In 2024 everyone wanted a chatbot. In 2026 everyone wants an agent, and most people asking for one cannot say what the difference is. That is not their fault; the word is used for everything from a scheduled script to a system that runs half a business. This article explains what an AI agent actually is, where it already earns its cost for a small company, and how to try one without risking anything you cannot undo.
What an agent is, in one paragraph
A chatbot answers. An agent acts. It is a language model given a goal, a set of tools it is allowed to use, and a loop: decide on a step, call a tool, look at the result, decide the next step, and stop when the goal is met or it needs a human. The tools are ordinary software, such as searching your order database, drafting an email, creating a ticket or updating a spreadsheet. The model supplies the judgement between steps. Everything useful, and everything risky, comes from which tools you hand it.
An agent is only as safe as the most powerful tool you give it, and only as useful as the data that tool can reach.
Five jobs where agents already pay for themselves
The best early uses share three properties: the work is repetitive, the result is easy to check, and a mistake is cheap to reverse. These five come up again and again in the businesses we work with.
- Support triage. Read incoming messages, look up the order, answer the routine ones and hand the rest to a person with a summary attached.
- Lead qualification. Reply to new enquiries within a minute, ask the three questions sales always asks, and book a call into the calendar.
- Document processing. Pull fields out of invoices, purchase orders or application forms and put them into the system they belong in, flagging anything unclear.
- Internal knowledge. Answer staff questions from your own policies, price lists and manuals, with a link to the source every time.
- Reporting. Gather numbers from three tools every morning and post a short summary, which is the same idea as the digest in internal dashboards people actually use.
Where agents still fail
Agents fail in predictable ways, and knowing them is most of the work of deploying one safely. They are confidently wrong when the information they need is missing. They follow instructions found inside the content they read, which means a cleverly written email can steer an agent that processes email. They drift on long tasks, losing track of the original goal after many steps. And they are poor at anything where the correct answer depends on an unwritten rule that lives in one manager's head.
- Never give an agent the ability to send money, delete data or sign anything without human approval.
- Treat every document and message it reads as untrusted input, not as instructions.
- Keep tasks short. Five well-defined steps beat fifty open-ended ones.
- Log every tool call, so you can see exactly what it did and why.
The Nepali angle: language and channels
Most small businesses in Nepal sell over Viber, WhatsApp, Messenger and the phone, not through a web form. An agent that only works in English on a website misses where the conversations are. Current models handle Nepali and Romanised Nepali far better than they did two years ago, but they still need testing on real customer messages, mixed scripts included. We cover the specifics in AI chatbots that understand Nepali.
What it costs to run
The build is a one-off cost; the running cost is per task. Each step an agent takes calls a model, and a single task can involve several calls. For routine work like triage or data extraction, a small, fast model usually costs a fraction of a rupee per task, and a well-designed agent uses the expensive model only for the steps that need it. The costs that surprise people are elsewhere: an agent stuck in a loop, a prompt that pulls in an entire document on every step, or a task nobody capped. Set a budget per task and an alert per day before launch, not after the first invoice.
How to start without betting the business
Pick one workflow that a person does every day, that takes under ten minutes each time, and where you can check the output in a glance. Run the agent in shadow mode first: it drafts, a person approves, and you measure how often the draft was right. When approval is a formality for most cases, let it act on those cases alone and keep the person on the rest. Expand only once the numbers say so.
- Week one: map the workflow and the tools it needs, read-only where possible.
- Weeks two and three: build and run in draft mode beside the person who does the job.
- Week four: measure accuracy, time saved and cost per task; decide whether to continue.
- After that: allow it to act on the easy cases, keep a human on the rest, and review the logs weekly.
Build, buy, or wait
If an off-the-shelf tool already handles your workflow, such as a helpdesk with built-in AI replies, buy it. Build when the agent needs to reach your own systems, follow your own rules or speak your customers' language, because that is where generic tools fall short. And wait if you cannot yet say which ten-minute job you want removed, because an agent without a defined job becomes an expensive demo. The longer argument for keeping trust while adding AI is in shipping AI features that users trust.
We design and build agents with guardrails, logging and cost limits as part of our AI and machine learning and automation work. If you have a workflow in mind, we can tell you in one conversation whether an agent is the right tool or whether a simpler script would do the job for less.
Abishek Bimali
Founder & Engineer
Abishek founded SiteCraft Innovation and leads its engineering. He writes about building web and mobile products that hold up in production, for teams in Nepal and abroad.
