AI assistant inside your own software
An assistant that reads your data and proposes actions rather than performing them.
Not a chatbot beside your system, but an assistant inside your application: it searches your own projects and quotes and does nothing without your approval.
Most AI integrations are a chat window that lets a language model guess. That works until someone copies a price, a stock level or a client name that never existed. In business software that is not a funny mistake, it is a wrong quote.
I build it the other way around. The assistant may not answer anything from memory: for every answer it calls a tool that pulls the real data out of your database, with the same permissions as the user asking. Calculations go through a calculation function, not through the model.
It does not write. If the assistant wants to create or change a task, client, ticket or quote, it produces a proposal that appears as a card on screen. Only when the user approves does anything happen. That is exactly how the assistant in Avitly works, my own platform for AV installers.
- Companies with their own application where a lot gets looked up and retyped
- SaaS owners who want to add AI without giving up their data
- Teams that want an assistant allowed to think along but not to change things on its own
Concrete deliverables,
no vague promises.
AI assistant inside your own software in the real world
This is how an assistant that does not guess works
Ask a question and watch what happens: first a tool that fetches your own data, only then an answer. If it wants to create something, that is a proposal you have to approve.
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.
Runs in your own browser. Nothing is stored or sent.
All demosWhat the track
looks like.
Decide what it may do
Which questions it should answer, which actions it may propose, and who may do what. This is the conversation that separates useful from dangerous.
Build the tools
For every kind of question a function that fetches the real data, with your permissions around it. No tool, no answer.
Proposals and approval
Creating and changing runs through a card the user confirms. What is not confirmed does not happen.
Measure and cap
Usage per organisation, a daily limit and a kill switch. You see what it costs before the bill arrives.
Frequently asked questions
about ai assistant inside your own software.
Does my company data go to an AI company?
Can it make things up?
Can it change things in our system by itself?
Which model are we locked into?
What does usage cost?
I also build
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.