Companies deploy AI at the speed of competition and govern it at the speed of committee. Wolfmark closes that gap, so when an auditor, a regulator or a board asks how you know, you have an answer that holds up.
Plenty of organisations want AI governance without a regulator requiring it. The disciplines are the same. What changes is depth, and that is set by your own risk tolerance rather than by an examiner. It is also cheaper to build before somebody external makes it mandatory.
“Reviewed by management” was a good control when a person produced the number. It isn't evidence when a model did.
There are two different problems here. Some controls were never designed to carry AI output. Others were designed well and aren't operating that way. They are fixed differently, and telling them apart is usually the first step.
Fourteen questions, about four minutes.
New risks need new control responses, and new obligations keep arriving in jurisdictions you answer to. Controls take months to build and longer to evidence, so an organisation that is behind today falls further behind without effective governance.
Already in force
The privacy regulator confirms the Privacy Act applies to AI. ASIC finds AI uptake outpacing governance at financial services licensees (REP 798).
Existing licensee obligations & directors' duties already put the onus on firms to govern AI.
APRA CPS 230 in force: regulated entities must manage operational risk, including material service providers.
An AI vendor behind a critical operation can be one of them.
National AI Plan: no standalone AI Act. Existing laws to be strengthened instead.
The obligations arrive through law you already answer to.
A US export-control directive took two frontier AI models offline for every customer, worldwide.
Who can switch you off?CPS 230 requires APRA-regulated entities to plan for it.
Organisations covered by the Privacy Act must disclose, in their privacy policy, automated decisions that could significantly affect people (APP 1.7 to 1.9).
A disclosure you publish yourself, and answer for.
Already in force
Executive order directs federal challenges to state AI laws.
Until a court or Congress decides, state AI laws still apply where you operate.
COSO guidance on generative AI: where a control relies on AI output, that output needs the same evidence as anything else the control rests on.
Guidance, not law. But COSO is the framework most SOX filers use to assess internal control.
California's rules on automated decisions apply to hiring, pay and other significant decisions about Californians. Colorado's AI law also starts, though a court challenge has paused enforcement.
Affected people can ask for notice, an opt-out or a human review.
First AI risk-assessment attestations due to California's privacy regulator.
An executive signs it, under penalty of perjury. Expect to be asked for the evidence.
reaches Australian and US companies that sell into the EU, or whose AI output is used there (AI Act, Article 2)
Already in force
Prohibited AI practices are banned, and AI literacy duties apply.
Obligations for general-purpose AI models and the penalty regime apply.
Transparency duties apply (Article 50), and enforcement powers begin.
High-risk obligations apply to stand-alone uses such as hiring and credit decisions (Annex III).
High-risk obligations apply to AI in regulated products, such as medical devices.
Notable upcoming events
10 Dec 2026 · Australia
41 days until privacy policies must disclose automated decisions that significantly affect people.
1 Jan 2027 · United States
63 days until California's rules on automated decisions about people apply.
2 Dec 2027 · European Union
13 months until high-risk obligations apply to stand-alone uses such as hiring and credit decisions.
Feb 2024 · Air Canada
A customer relied on the airline's website chatbot, which gave him the wrong bereavement-fare policy. The tribunal held the airline liable and rejected its argument that the chatbot was responsible for its own words.
The lesson
Your chatbot speaks for the company.
Treat what customer-facing AI says like any published policy: owned, reviewed and tested before it goes live.
Jul 2025 · Replit
During a declared code freeze, a coding agent deleted the project's database. The platform then rolled out automatic separation of development and production databases.
The lesson
An instruction is not a control.
Keep development agents out of production by access design, and enforce a freeze technically, not by asking. Segregation of duties is just as relevant with agents as it is with humans.
Aug 2025 · Salesloft Drift
Attackers stole the tokens that connected an AI chat agent to customers' Salesforce environments, then used them to search those environments for passwords and access keys.
The lesson
Every AI integration holds a set of keys.
Know key AI integrations that hold access to your systems, limit what each token can reach, and watch what it does.
Oct 2026 · OpenCode
Attackers using AI agents breached a developer platform's console and took the AI provider keys customers had stored there. Some users found hundreds of dollars in charges overnight.
The lesson
An AI key is a credential with a spending limit.
Hold provider keys centrally, scope and rotate them, and set spend alerts that someone owns.
Your customers
Enterprise buyers ask for ISO/IEC 42001 alignment and AI controls in vendor due diligence before any regulator does.
The voluntary yardstick
Australia's Guidance for AI Adoption sets out six essential practices as the expected standard of care. Voluntary standards have a habit of becoming the measure after something goes wrong.
Wolfmark finds where AI is running, works out which uses actually matter, tests whether your controls can carry them, fixes the design, wires it into how you really work, and sets it up so it keeps running afterwards.
What is in scope, who owns AI risk, and which obligations apply.
An inventory of AI in use, including the AI nobody declared.
A risk tier for every use, so effort goes where the exposure is.
A gap assessment that separates design gaps from operating gaps.
Operating model, accountability, policies and review gates.
Controls attached to the real systems, pipelines and people.
Monitoring, reporting, internal review and continual improvement.
The approach lines up almost step for step with COSO's own roadmap in Achieving Effective Internal Control Over Generative AI (February 2026). It was built independently. The fact that they converge is the point.
Mapped to ISO/IEC 42001 · NIST AI RMF · EU AI Act · COSO · SOX 404
The method is the same for every client. The depth is not: a SOX filer gets financial reporting depth, a company that builds AI products gets provider depth, and a 40-person company gets a program sized for a 40-person company.
Every engagement starts with an initial conversation: thirty minutes, no obligation, on where you stand today, what you are trying to achieve, and what you are looking for. You get an honest answer on whether you need help, and with what.
Where AI is running, which controls rely on it, and what would not survive an auditor. Useful whether or not we work together after.
The design and the build: operating model, policies, review gates, and controls wired into how you actually work. The gaps fixed in the order that matters.
Ongoing oversight: inventory, testing and board reporting on a cadence, so the program stays honest after the project ends.
John Lum spent twenty years in internal controls and audit: at PwC, in pharma, and as head of internal controls at a consulting firm. AI is now landing in places controls were never built for: the financial close, customer decisions, vendor tools that switched AI on quietly. Wolfmark was built for that gap.
I haven't been doing AI governance for twenty years. Nobody has. I've been doing governance for twenty years, and AI is the newest thing it has to cover.
The person who scopes the work is the person who does it.
It can't be bought, inherited from a vendor, or deferred to a regulator.
Thirty minutes. No obligation.
