What an AI agent can do for a small business, and where it should stop
Agents that answer enquiries and handle routine steps, how we test them before they talk to your customers, and the jobs we keep for people.
By Anubhav Kayal, Asterisk Studio

Photo: Zulfugar Karimov on Unsplash
An AI agent is a program that reads a message, decides what to do, takes a step such as looking something up or booking a slot, and checks the result before replying. A chatbot answers. An agent can also act.
Jobs agents do well
The same twenty questions come in every day: timings, location, price range, delivery, whether something is in stock. An agent on WhatsApp or your website can answer those at eleven at night and pass the serious buyers to you in the morning with their details already collected.
Speed matters more than most owners expect. In a HubSpot survey, 66% of customers said they expect a reply within five minutes. Nobody serving customers at a counter can manage that on WhatsApp all day. An agent can, and Salesforce found that 78% of small and medium businesses in India are already using or trying AI.
For a real estate project, that might mean asking a buyer which flat size they want and their budget, sharing the floor plan, and booking a site visit in the sales team's calendar. For a shop, it might mean reading an order message, checking stock in a sheet, and sending the payment details.
Agentic workflows go a step further and string several of these together. An order comes in, the agent checks stock, raises the invoice, updates the sheet and sends a follow-up two days after delivery, and a person only steps in when something doesn't fit.
Where we keep a person
Anything involving money going out, refunds, complaints or promises the business has to keep goes to a human. The agent says it's passing the conversation on, and it does. A wrong answer about shop timings costs nothing. A wrong answer about a refund costs a customer.
Customers want the same thing. In Salesforce's 2024 survey, 45% of people said they're more likely to use an AI agent when there's a clear way to reach a person.
Testing before it goes live
At the ForgeAI hackathon we built a system of four AI agents that process insurance claims, plus a harness to test them. Every time a trace showed an agent getting something wrong, that mistake became a written rule. The rules were checked on every live request and replayed before each release, so a fixed mistake couldn't quietly come back. It placed third in its track.
We use the same approach for client agents. Before launch we collect real questions from your WhatsApp history, write down the right answers, and run the agent against all of them. When it gets something wrong after launch, that case is added to the test set.
Start with one job
Pick the question you're most tired of answering, or the step that eats an hour of someone's day. We'll build an agent for that one job, run it for a month, and show you what it handled and where it handed over.