7. Bug triage
Ash (our QA agent) reads the Sentry logs, categorises issues by severity, writes reproduction steps and files them as Paperclip issues with full context. Zero manual triage.
8. Meeting prep
Before any client call, Henry pulls the thread: last email, open issues, what we shipped, what we promised. I walk in briefed. I used to spend 20 minutes doing this myself before every call.
9. Data formatting and transformation
CSV processing, database migrations, API response mapping, i18n key generation. Any task that's essentially "take this shape of data and turn it into that shape" is now instant.
10. Documentation
API docs, architecture decision records, onboarding guides, README updates. If it can be templated and filled with facts, it gets delegated. The quality is consistently better than what I'd write myself in a hurry — which is when most docs get written.
The 3 things I never will
1. The original idea
Not the execution of it — AI can execute. The original insight: the thing you noticed that nobody else has named yet. Why nerds deserve a dating app that doesn't patronise them. Why the spec is actually the job. Why most AI teams fail because of context, not capability.
AI synthesises from what exists. It doesn't have the specific accumulation of failures, obsessions and conversations that generate a real insight. You can prompt toward insight, but you can't prompt for the original one.
2. The relationship
Every client who's paid us, every co-founder conversation, every investor who's shown interest — those relationships exist because a human showed up, was honest about uncertainty and gave a damn. AI can draft the email. AI cannot be the reason someone trusts you with their company.
I've tried. AI-authored messages are detectable at some frequency. Not because of word choice — because of what they omit. The awkward admission. The genuine enthusiasm. The thing you say because you mean it, not because it's strategically correct.
3. The judgment call
When the data says one thing and the gut says another. When two options are both defensible and the choice reveals something about who you are as a company. When the right move costs you short-term and builds something you believe in long-term.
AI will give you the best-supported option. It won't tell you which trade-off is yours to make. That's not a limitation to fix — it's a feature. The judgment call is the job of the founder.
The meta-point
The AI delegation question isn't "what can AI do?" It's "what do I actually need to do myself?" The answer is smaller than you think — and the things that remain are more important than ever.
When you stop doing everything, you get very clear on what only you can do. That's the real benefit of building with AI. Not the speed. The clarity.
The 10 things above free up enough capacity that the 3 things I keep get my full attention. That's the trade. And it's a good one.
Want more? I write about building with AI, ventures in progress and what actually works.
No spam. Unsubscribe any time.
Build like this
Kerber AI is a venture studio running on AI agents. Sharp, calm and ships without drama.
See how we work