How to delegate work to AI agents: tasks, autonomy levels and review loops

The short answer: hand over work that happens often, follows rules you can write down and is cheap to check. Give each task a brief with a clear definition of done, decide up front whether the agent may only draft, may act after your approval, or may act and report afterwards, and review its work at set points rather than all day. The model behind the agent matters far less than the system around it — who approves what, where the agent reports, and what it does when it gets stuck.

This guide is written for Australian owner-operators and small teams who already know AI can write a decent email and now want it to take real work off their plate: receipting, follow-ups, research, bookings, reports. It is the hub of our delegation series, so each section points to a deeper article when you are ready for the detail.

Asking an AI versus delegating to an AI agent

Asking a chatbot is still you doing the work, with help. You paste in context, read the answer, copy it somewhere else and remember to follow up. Delegating is different: the agent owns a task through to a result, works inside your actual email, files and CRM, and comes back with evidence of what it did.

Compare two instructions. “Draft a reply to this tenant” is a request. “Every weekday morning, match yesterday’s rent deposits to tenant ledgers, receipt the exact matches and send me a list of anything that doesn’t match” is delegation. The second one has an owner, an input, a rule, an output and an exception path. If you cannot write those five things down for a task, you are not ready to delegate it — to an agent or to a new staff member.

Choosing which tasks to hand over first

Start with work that is frequent, rule-based and easy to verify. Do not start with the task you hate most. The most hated task is usually hated because it is full of judgement — a difficult client, a pricing exception, a dispute — and that is exactly where an agent with a fortnight of context will disappoint you.

Score candidate tasks on four questions: how often does it happen, are the rules clear, can you check the result quickly, and what does a mistake cost? The table below shows how that plays out for a typical small business. The verdicts are our view, not a law of nature; your own risk appetite should move them.

Work most businesses should keep manual

• Conversations where you need to read the room: complaints, apologies, bad news, anything emotional.
• Pricing exceptions, discounts and negotiations, where the rule is “it depends”.
• Advice that is regulated or professional in nature. An agent can assemble the facts; the advice is yours.
• Moving money out. Agents can prepare a payment run or a disbursement list; a person should release it.
• Anything where you cannot describe what the right answer looks like. If you can’t, neither can the agent.

One honest caveat: if your business is mostly hands-on craft and you have less than an hour of admin a week, you probably don’t need an agent platform at all. A good template library and a calendar will do. Delegation pays off when there is a steady stream of repeatable work that currently lands on you in the evenings.

Three autonomy levels, set per task

Every delegated task should carry one explicit autonomy level. Leaving it implicit is how people end up surprised that an email went out, or annoyed that the agent asked permission to rename a file.

Set levels per task, not per agent. The same agent can be at level 3 for filing receipts and level 1 for replying to complaints. Promote one task at a time, and only on evidence: for example, a run of approvals where you changed nothing. Pick your own threshold and write it down. Demote just as deliberately — if an error reaches a customer, that task drops a level until you understand why.

Some actions should never leave level 2 in a small business: messages to customers, payments, publishing and deleting. We explain why, and how to make those approvals take seconds rather than minutes, in approval gates for AI agents.

Writing a brief the agent can actually execute

Most failed delegation is a bad brief, not a bad agent. “Sort out the overdue invoices” leaves every important decision unmade: which invoices, how overdue, what tone, what happens if someone has already paid, and when to stop. A good brief has six parts.

1. Outcome — the result in one sentence, from the customer’s or business’s point of view.
2. Inputs — where the agent should look: which inbox, folder, CRM view or spreadsheet.
3. Constraints — tone, limits, things never to do, who must not be contacted.
4. Definition of done — what has to be true for the task to count as finished.
5. Stop conditions — situations where the agent must pause and ask instead of guessing.
6. Report format — what you want back, how long, and where.

That brief takes three minutes to write and saves a week of back-and-forth. We break down each part, with bad-versus-good examples, in how to write instructions for AI agents.

Your first week: earning trust deliberately

Trust should be earned in small, visible steps. In the first days, connect the agent to the systems it needs with the narrowest access that works, give it one research task and one drafting task, and read everything it produces. By the end of the week you should know where its judgement is good, where your rules were vaguer than you thought, and which one task is ready to move to approval mode.

The most common first-week mistake is the opposite: connecting every system on day one, asking for a big project, and then losing confidence when the agent makes a reasonable guess you didn’t want. Our day-by-day plan for the first week with an AI assistant lists what to hand over first, what to hold back and what to look for in its output.

Review loops that don’t turn into babysitting

The point of delegating is to stop watching the work. That only happens if review has a shape. We use three loops, each with a different cadence and purpose:

• Approval requests, in real time. Short, specific, answerable with a tap: here is the draft, here is why, approve or edit. If an approval needs you to open three other systems, the request is badly designed.
• A daily exception report. One message a day listing only what needs you: things the agent couldn’t finish, things that broke a rule, anything waiting on your decision. Routine success goes in a one-line count, not a list.
• A weekly review, about half an hour. Look at edits you made to drafts, exceptions that repeated, and rules you changed in your head but never wrote down. Update the briefs.

The daily report is the loop most people get wrong — either it is a wall of “all done” that nobody reads, or it doesn’t exist. We show what to include and leave out, with a template, in designing a daily exception report.

Escalation: what the agent does when it gets stuck

A well-delegated agent stops and asks rather than guesses. That sounds obvious, but it has to be designed in. Write down your escalation triggers once and include them in every relevant brief:

• The instruction is ambiguous, or two rules conflict.
• Data the task depends on is missing or doesn’t reconcile.
• An amount is above a threshold you set.
• The customer sounds upset, or uses words like complaint, refund, lawyer, tribunal or ombudsman.
• The agent hits an access error or a system is down.
• The task would touch a person or record you have flagged as sensitive.

Then decide where escalations go. For a sole operator it is simply the chat thread with the agent. For a team, name an owner per area and a fallback when that person is away. What should never happen is a silent best guess on money, legal matters or a customer relationship. An escalation that waits until tomorrow morning is a minor delay; a confident wrong answer sent at 9pm is a problem.

Splitting bigger requests across agents and people

Single tasks are easy to delegate. Requests like “get the month-end owner statements out” or “prepare for Thursday’s tender meeting” are really several tasks with dependencies, some for agents and some for people. That is the orchestrator’s job: break the request into pieces, give each piece a clear owner, run independent pieces in parallel, and bring the results back into one thread.

In Cortex that orchestrator is Fulcrum Sessions. It plans the request, splits it into agent tasks and human hand-offs, runs specialists side by side where it makes sense and reports back in the same conversation. How that works — and where parallel work helps or hurts — is covered in multi-agent orchestration explained.

Measuring whether delegation is working

Resist the urge to quote a big “hours saved” number on day ten. Measure things you can actually observe, and count your own review time as a cost. The method in measuring hours saved by AI agents uses before-and-after sampling; the signals below tell you sooner whether you are on track.

Delegation mistakes we see most often

• Vague outcomes. “Help with marketing” is not a task. “Draft three replies to this week’s Google reviews for my approval” is.
• Broad access on day one. Give the agent what the task needs and no more. Our security series covers least-privilege access for AI agents.
• No stop conditions. Without them the agent will fill gaps with reasonable guesses, which is precisely what you didn’t want.
• Rubber-stamp approvals. If you approve without reading, you have built level 3 with extra steps and none of the safety.
• Rules that live in your head. When you correct the agent, ask it to update the written rule, then check it did.
• Delegating judgement disguised as admin. If every instance is “a bit different”, it isn’t ready.

Where Cortex fits, and where it doesn’t

Fulcrum Cortex is built around the approach in this guide: specialist agents work in your email, files and CRM, you direct them from the chat app you already use, they draft by default, and nothing outward-facing goes out without your OK. When you want a real person in the loop as well — for example to cover approvals while you are on leave — there is a human-in-the-loop add-on, priced on the pricing page.

It is not the right tool if you want agents that act on customers with no approvals at all, if your data cannot leave a specific on-premises system, or if your admin load is genuinely small. If you are somewhere in between, start with one task and the three autonomy levels, and see how far it goes. You can see how Sessions plans and reports work on the Sessions page.