AI Agents for Business: What They Do Well, and Where They Fail

In short: an AI agent is software that uses a language model to work through a task in steps: read a request, look things up in your systems, draft a reply or take an action, and hand over to a person when it is unsure. Today agents work best on narrow, repetitive tasks with clear rules and a person signing off.

Every week someone asks me whether they should "get an AI agent". The honest answer is almost always: yes, for one job, done carefully. Not for everything at once.

What is an AI agent, in plain words?

A chatbot answers. An agent does. A chatbot replies to the question in front of it. An agent can open your order system, find the right policy, fill in a form, write the reply, and then stop and ask a person before anything that matters goes out.

Think of it as a fast, tireless junior colleague on their first week: good at the routine, needs clear instructions, and should not be left alone with money, angry customers or anything legal.

Where AI agents work well today

• First response on a support queue. Reading the message, working out what it is about, pulling the order or policy, and drafting a reply for a person to approve.
• Sorting incoming email and WhatsApp. Tagging, routing to the right person, and flagging the urgent ones.
• Answering from your own documents. Policies, manuals, contracts and reports, with the source page shown every time so people can check.
• Preparing reports. First drafts of weekly updates, donor reports and minutes from your notes and spreadsheets.
• Follow-ups. Reminders, payment nudges and appointment messages, written in your tone.
• Internal helpdesk. "How do I claim travel?" "Which form for leave?" Answered in seconds, in English or Hindi.

Where they still fail

• Judgement. Exceptions, grey areas and decisions with consequences.
• Upset people. An angry customer needs a person, quickly.
• Facts that are not in your documents. Agents can sound confident while being wrong. Make them show their source, and say "I don't know" when there isn't one.
• Money and commitments. Refunds, payments and promises should always wait for a person's approval.
• Long jobs with no checkpoints. The longer the chain of steps, the more small errors add up. Break work into short steps with checks.

What does an AI agent cost to run?

Building it is only part of the cost. Ask for a monthly estimate based on your volumes, covering:

• the AI model's usage, usually charged per task or per amount of text;
• hosting and the tools it connects to;
• your team's time to check its work, especially at the start;
• someone to adjust it when your policies, prices or products change.

Two habits keep costs sane. Reuse answers to questions people ask again and again instead of paying the AI every time, and set a daily spending cap with a polite message when it is reached. I run the AI tools on this website exactly that way.

How to start safely: five steps

1. Pick one queue or one task. Not "customer service". Something like "replies to delivery-date questions".
2. Collect 20 to 50 real past examples, with the reply your best person gave.
3. Write down what "done well" means: correct facts, right tone, the cases it must hand over.
4. Run a pilot with a person approving every output. Measure how many drafts go out unchanged.
5. Widen only when the numbers say so. Then add the next task.

AI agents for NGOs and small organisations

The best first tasks are usually the paperwork that steals time from the actual work: donor and grant report drafts from field notes, weekly summaries of long WhatsApp groups, and a searchable archive of years of studies where every answer points to its source. Budgets are tight, so start with a pilot on tools you may already pay for.

Want to see AI on real tasks first?

I built a free guide that shows ten everyday tasks for nine professions, with a one-click AI example for each, in English, Hinglish or Hindi. It is a quick way to judge the quality before you spend anything: try it on your profession. When you are ready to talk about your own office, start here: how I work as an AI consultant.

Questions people ask

Is an AI agent the same as a chatbot?

No. A chatbot replies to a question. An agent works through a task in steps: it can look things up in your systems, fill in forms and draft replies, then hand over to a person when it is unsure.

Will an AI agent replace our staff?

It should not, and I would be wary of anyone who promises that. Agents take the routine first steps. Judgement, difficult conversations and anything involving money or commitments stay with your team.

Can an AI agent work on WhatsApp?

Yes. Many agents connect to WhatsApp Business. Start with sorting and drafting replies that a person approves, then widen once it has proved itself on real messages.

Can it reply in Hindi or Hinglish?

Often, yes. Test it early with your real messages, because quality varies by task and by model. Keep a person in the loop for anything sensitive.

Do we need our own servers?

Usually not. Most pilots run on hosted AI services. What matters more is deciding what data reaches the AI, where it is processed, and whether the vendor keeps or trains on it.