What AI agents can do today, and where they stop

AI agents can already do a useful amount of office work: read a brief, inspect documents, collect information, compare records, draft material, prepare updates and put the result into a usable deliverable. They are especially good at work where the inputs are visible, the rules can be written down and a person can check the result before anything consequential happens.

They are much less reliable when the work depends on judgement that has not been written into the brief, on the history of a relationship, or on an action that cannot easily be taken back. The useful boundary is practical: which parts can the agent carry out, where should it stop, and who signs off before the work leaves the workspace?

Reading and sorting are already strong delegation candidates

Give an agent a folder of material and a specific question, and much of the first-pass work can be delegated. It can read meeting notes, pull dates out of a contract pack, group support requests by topic, find repeated complaints, compare two versions of a document or turn a long thread into a short briefing.

This works because the source material is there to inspect and the task has a visible finish line. "Find every renewal date and put it in a table" is a cleaner assignment than "tell me what I should worry about." The first asks for extraction and organisation. The second asks the agent to decide what counts as important, which depends on context that may not exist in the files.

Careful delegation helps here. A brief should say what sources the agent may use, what fields matter, how uncertainty should be marked and what the final deliverable should look like. If a missing figure must be left blank rather than inferred, say so. That small instruction can prevent a tidy report from becoming a confident guess.

Drafting works best when the agent is not the final editor

Drafting is another good fit. An agent can prepare a reply to a supplier, turn notes into a project update, write a first version of an outreach message, reshape a long memo for another audience or prepare a meeting brief from material already on hand.

The weak point appears when a draft quietly becomes a decision. Consider an agent preparing an email about a delayed project. Writing a clear account of the delay is ordinary drafting. Deciding whether to blame a contractor, promise a new delivery date or offer compensation is different. Those choices can change a relationship or create an obligation.

A sensible setup lets the agent get the message almost ready, then stops before it is sent. The person reviewing it should see the actual proposed action, not a vague request such as "approve email." In Delagents, work is set up around that split: a person writes the brief, the agent carries out the work, and actions with an effect outside the workspace wait for sign-off. The product overview shows how those delegations, checkpoints and deliverables fit together.

Comparison and checking need a rule for what counts as a mismatch

Agents are useful when two sets of records need to be compared. They can match invoice lines against a ledger export, check whether names in one list appear in another, find changed clauses between document versions or flag a support ticket whose requested refund does not match the order value.

The trap is hidden policy. Suppose two invoices differ by a few cents because of tax rounding. Should that be ignored, corrected or sent for review? A person may know the company's usual treatment without ever having written it down. The agent does not share that unwritten history.

Good briefs turn those habits into rules. State the tolerances. Define the fields that must match exactly. Say which exceptions can be filed automatically and which should be brought back to a person. The agent can then do a large amount of checking without pretending that every discrepancy has the same meaning.

Filing and record updates are useful until the change becomes destructive

Agents can also handle routine filing: rename documents, assign categories, prepare records for an internal system, add notes, create a new entry or assemble a packet for someone else to review. These tasks are often dull for a person and easy to describe.

The boundary becomes sharper once the work can damage existing records. Creating a new draft record is different from overwriting one someone else created. Adding a proposed category is different from deleting the old one. A competent agent may understand the instruction perfectly and still be the wrong actor to make an irreversible change without a person looking at it first.

Relationships and judgement are where the handoff gets awkward

Some work looks simple on paper because the difficult part lives in someone's head. A founder may know that one supplier can take a blunt message and another needs a phone call. An operations lead may know that a late invoice is harmless because of a standing arrangement. A recruiter may know that an unusual gap on a CV was already explained in an earlier conversation.

An agent can use information it has been given. It cannot safely invent the unwritten part of the relationship. That is why first contact, terms, permissions and public commitments deserve a hard stop rather than a soft suggestion.

The same applies to actions where the cost of a bad call is hard to reverse. Paying, issuing refunds, widening access, accepting terms, changing roles, publishing publicly and acting in another workspace should stay with a person. So should deleting or overwriting records the agent did not create. These limits are about authority: the agent may be capable of preparing the action while a person remains responsible for deciding whether it happens.

Choose the next delegation by looking for a reversible middle

A useful first delegation has a clear input, a visible output and a point where the agent can stop before the outside world changes. Take a weekly supplier review. The agent can read the invoices, compare them with purchase records, flag mismatches, draft questions and prepare a short file for review. A person then decides which question gets sent and whether any payment should move.

Before handing over a task, write down four things: what the agent may read, what it is expected to produce, which judgement calls must come back to you and which actions require explicit approval. Put a budget around the delegation too. If the agent needs tools or integrations, decide their scope before the work starts.

Delagents is built around that division of labour. Each delegation has a hard budget cap, and the activity log records tool calls, approvals and policy changes. Actions that affect the outside world wait in an approvals inbox with the exact proposed action, the cost so far and a deadline. Start with one task you currently do by reading several inputs and producing a draft, comparison or filing decision. Write the brief so the agent can complete that middle portion on its own, then keep the final consequential step for yourself.