How to build an agent in your company.
One complete example: a hardware store that quotes by email. Four pieces, ten test emails and one person who approves before anything goes out.
What does an agent do that a chat does not?
With a chat, the person still does the work: copy, ask, paste and send. The AI only answers what it is told.
An agent picks up the request where it arrives, looks up what it needs and leaves the reply ready for review. The person goes from doing the task to approving it.
With a chat
- 1Copy the email.
- 2Ask for the price.
- 3Copy the answer.
- 4Send it by hand.
With an agent
- 1The email arrives.
- 2It checks the price list.
- 3It works out the total.
- 4It leaves the draft ready.
Every agent has four pieces.
- 1
Instructions
What it does and when it asks
- 2
Knowledge
What it needs to know
- 3
Tools
Where it reads and where it writes
- 4
Review
Who approves before it goes out
If one fails, the agent fails. Without instructions it improvises, without knowledge it makes things up, without tools it cannot act and without review nobody catches the mistake in time.
A task that repeats every day.
The first agent is chosen for being boring: a task someone does twenty times a day, always the same way.
The test fits on one page. Write down the steps. If any of them means reading and understanding text someone else wrote, like an email, the task is for an agent.
If its steps fit on one page, it works
- 1Read the quote request.
- 2Look up each product on the list.
- 3Reply with the total.
If the steps never change and there is nothing to read, a simple automation is enough.
It is built in the tools the company already uses.
None of them needs a development team or a new system. The choice depends on where your team already works.
Zapier, Make or n8n come in when the agent has to act on its own after an event, like a new email, without anyone opening anything.
Three options
Claude Cowork
if the company uses Claude
ChatGPT workspace agents
on the Business and Enterprise plans
Zapier, Make or n8n
so it reacts as soon as an email arrives
Write them as if for someone new.
Think of the first day of someone very capable who does not know the company. Anything you leave out, they will assume.
Three parts: who it is, what it does and when it stops to ask. The third is the one most often forgotten, and the one that prevents the expensive mistakes.
You are the quoting assistant at Ferretería Andina. When a price request arrives, look up each product on the list, add up the total with VAT and draft the reply. If a product is not on the list, do not make up a price: let María know.
- Who it is
- What it does
- When it asks
It answers with the company's own data.
An AI model knows a lot about the world and nothing about your hardware store. This week's price of cement is only in your files.
It gets the same documents a person would check to do the task, and they are kept up to date. With September's list, it quotes September's prices.
- XLSXlista-de-precios-octubre.xlsxTo quote with real prices
- PDFcondiciones-de-entrega.pdfTo answer lead times and shipping
Without these files, it answers with made-up prices.
It only reaches what it needs.
Every access is an open door. To quote, the agent needs to read email and the price list. Nothing else.
It is the rule already applied to anyone who joins the company: the permissions for their task, and more only when the work calls for it.
Agent access
- Yes
- Price list
- Yes
- Bank account
- No
If it does not need it for this task, it does not touch it.
Ten real emails before switching it on.
Before connecting it, it is given ten requests that actually came in. The easy ones and, above all, the odd ones.
Every mistake is fixed in the instructions so it does not come back. Then all ten run again.
Email 3: asks for something not on the list
Let María know
Correct
Email 7: asks for a discount
Made up a 10% discount
Error
Added to the instructions: “Do not offer discounts. Pass the request to María.”
For the first few weeks, everything goes out as a draft.
Whoever used to write the quote now reads it, corrects it if needed and approves it. That person is what makes the agent reliable.
When weeks go by without corrections, the team decides which cases can go out on their own. That decision is made by someone with a first and last name.
Draft for Constructora Ríos
20 bags of cement at $8.40 each. Total with VAT: $193.20.
A person reads it and approves it. The agent sends nothing on its own.
How much time does it give back to the team?
This is the example's arithmetic: twenty quotes of ten minutes each take 3 h 20 min of a working day. Reviewing each draft takes a fraction of that.
With the first agent running, the next repeated task is chosen. That is how it gets built: one piece at a time.
Quotes per day
20 × 10 min
The team gets back
0 h 0 min
Of an 8 hour working day
Once it works, move on to the next task.
Shall we build the first one in your company?
We choose the task with you, write the four pieces and test it with your real emails.