Writing | Agron Fazliu

A note, this week

When software can act, access is the first design problem

24 September 2026

A system that drafts text is one thing. A system that can open records, call tools, send messages or change data is a different kind of object, and most of the rules an organisation already has were written for the first one.

The question stops being which model is in use. It becomes who this thing is when it acts, and what it is permitted to do on its own. Those are not model questions. They are questions about identity, permissions, suppliers, records and the ability to undo something.

Which is why this cannot live in a policy document. A policy can say that AI use must be responsible. It cannot tell you which account performed an action at three in the morning, or whether that action can be reversed. Only the design can do that.

The uncomfortable part for most organisations is that the decision is already being made, quietly, by whoever connected the tool. Deciding it deliberately costs an afternoon. Discovering it afterwards costs considerably more.

See the six levels of authority, and what each one costs

Start here

Three that need no background at all.

Why AI is a question of governance before it is a question of technology

What an organisation can actually integrate, control and be answerable for, versus what it has been told the technology can do.

Read it

Why two things moving together does not mean one caused the other

The mistake that quietly sends product investment in the wrong direction, explained with ice cream and murder rates.

Read it

How to use data about people without it becoming surveillance

Where the line sits between understanding how an organisation works and watching the individuals inside it, and why that is a design problem rather than a legal one.

Read it

Questions I am working on now

Unfinished thinking, stated honestly.

These are the things I do not yet have a settled answer to. They are what the next few years of the work probably look like.

What changes when a system can act rather than only answer? Most of the rules we have were written for software that recommends, not software that does.
Which sensitive work genuinely needs models running on your own hardware, and which organisations are simply reassuring themselves at considerable expense?
When is buying another platform worse than simplifying the ones you already have? The second option is almost never on the agenda, because nobody sells it.
When is the right technology decision to remove something? Modernisation is usually treated as adding a platform. Some of the best work is subtraction: retiring systems, contracts and interfaces that stopped justifying their complexity.
How should a board judge an AI proposal when what the technology can do changes every few months and the business case does not?
Which parts of technology leadership have to be permanent, and which are better brought in for a defined stretch and then let go?

If you have a view on any of these, I would rather hear it than not.