Every business is different. Why are AI agents made to measure?

Because an AI agent is not software out of a box; it is a new employee. And you induct an employee into the business, not the business into the employee.

The same profession, two different businesses

Take two accountancy practices. The first collects invoices by email and enters them into the accounting software straight away. The second gets photos of receipts through a messaging app and keeps its records in a spreadsheet. The same profession. Completely different document flow, exceptions and client habits.

A ready-made "AI for accountants" tool assumes one flow. If yours is different, you have two options. You can rearrange the business around the tool. Or go back to the old methods after a month. Neither is good.

You fit the technology to the people, not the people to the technology.

An agent is an employee. You induct an employee

A new person in the team gets a job description, access rights, a line manager and a probation period. An AI agent needs exactly the same:

• A remit. One specific task, for example: it answers requests for quotes.
• Tools. A mailbox, a price list, a calendar, the sales system. Only the ones it really needs.
• A line manager. A person who approves the important steps: sending a quote, a change in the system, a payment.
• A probation period. We count how many cases the agent handled well and improve it using examples.

Several narrow specialists work better than one agent that does everything. One reads invoices. Another watches enquiries. A third prepares quotes. Each has its own tools and its own limits.

What we tailor with: LangChain and LangGraph

We build agents on open tools from the LangChain family. The LangChain library itself appeared in October 2022, a month before ChatGPT was launched. Today it has more than 700 ready-made connections to models, databases and services.

For us the most important is LangGraph, released in 2024. It lets you describe an agent's work as a graph: steps and the arrows between them. In other words, the way you would draw a process on the office whiteboard. According to its makers, its users include Klarna, Uber and J.P. Morgan.

Look at the diagram. A request for a quote lands in the mailbox. The agent reads it and recognises what it is about. It prepares a simple reply straight away. For a quote it reaches for the price list and the customer's history. It hands an unusual case to a person at once. At the end there is always the same step: a person approves. In another business the same diagram will have different steps and different exceptions. And that is exactly the point.

What your business gets out of it:

• A person in the loop. The agent stops at the appointed place and waits for your "yes".
• Memory and resuming. The state of the work is saved after every step. After a break or a failure the agent picks up where it left off.
• A full trail. Every step, every tool used and every cost is recorded. You know what the agent did and why.
• A replaceable model. The same agent can work on a large cloud model or on a model running in our own UK or EU data centre. That is decided by how sensitive the data is, not by a supplier's whim.
• Open-source software. An agent built for you is not tied to a single AI vendor.

What is ready-made and what is made to measure

Not everything is built from scratch. The building blocks are ready: reading documents, working with a mailbox, searching the company's knowledge, writing to a spreadsheet or a system. What is made to measure is what matters most: the order of the steps, the exceptions, the language the agent uses when writing to your customers, and the places where it asks for your consent.

What it looks like in practice

1. A conversation and a process map. We sit down with the person who does this work today. We draw the steps, the exceptions and the places where a person decides.
2. One task to start with. We choose something repetitive and measurable.
3. An agent with tools and limits. It gets access only to what it needs. A person approves the important steps.
4. A probation period on real data. We measure the result and make improvements.
5. Further tasks. They go faster, because the agent already knows the business.

We work the same way ourselves. At Smartx the repetitive work is done by agents, and people approve the result.

An AI agent is an employee who does not sleep, does not fall ill and does not forget. It just needs a proper induction.

What this means for you

1. Start with a question, not a purchase. What do you do by hand in the business, even though you do not have to?
2. Choose one task with a measurable result. Number of cases, response time, number of errors.
3. Insist on three things. A person approves the important steps. Everything is recorded. The data stays where you agree it will.