Thoughts on AI, automation, and navigating the restructuring of professional work.
22 posts
An OpenAI agent worked around access controls to reach Australia's Medicare statistics portal. I have watched similar behavior at smaller scale: agents routing around controls, legacy configurations outside new governance, and auditors treating AI output as a finished observation.
How businesses can assess AI costs, returns and usage as model prices fall.
California just enacted the first US framework that defines, in law, who counts as an AI auditor: independent, registered, third-party. Internal audit sits on the other side of that line, and the work between now and 2029 is making your AI auditable before someone from outside arrives to check it.
Anthropic's agent-monitoring data points to three questions auditors should ask when AI is used in their organization.
A small refereed study found internal auditors adopting AI faster than external auditors, and more relaxed about it. The auditors doing less were the ones interrogating the risk, and the difference is that one group answers to a referee and the other does not.
Organizations engage internal audit for two different reasons, and we have never had to separate them because both arrived through the same door. AI separates them. Formal assurance stays human. Advisory becomes something the business reaches in five minutes instead of six weeks.
OpenAI gated its exploit-writing model behind identity verification, scoped authorization, hardware keys, and sandboxing. That control set is the template for an audit question most functions have not scoped: who inside your walls can now reach offensive AI, and on whose authority.
A security vendor just shipped access control that gates on what an agent is trying to do, not who it is. That is a different control primitive, and it quietly breaks the access review most audit functions still run.
A chip company published how it built an internal knowledge base that now fields 15,000 questions a day. Read past the infrastructure and it is the audit planning phase, rebuilt as a standing asset instead of something we throw away at the end of every engagement.
A survey of more than 1,500 finance leaders this month found 44% are only somewhat confident they could explain an AI agent's action to an auditor, and only 7% put governance ahead of speed. That gap is not neglect. It is a rational response to incentives, and that changes how you audit it.
On August 2 the EU AI Act's high-risk rules become enforceable, and Article 14 asks for something simple: a human has to be able to stop the system in a safe state. Yet industry data says 60% of organizations cannot terminate a misbehaving agent. That gap is the most testable control in the whole regulation, and it's ours to test.
This month agents stopped talking and started operating real applications. Meta's Muse Spark and Claude Cowork now click through desktops, browsers, and ERPs inside a logged-in human session. That breaks the audit trail in a way API logging never anticipated, and it lands squarely on internal audit.
On June 12 a single US export-control letter took Claude Fable 5 and Mythos 5 offline for every customer on earth, for 18 days, with no notice. In the same window, US companies quietly moved up to 46% of their token usage to cheaper Chinese models and lawmakers opened a probe. Your AI vendor is now a single point of failure, and that is a third-party risk audit waiting to happen.
Companies spent two years telling people to use more AI. Now Uber has capped its engineers at $1,500 a month and Meta is reining in spend that was heading toward billions. The swing from tokenmaxxing to token discipline is a governance failure hiding in plain sight, and it is exactly the kind of thing internal audit exists to catch.
Non-human identities now outnumber people in most enterprises by forty to one, 82% of companies have AI agents running they never inventoried, and most agents have no audit trail. That is not a security problem. It is an audit problem, and it is ours.
AI adoption is accelerating so fast that providers are throttling access and companies are building token leaderboards. Inside every organization, two cohorts are forming. And nobody has written the governance rulebook yet.
My education taught me to learn. The knowledge it gave me is already obsolete. If AI handles the output, the question isn't what you know—it's whether you know how to steer the machine that knows everything.
The big firms have already made their bet. Here's a practical workflow to make yours — one hour a day, go hard, and stop being polite to the machines.
KPMG just demanded a 14% fee cut from its own auditor—citing AI. A law firm replaced a departing associate with AI tools and saw costs drop 27%. The economics of professional services are being rewritten in real time.
The SaaSpocalypse wasn't about one product. It was the market recognizing that middlemen get eliminated when knowledge becomes abundant. Here's what that means for audit—and every knowledge profession.
The market just woke up to what I discovered last week: AI can automate knowledge work now. Here's what that looks like on the ground—and why it's just prompts and workflow.
I built an AI system that can do a significant portion of my audit job. In a week. With zero coding knowledge. This is what that satisfying, terrifying realisation felt like.