Put an AI agent to work in your business

An AI agent is software that takes a request, finds out what it needs to know, does something about it and tells someone what it did. The request can be a WhatsApp message, an email, a scanned invoice or a change in one of your systems. The action can be a reply, a booking, an updated record or a draft waiting for approval.

What makes it an agent rather than a chatbot is that it uses tools. A chatbot can only repeat what it was given. An agent can look up an order in your store, ask the courier for the latest scan, check a customer’s history in the CRM and then decide what to say, all within rules your team has agreed.

That is also why an agent is only as good as the access and the rules around it. Most of our work is not the AI model itself but the plumbing and the judgement around it: which systems it can read, what it is allowed to change, and when it must stop and hand over to a person.

The best first task is one that comes up many times a day, follows clear rules and is easy to check. Order and booking questions, first-line support, invoice and form entry, and follow-ups on enquiries are the usual places to start.

We look for work that people do today by switching between screens and copying information from one to another. If a person can describe the steps and the exceptions, an agent can usually do most of the cases and pass the rest on.

Tasks that rely on taste, negotiation or a long relationship are better left with your team. An agent can still help there by preparing the facts, drafting a first reply or reminding someone that a follow-up is due.

In practice the first agent often looks like this: an online store hands order and delivery questions to an agent on WhatsApp; a clinic lets an agent book, move and remind appointments; a distributor lets an agent read supplier invoices into its accounts; a software company lets an agent answer first-line tickets from its runbooks. Each starts small and grows once the team trusts it.

An agent needs three things: access to the information it should use, the actions it is allowed to take, and rules for when to stop. We work these out with the people who do the job today, because they know the exceptions no document mentions.

Information usually means your product catalogue, prices, policies, past tickets or chats, and the records in your systems. Actions are the few things it may do on its own, such as replying, booking or updating a status. Rules cover the rest: what it must never promise, which topics always go to a person, and how it should hand over.

You also need someone on your side who owns the agent: a person who approves its rules, reads its reports and decides what it should learn next. It does not have to be a technical person; it should be someone who knows the work.

Customers write the way they talk. In India that often means Hindi or Gujarati written in English letters, mixed with English in the same sentence. An agent can understand that and reply in the same style, or switch to the language the customer prefers.

We build agents that work in English, Hindi and Gujarati, and in other languages when a business needs them. Each language is tested on real past messages before launch, because a reply that is correct but sounds wrong still loses the customer’s trust.

The same applies to tone. An agent for a clinic should sound calm and careful; one for a fashion store can be light and quick. We write the tone into its instructions with examples from your own team.

A good agent knows its limits. It answers what it is sure of, drafts what a person should approve, and hands over everything else with the history, the files and a suggested next step.

Most businesses start with a person approving every reply. As the agent proves itself on real cases, you widen what it may do on its own, one kind of case at a time. Refunds, price changes, medical or legal questions and anything unusual can stay with your team for good.

Every run is logged: what came in, what the agent read, what it did and why. That record lets your team check any answer, and it is how we find and fix mistakes.

Agents work through the software you already have: your online store, CRM, accounting or ERP software, booking system, helpdesk, WhatsApp, email or Slack. Modern tools with an API connect quickly; older or in-house software may need a small bridge, which we build.

The agent gets its own account in each system, with only the permissions its task needs, never a person’s login. That way every change it makes is visible, limited and easy to switch off.

Where your own data is involved, such as product documents or past tickets, we set up a searchable store of it so the agent can find the right page and show where an answer came from.

Before we build, we agree what success looks like for the task: how many cases the agent should handle on its own, how fast customers should get an answer, how many hand-overs are acceptable and how mistakes will be counted.

During the pilot the agent runs beside your team on real cases, and we compare its answers with what your team would have done. Launch happens only when those numbers look right to you.

After launch the same measures go into a simple monthly report, so you can see what the agent handled, where it handed over and what changed.

The report also shows what the agent could not do. Those hand-overs are the best guide to what it should learn next: a missing policy, a question nobody wrote down, or a system it still cannot reach.

We choose the model for each task rather than using one for everything. Large models from providers such as OpenAI and Anthropic are strong at understanding messy requests and writing natural replies; smaller or open-source models can be cheaper and faster for narrow jobs such as reading a fixed form.

We use the providers’ business services, where your data is not used to train their models, and we can keep data in the region your business needs. Because the agent’s rules, tools and tests are ours rather than the model’s, we can change the model later without rebuilding the agent.

An agent is not finished at launch. Your products, prices and policies change, customers ask new questions, and the agent needs to keep up.

We run the agent with you: we watch the hand-overs, correct mistakes, update what it knows and keep its tests passing. When the first task is settled, we add the next one, often in the same systems, which makes each new task quicker than the last.