How AI agents work, in plain words

Chatbot, automation, agent

Chatbot: Answers questions from a script or a set of documents. It talks; it does not act in your systems. Useful for FAQs, limited for real work.
Fixed automation (Zapier, n8n, Power Automate, RPA): Runs the same steps every time, exactly as someone designed them. Reliable while the input is tidy, and it stops when an email is worded differently or a PDF changes layout.
AI agent: Uses a language model to understand messy input, such as emails, scans, voice notes and photos, and a fixed set of actions to do the work in your systems. It copes with variation and asks a person when it is unsure.

What is inside an agent

The language model: The part that reads and writes language: the same kind of model behind ChatGPT and Claude. It understands an email, a PDF invoice, a voice note or a photo of a form. On its own it knows language in general, not your prices, customers or stock.
Your data, looked up live: When the agent needs a price, a stock level or an order status, it reads it from your systems at that moment. That is why its answers match your records instead of the model's memory, and why it can say "I don't know" when your records are silent.
Actions (tools): A short list of things the agent may do, each one built and tested for you: create an order in the ERP, send an email, book a slot, update a row in a sheet. The model chooses which action fits; the action does the work. Anything not on the list, it cannot do.
Rules and approvals: Which actions may run on their own and which wait for a person, amount limits, working hours, who gets asked about what. You set them, and you can change them.
The log: Every input, decision and action is recorded with the time and the reason, so you can check what the agent did and why.
Where it runs: In the cloud or on your own servers. The agent uses its own account in each system, with only the rights it needs.

One job, step by step

A supplier invoice, from the inbox to the accounting system.

A supplier emails an invoice as a PDF to your accounts inbox.
The language model reads the PDF: supplier, invoice number, lines, amounts, tax.
The agent uses its "find purchase order" action to fetch the matching order and goods receipt from your ERP.
It compares them. Eleven lines match; one line is 4% above the ordered price.
Your rule says price differences over 2% need a person, so it sends the accountant a short message with the line, both prices and the order.
The matching invoices are prepared as entries in your accounting software, waiting for approval. Nothing is paid by the agent.
Every step above is in the log.

How it connects to your systems

API: Most modern software (QuickBooks, Xero, Shopify, HubSpot, Salesforce, Odoo, Google Workspace, Microsoft 365) has an API: an official way for programs to read and write data. This is the first choice.
Email inboxes: Many jobs start with an email. The agent reads a shared inbox through Gmail or Microsoft 365, and replies or files messages.
Exports and imports: Older systems that only export files to Excel or CSV still work: the agent reads the export and prepares an import file.
Database read access: For in-house systems, read-only access to the database is often the simplest and safest route.
Screens, through a browser: When a system has no API at all, such as a supplier portal or a government site, the agent can work through its screens in a browser the way a person does, with its own limited account.

Running it on your own servers

For privacy, data protection or simply control, an agent does not have to use a cloud AI service.

What runs on your servers: Everything, if you want: the agent, its database, its logs and the language model itself. Open models such as Llama, Qwen, Mistral and Gemma can run on your own hardware, so documents, payslips and customer records never leave your network.
What you need: A server with a suitable graphics card (GPU) in your office or data centre, or a private cloud you control. How big depends on how much work the agent does; we size it after measuring your volumes.
The trade-off: Open models that fit on one server are smaller than the largest cloud models. For reading documents, pulling out fields, sorting email and answering from your own data they are usually enough. For harder reasoning we test both on your real cases before choosing.
Mixed setups: Sensitive steps can stay local while the rest uses a cloud model. For example, payslips are read on your server, and only figures with names removed go to a cloud model for a summary.
Without internet: Internal jobs, such as production planning on the factory network, can run with no internet connection at all.
What still passes outside: Jobs that use outside services still go through them: a WhatsApp message passes through Meta, an email through your mail provider, as they do today. Local hosting keeps the agent, your data and its decisions inside.

What keeps it safe

It sees only what you allow: its own accounts, with limited rights, in each system.
It can do only the actions on its list, and each action has limits.
Payments, signatures, hiring and anything else you choose wait for a person.
It answers from your data and says when it does not know, instead of guessing.
Everything it does is logged and can be reviewed.
The business APIs of the major AI providers do not use your data to train their models by default. Where that is not enough, the agent can run with a model on your own servers.

When an agent is the wrong tool

One-off tasks. Building an agent pays off on work that repeats every day or week.
Work without clear rules, where two good employees would decide differently and nobody can say why.
Decisions that carry legal or personal responsibility, such as hiring or medical advice. The agent can prepare them; a person decides.
Very small volumes, where an email template or a simple automation does the job.

How we start

Understand the job

We go through the work with the people who do it today: inputs, systems, rules, exceptions.
First version on your data

We build the agent on your real documents and systems and measure it on cases you already handled.
Run with approval

At first it prepares and a person approves everything. Routine cases run on their own only when you decide they should.
Keep it working

We look after it as your processes and systems change, and add the next task when you want.

Questions

Will an AI agent replace my staff?

It takes the repetitive part of a job: copying, checking, chasing, answering the same question. People keep the decisions, the exceptions and the relationships with customers and suppliers.

Can an AI agent make mistakes?

Yes, like any person or program. The design keeps mistakes small: it answers from your data, acts only within set limits, asks when unsure, and logs everything so a pattern of errors shows up and gets fixed.

Do we have to change our software?

No. The agent works with the email, spreadsheets and systems you already have. That is most of the work of building one.

Is our data used to train the AI?

Not by default with the major providers' business APIs. If your data must not leave your premises, the agent can run with a model on your own servers.

Can the whole agent run on our own servers, without OpenAI or any cloud AI?

Yes. The agent, its data and an open language model such as Llama, Qwen or Mistral can all run on your hardware. Local models are smaller than the largest cloud ones, so we test on your real cases first to confirm they do the job.

How is this different from ChatGPT?

ChatGPT is a general assistant you talk to. An agent is built for one job in your company: it is connected to your systems, follows your rules, does the work without being asked each time, and keeps a record.

What does it cost to run?

There is the cost of building it and a running cost for the model and hosting, which depends on volume. After a first call we give you a clear estimate for your case.

What agents do in practice

Which job would you hand to an agent?

Write a few lines about how the work is done today. We reply to set up an online intro call.