Founding Batch

January 2027 is priced at founding rates. Fees rise 30% from Batch 2

Apply at Founding Rates
Zrooth
Menu
← All Insights

Insight · Agentic AI

What You Can, and Cannot, Delegate to an Agent

2 September 2026 · 9 min read

You can safely delegate to an AI agent work that is bounded, repeatable and easy to check, where a wrong answer is visible and cheap to reverse. You should not hand over work that is open-ended, that commits money or people, or where a quiet mistake compounds before anyone notices. The guardrails you build around the agent matter more than the agent itself.

This article is for founders, operators, working professionals and SME owners in India, the UAE, Kenya and online who are being told that agents can "run the business". It sets out a practical test for deciding what to hand over, the guardrails that make delegation safe, and the decisions that should stay with a person however good the tools become.

What Is an AI Agent, in Practical Terms?

An AI agent is a system in which a language model does more than answer a question. It decides which steps to take, calls tools (search, a spreadsheet, an email account, a database, a payment system), looks at the result and decides what to do next. Anthropic's engineering team draws a useful line here in Building effective agents (December 2024): in a workflow, the model and tools follow predefined code paths; in an agent, the model directs its own process and tool use.

That difference is the whole question. A workflow does what you designed, in the order you designed it. An agent chooses. Choice is what makes agents useful on messy tasks, and it is also what makes them risky: the more an agent can choose, the more ways it can be wrong without you seeing it. The same article recommends starting with the simplest solution that works and adding autonomy only where it clearly helps, which is sound advice for any business.

The Test: What Happens When It Gets It Wrong?

The useful question is not "can an agent do this?" It is "what happens when it does this wrong, and how would we find out?" Before you delegate to an AI agent, answer four questions about the task.

  1. Is the task bounded? Can you describe the inputs, the allowed actions and the finished output in a few sentences? "Sort incoming invoices into supplier folders and extract the totals" is bounded. "Manage our suppliers" is not.
  2. Is a mistake visible? Will a wrong result show up quickly, in a place someone looks, or will it sit quietly in a record until it causes damage?
  3. Is a mistake reversible? Can you undo the action cheaply? A draft can be deleted. A sent email, a paid invoice or a changed customer record is harder to take back.
  4. Who owns the outcome? Name the person who answers for the result. If nobody can be named, the task is not ready to hand over.

If the answers are yes, yes, yes and a named person, the task is a good candidate. If any answer is no, you either redesign the task (narrow it, add a check, make the action reversible) or keep it with a person.

Tasks You Can Usually Delegate to an AI Agent

The following are good candidates in most small and mid-sized businesses, provided the guardrails in the next section are in place.

Preparing, Not Sending

Drafting replies to routine enquiries, first versions of quotes, meeting summaries and follow-up emails, all left in a queue for a person to approve. The agent removes the blank-page work; a person keeps the final word.

Sorting and Extracting

Reading incoming documents, classifying them, pulling out the fields that matter (dates, amounts, names, reference numbers) and filing them. Errors are easy to spot with a sample check, and nothing leaves the business.

Reconciling and Flagging

Comparing two lists (bank lines against invoices, stock counts against orders) and flagging the mismatches for a person. The agent does the tedious matching; a person decides what each mismatch means.

Researching and Assembling

Gathering public information on a supplier, a market or a regulation and assembling it into a brief with its sources listed, so a person can check the sources rather than trust the summary.

Status and Chasing

Watching a queue and sending reminders against a written rule ("if a signed form has not arrived within three working days, send the standard reminder"). The rule is yours; the agent only applies it.

What Should You Never Delegate to an AI Agent?

Some work should stay with a person, not because the tools are weak, but because the cost of a quiet error is too high or the responsibility cannot be handed over.

  • Moving money. Paying suppliers, issuing refunds or changing bank details. An agent can prepare the payment run; a person approves and releases it.
  • Hiring, firing and disciplinary decisions. These affect people's livelihoods and are subject to employment law in every market. An agent may help organise applications; it should not decide.
  • Legal, tax and regulatory commitments. Signing contracts, filing returns, making statements to a regulator. These need a qualified person who answers for them.
  • Anything involving personal data you have not mapped. India's Digital Personal Data Protection Act, the UAE's federal data protection law and Kenya's Data Protection Act all place duties on whoever handles personal data. If you do not know where the agent sends data, do not let it touch that data.
  • Open-ended goals. "Grow sales" or "fix customer service" give an agent no boundary. It will act, but you cannot tell whether it acted well.
  • Decisions you could not explain afterwards. If a customer, auditor or investor asks why something happened and the only answer is "the agent decided", the task was delegated too far.

The Guardrails That Make Delegation Safe

The OWASP Top 10 for LLM Applications (2025) names the core risk "Excessive Agency": an AI system takes harmful actions because it was given too much functionality, too many permissions or too much autonomy. Its recommended mitigations translate directly into business practice.

Give the Agent Only the Tools It Needs

If the job is reading invoices, the agent needs read access to one inbox and write access to one folder. It does not need your whole email account, your accounting login or a general command line. Narrow tools that do one job are safer than broad ones.

Restrict Permissions

Use read-only access wherever the task allows. Create a separate account for the agent rather than lending it a person's login, so its actions are logged under its own name and can be switched off without locking anyone out.

Put a Person on High-Impact Actions

This is the maker-checker step: the agent makes, a person checks. Sending anything to a customer, deleting records, changing prices and paying money should all wait for a human approval.

Keep a Log You Can Read

Every action the agent takes should be written to a log that a non-technical manager can read: what it saw, what it did, when. Monitoring does not prevent mistakes, but it lets you catch them before they compound.

Set Stopping Conditions

Limit how many steps, retries or messages an agent can take in one run, and what it does when it is stuck: stop and ask, not guess. Anthropic's guidance makes the same point about stopping conditions and checkpoints for human feedback.

For a wider frame, the US National Institute of Standards and Technology publishes the voluntary AI Risk Management Framework (version 1.0, January 2023) and a Generative AI Profile (July 2024). Its four functions, Govern, Map, Measure and Manage, are a useful checklist even for a ten-person business.

How to Delegate to an AI Agent Step by Step

  1. Map the process first. You cannot delegate a task you have not described. Start with where the business leaks time and money and pick one step from the map.
  2. Write the task as a brief. Inputs, allowed actions, forbidden actions, expected output, and what to do when unsure. If the brief runs to pages, the task is too big.
  3. Run it in shadow mode. Let the agent do the work alongside the person who does it today, without acting on its output. Compare results for a week or two.
  4. Measure the errors, not just the speed. Count how often it is wrong, how wrong, and how long each correction takes. Use the method in costing a process before and after AI.
  5. Hand over the low-risk part first. Let it act on the cases where it was reliable in shadow mode and route the rest to a person.
  6. Review on a schedule. Read a sample of the log every week. Widen the agent's scope only when the evidence supports it.

Before you commit to a particular product, test it against the real task, as described in judging an AI tool against a real task. Many tasks turn out to need a simple workflow, not an agent at all, and some can be built without code, as set out in building a working tool without writing code.

An Illustrative Example

The following is an illustrative example, not a real case. A small logistics business in Navi Mumbai receives delivery confirmations by email from a dozen carriers, each in a different format. Today one person reads each one, updates a spreadsheet and emails the customer.

A sensible design lets an agent read the confirmations, extract the consignment number and delivery time, update the spreadsheet and draft the customer email. The agent cannot send. It flags any confirmation it cannot read or that does not match an open consignment. The person now reviews a queue of drafts and a short list of exceptions. If the agent misreads a date, the error sits in a draft, where a person sees it, rather than in a customer's inbox.

The same business would not let the agent chase late payments or adjust carrier rates. Those touch money and relationships, and a quiet mistake there costs far more than the time saved.

Delegation Is Not Abdication

The person who set up the agent still owns the outcome. Regulators are increasingly explicit about this in their own sectors, and it is good practice everywhere: if you rely on a tool, you answer for what it does. The best operators keep the boring, load-bearing decisions close and hand the agent the repetitive middle. That is also how teams stay small without losing control, a theme explored in running every department with a lean team.

Learning to Design Agents With Guardrails

Deciding what to delegate, and building the guardrails around it, is part of the practical work in the AI Entrepreneurship Programme. You can see where it sits in what you learn and in how the twelve weeks work: eight weeks of guided training, then four weeks of remote practical work with an assigned assistant. If you run a business and would rather have agents designed and supervised with you, Zrooth Business Automation starts from the same discipline.

The programme is business consulting and training. It does not promise a job or a business outcome.

Sources

Frequently Asked Questions

What is the difference between an AI agent and an automation workflow?

A workflow follows steps you defined in advance, in a fixed order. An agent decides its own steps and which tools to call. Workflows are easier to test and predict, so use one whenever the task allows.

Can an AI agent send emails to customers on its own?

It can, technically, but for most businesses it should not at first. Let it draft into a queue for a person to approve, and allow direct sending only for narrow, proven cases such as a standard acknowledgement.

Should an AI agent have access to my accounting or banking system?

Read-only access to produce reports or reconciliations can be reasonable. Permission to make payments or change bank details should stay with a person, with an approval step the agent cannot bypass.

What is shadow mode and how long should it run?

Shadow mode means the agent does the work in parallel with a person without its output being used. Run it until you have seen enough real cases, including awkward ones, to judge its error rate honestly; for a daily task that is often a few weeks.

Who is responsible when an AI agent makes a mistake?

The business that deployed it, and in practice the person who set it up and approved its scope. Vendors' terms rarely accept liability for how you use their tools, so plan as if the responsibility is yours.

Do data protection laws affect what I can delegate to an agent?

Yes. India, the UAE and Kenya each have data protection laws that apply to whoever handles personal data. Know where the agent sends data and which providers process it, and take qualified legal advice where you are unsure.

What is excessive agency?

It is the OWASP term for an AI system that can take harmful actions because it has too many tools, too many permissions or too much autonomy. The fix is to narrow all three and add human approval for high-impact actions.

How do I know an agent is still working well after launch?

Read a sample of its log every week, track how many cases it routes to a person, and watch for drift: a rising correction rate usually means the inputs have changed.

Is it cheaper to delegate to an agent than to hire?

Sometimes, but not automatically. Count the tool costs, the set-up time, the checking step and the cost of errors. The net figure is often smaller than the headline, and some tasks are cheaper done by a person.

Pathfinder · Free

Where Do You Fit in the AI Economy?

Type your industry and the work you do. See the live businesses you would learn inside, your twelve weeks, and three venture ideas made for you.

Find My Path

Join the Discussion

Agree, disagree or add your own experience. Every comment is reviewed before it appears; spam and abuse are removed, and disagreement is welcome.

    Add Your View

    0 / 2,000
    Links send a comment to manual review. Please do not share personal details about other people.