Aria

When an organisation decides to adopt AI, the default path is buying licences: twenty dollars per person per month, everyone with their own chat window. That solves access and nothing else. Aria is the answer to what happens with everything else.

It’s a conversational assistant for general business use, built on the Claude API and deployed internally.

Cost: from seats to requests

A per-person licence charges for the possibility of using the tool. In any real organisation that possibility is distributed very unevenly: some people use it daily, and a majority open it twice a month. Paying per seat means paying the same for both.

Connecting directly to the API moves the cost to actual consumption. The organisation pays for what genuinely gets used, and granting someone access no longer carries a fixed monthly price, which, paradoxically, makes it possible to give access to more people.

The data that walks out unnoticed

This is the point that never appears in the pricing comparison. Someone needs to make sense of a report, pastes the whole spreadsheet into a public chat, and gets their answer. The problem is what stayed on the other side.

Aria operates with Zero Data Retention on the provider connection: the content of queries and attached files isn’t stored once the request has been served. The practical difference is that a confidential document can be used to get work done without that use meaning it left the company’s control.

This isn’t a feature bolted onto a chatbot at the end. It’s the condition that makes it acceptable for the thing to exist at all.

Scope: a work tool

A corporate assistant with no topic boundaries is, in practice, an entertainment subscription paid for by the company. Aria is scoped to work queries, and that boundary does two things at once: it keeps spending tied to operations, and it gives the assistant a clear identity. People know what it’s for.

Access: who needs it, not the whole directory

AI tools get distributed with the same logic as any sensitive system: by need. Aria has its own access control, so the organisation knows at any moment who can use the assistant, and can review that list, instead of finding out when the invoice arrives.

Why I think this is the right project

Conversations about AI in companies tend to stall on which model is best. The hard part comes earlier: defining the conditions under which an organisation is willing to let its information pass through a model. Aria is that answer turned into software, and it’s the same logic I applied to the Microsoft Fabric migration: technology only starts being useful once somebody has decided the rules.

Confidential project. There’s no repository or public demo.