What tasks should you delegate to an AI agent first?

Agents & delegation/6 min read

Delegate tasks that pass three filters: they happen often, they have a definable good result, and a mediocre first attempt costs you nothing. Research, first drafts, summaries, and monitoring pass all three. Decisions, relationships, and anything you cannot verify do not. Here is the full starting list and how to hand over task number one.

What makes a task right for an AI agent?

Run every candidate task through three filters.

Frequency. A task you do once is faster done than delegated, because writing a good brief takes effort. A task you do every week pays that effort back forever. Weekly research, recurring reports, and standing summaries are the textbook cases.

Definability. If you cannot describe what a good result looks like, the agent cannot produce one, and neither could a new hire. "Summarise these five reports into one page with the three decisions I need to make" is definable. "Keep an eye on things" is not.

Cost of a bad first draft. Early outputs will be imperfect while your briefing improves. Pick tasks where imperfect costs you a shrug, not a client. A rough research pack wastes ten minutes of review. A rough email to your biggest customer wastes a relationship.

Notice what is not on the list: difficulty. Agents handle hard tasks fine. It is vague, rare, and high-stakes tasks that fail, and they fail with humans too.

Which tasks should you start with?

The reliable first wave, in rough order of payoff:

Research packs. "Collect what was published on X this week, summarise, flag the three items worth my attention." This is the single highest-value starter because it is frequent, definable, and the raw material was going to eat your evening anyway. It is also where my own agent started: overnight research packs waiting with my morning coffee.

First drafts. Emails, posts, documents, briefs. Not final versions, first drafts. The agent moves you from blank page to editing, which is where your judgement actually earns its keep.

Summaries and digests. Long reports, meeting transcripts, overflowing inboxes, forty open tabs. Compression is the task agents do best out of the box, with the least briefing skill required from you.

Monitoring. Prices, competitors, mentions, regulatory updates, anything you check manually out of low-grade anxiety. Give the agent the checklist and the threshold for waking you up.

Repurposing. One piece of content into five formats. The thinking is already done, and it is yours; the agent handles the reshaping.

Data tidying. Messy lists into clean tables, inconsistent formats into one format, scattered notes into a structured document. Boring, frequent, perfectly definable.

This is not a niche skill to acquire someday. Gartner predicts that by 2029, at least half of knowledge workers will need to work with, govern, or create AI agents on demand. The people starting with research packs today are building that skill on easy mode.

What should you not delegate first?

Four categories stay with you, at least until your briefing and review skills are proven.

Judgement calls. The agent prepares the decision. It does not make it. Anything involving trade-offs you would struggle to explain to a colleague is not ready for delegation, because you cannot brief what you cannot articulate.

Relationships. Messages where the reader knows your voice, or where trust is the actual product. An agent can draft; the send button is yours.

Anything you cannot verify. If you lack the knowledge to check the output, you are not delegating, you are gambling. Delegate inside your competence first, expand later.

Sensitive data and sign-offs. Legal, financial, and reputational calls need a human signature, and personal data needs boundaries set before the first task, not after the first incident. This is not fear of the technology. It is the same access control you would apply to any new contractor in week one.

The data on skipping these rules is brutal. Forrester and Anaconda research from 2026 shows 88 percent of AI agent pilots never reach production, and among deployments with negative returns, 41 percent of failures trace back to unclear success criteria and 33 percent to insufficient tool or data access. Read that list again: not one of the top causes is model quality. Agents fail on scoping and ownership, which means they fail on management, and that is fixable at your desk tonight.

How do you hand over the first task?

Write a brief with five lines. This is the whole method.

Goal: what you want and why, one sentence each.

Format: what the output should look like, ideally with an example.

Boundaries: what the agent must not do, touch, or assume.

Quality bar: what makes the result good, and what to check before returning it.

Escalation: what to do when unsure, which is ask, not improvise.

If that list feels like management rather than technology, that is the point. The five lines are the same ones a good operations manager gives a new team member, and your agent will show you very quickly which kind of manager you are. Literally: An AI agent will show you what kind of manager you really are.

How do you know it is working?

Two signals, one per month.

First, review time falls. Early on you will fix a lot. If your briefs are improving, the fixing shrinks week over week. If it is not shrinking, the problem is almost always the brief, not the agent.

Second, the task disappears from your head. The real payoff of delegation is not the hours, it is the mental slot. When you stop thinking about the weekly research between Tuesdays, the delegation worked.

When both signals are green, add the second task. One at a time. The people running five-agent stacks all started with one research pack, one brief, and one review loop, and the whole first step fits in a single evening: Build Your First Agent in an Evening.

FAQ

What is the best first task to give an AI agent? A recurring research pack: collect, summarise, and flag what deserves your attention on a topic you already track weekly. It is frequent, easy to define, and a weak first attempt costs you nothing but a short review.

What tasks should you never give an AI agent? Final decisions, messages where trust is at stake, tasks you lack the expertise to verify, and anything touching sensitive data without preset boundaries. The agent prepares; a human signs.

How many tasks should I delegate at the start? One. Run it until your review time drops and the task stops occupying your head, then add the next. Stacking five delegations on day one multiplies briefing errors, not output.

Do I need technical skills to delegate to an AI agent? No. The skill is writing a five-line brief: goal, format, boundaries, quality bar, escalation. It is a management skill, and it transfers directly back to delegating to people.

Author box

Natalya Rojkova is a finance leader who runs capital controls for multi-billion dollar data centre builds by day and builds with AI agents every evening. She writes about delegation, agents, and the new layer of management.

CTA block (Agents & delegation pillar)

One task. Five lines. One evening. Build Your First Agent in an Evening is a free guide that walks you through setting up your first working agent, brief included. [Kit form: evening-guide]

Sources referenced in text (for the page footer, optional)

Gartner press release, August 2025: at least 50% of knowledge workers to work with, govern, or create AI agents by 2029.
Forrester / Anaconda research, 2026: 88% of agent pilots never reach production; failure root causes 41% unclear success criteria, 33% insufficient tool or data access.