An AI agent that prepares the work.
You press approve.
For small and mid-sized companies where the same tasks come back every day: turning a request into a quote, working through the inbox, keeping track of what needs ordering. I build an agent that does that groundwork inside your own system, on your own data, and that sends or changes nothing before a human has approved it.
An assistant inside your own application
Pick a question the way an employee would ask it:
Every answer comes from a tool that fetches your own data, with the permissions of the person asking. Creating and changing never happens by itself.
The work that comes back every day.
A quote is an hour of retyping
The request is in an email, the prices in the price list, the client in the system. Someone puts it together by hand, every single time.
An inbox nobody keeps up with
Order questions, complaints and requests all land on one address. What is urgent sits between the newsletters.
Stock that runs out before anyone notices
Someone only sees an item is gone when a customer orders it. Then the supplier's lead time is your problem.
A chatbot that only talks
It gives an answer, but you still have to make the quote, write the email and place the order yourself. Nothing is off your plate.
Worried it will act on its own
An agent that emails or orders by itself sounds handy until it sends the wrong price to the wrong client.
No idea what it will cost
Usage is billed per request. An agent running all night without a limit is an open-ended bill.
How I build an agent you can trust.
Decide what it may do
Which task it prepares, which data it may read for that and which actions it may propose. This conversation makes the difference between useful and dangerous.
Tools on your own data
For every step a function that fetches the real data, with the user's permissions. What the agent cannot find through a tool, it does not know. Prices and margins go through a calculation function, not the model.
Proposing instead of executing
A quote, a reply or an order appears on screen as a proposal. Only when someone presses approve does anything happen.
Limits and an off switch
Usage per organisation, a daily limit and a switch that turns the whole thing off. You see what it costs before the bill arrives.
What an agent does that a chatbot does not
A chatbot answers a question. An agent carries out a series of steps: it reads a request, looks up the client and the products, works out the price and prepares a draft. The difference is not the model but what it is allowed to touch.
If you want an assistant that talks to visitors on your website, adding AI to your website is the right page. If you first want to know where AI pays off most in your business, start with the step plan on AI integration for your business.
| Chatbot | AI agent | |
|---|---|---|
| What it does | Answers a question | Prepares a task in steps |
| Where the data comes from | Texts and pages | Your own database, through tools |
| What is left afterwards | An answer in a chat window | A quote, email or order as a draft |
| Who decides | The reader | A human who presses approve |
What an agent can do for you
The tasks that work have three things in common: they come back often, the data is already in a system, and a human can judge the result in half a minute.
If the system underneath is still missing, that comes first: a quoting system or custom inventory software.
A quote from a request
A request comes in by email or form. The agent recognises the client, looks up the products in your price list, lets the calculation function set price and margin, and prepares a draft quote. You read it, adjust and send.
Customer service: sorting the inbox
Incoming email gets a subject and an urgency and lands with the right person. For an order question the agent fetches the status and writes a draft reply with the real data in it. A colleague reads it and presses send themselves.
Stock: flagging and preparing orders
The agent looks at stock, open orders and what normally goes out each week, and flags what is running low. It prepares a purchase order per supplier that someone checks before it goes out.
Looking things up in your own data
Which quotes are still open for this client, what was delivered last month, where is this project stuck. A question in plain language, answered with the real data alongside.
AI automation for small businesses: when it pays off
An agent pays off in a company of five to fifty people where a few people spend much of their day on groundwork: looking up, retyping, putting things together. There, automating one task well is worth more than a broad AI project.
And sometimes you do not need an agent. If a task is exactly the same every time, a regular automation or integration is cheaper and more predictable. If it is about writing or summarising for individual staff, a ChatGPT or Copilot subscription will do. I will tell you so.
- The task comes back every day or every week
- The data is already in a system, or can be
- There is language in it: email, requests, descriptions
- A human can judge the result quickly
- Mistakes can be caught before they go out
- Someone does the work by hand today
The rules that keep the agent safe
An agent that may email, order or change prices on its own is a risk no time saving is worth. So I build it the other way round: it may read and propose, a human decides.
Only the request and the data needed for that one task go to the provider you choose. Conversation content is not stored by default; usage is measured.
- Answers only from what a tool returns
- Calculations through a function, not the model
- The same permissions as the user, no more
- Every write action is a proposal someone approves
- Daily limit per organisation and one switch to turn it all off
- Keys on the server, never in the browser
The proof: the assistant in Avitly
Avitly is my own platform for AV and IT installers, which I use daily myself. The assistant in it works exactly as described above: it searches projects, quotes and equipment through tools, with the user's permissions, and points out on the floor plan what has not been quoted yet.
If it wants to create or change a task, client, ticket or quote, a card with the proposal appears. What is not confirmed does not happen. Claude, Gemini or GPT can sit behind it; switching is a setting, not a rebuild.
What it costs
The build is a fixed amount you hear up front, excluding VAT. You pay for usage yourself at the model provider, capped with a daily limit. How pricing works is on pricing. If there is no application yet for the agent to live in, it starts with custom software.
Agent in an existing application · € 2,950 to € 5,500
Your system already exists and the data is in one database. The agent gets a tool per step and the user's permissions. Lead time three to six weeks.
Agent on data from several systems · € 4,500 to € 8,500
The price list is in the accounting package, stock in the webshop, clients in a CRM. First the integrations, then the agent on top.
New application with an agent in it · € 6,500 to € 12,000
There is no system yet for the agent to live in. Then it becomes a software project, with the agent as one of the parts.
Clients rate me 5 stars on average.
“I am extremely happy with the result of my new website for DoubleStreet Australian Cobberdogs, made by Webframer. From first contact, Webframer radiated expertise and creativity.”
“Webframer helped me enormously with the Veldvisuals website. It is not only good to spar with Bas about building the site, but also in the aftercare and whenever changes have to be made. Recommended!”
“Great to work with. Listens carefully to what we want. When we ask for changes, they are made expertly.”
“Really pleasant to work with and short lines of communication!”
“The website looks great!! Definitely recommended!”
“Excellent! Fast service and good contact!”
Translated from the Dutch originals on Google.
I am Bas Voets. Since 2021 I have been building websites, client portals and business software for small and mid-sized companies, from Hazerswoude-Dorp in the Netherlands. From first call to launch you talk to me, not to an account manager.
What is the difference between an AI agent and a chatbot?
What does AI agent development cost?
Can the agent email or order on its own?
Can the agent make things up?
Which AI model are we tied to?
Do I actually need an agent?
Tell me what you want to build.
A half-hour call. You get an honest answer about what it costs, how long it takes and whether building it is smart at all.