The AI governance conversation is older than you think

A recent Financial Times webinar on AI governance in Asia made a point that deserves more attention than it usually gets:  as regulatory frameworks lag behind AI adoption, boards must move quickly enough to capture AI's benefits without exposing themselves to legal, regulatory, and operational risk.  The maxim is not, "Move fast and break things," but, "Move fast, safely."

That's being framed as something new.   It isn't.

But it is timely.  Upcoming changes in Australia's privacy legislation will require organisations' privacy policies to disclose when automated processes are used substantially to make decisions that affect end users' access to services (read: AI-based workflows are deeply integrated into business operations).  Just as waves of data protection regulation globally have converged around GDPR-like principles, similar extension and normalisation of regulation is likely in response to the current AI zeitgeist - largely through the application of existing principles of security, control and explainability, not net-new obligations.

Boards have always held ultimate accountability for risk management and corporate stewardship.  That doesn't change because the system making decisions involves a large language model rather than a spreadsheet or a human.  Whether decisions are made and actions taken in an organisation by consensus, by whoever shouts loudest, by a logical flow codified in software, or - in our brave new world - by an AI agent operating on behalf of a business unit, the board is still accountable for the outcomes.  The mechanism of action changes.  The obligation doesn't.

The challenge is that AI looks different enough to feel like uncharted territory.  It isn't.  And boards that treat it as alien will either overcorrect into paralysis or reach for application-layer fixes when the problem runs deeper.  Neither is good governance.

---

Strip it back to first principles

At its core, any system that processes information and produces an outcome has three components: an input, a processing step, and an output.  AI is no different.

When you strip the conversation about AI governance back to these components, most of it becomes familiar ground - with some important qualifications at the edges.

The input: Data

The data that feeds AI systems is the same data that organisations have always had - structured records, transactional systems, CRM data, financial reporting.  The foundation hasn't changed.  What's expanded is the variety:  We're now pushing deeper into unstructured inputs.  Documents, emails, call recordings, images, scraped web content.  This is the "fourth V" of big data - variety - becoming commercially significant in ways it wasn't when most data governance frameworks were written.

The FT webinar cited a Hong Kong compliance check finding that around 80% of organisations are already using or experimenting with AI in daily operations.  That uptake is happening largely on top of both structured data systems and an unstructured document layer.  The governance gap isn't primarily in the AI layer - it's in the input layer that the AI is consuming.

The foundational question for boards on inputs is:  Do you know what data and documentation you have, where it came from, and under what terms it can be used?  That's not an AI question.  It's fundamentally the same question that data protection frameworks across the region have required organisations to answer for years.

Australia's Privacy Act 1988, via the Australian Privacy Principles, has long required organisations to understand what personal data they hold, why they hold it, and how it can be used.  The Privacy and Other Legislation Amendment Act 2024 goes further:  from December 2026, organisations must disclose in their privacy policies when they rely on automated processes to make or substantially assist in decisions that affect individual rights - access to services, credit, employment. 

The obligation to know your data is becoming an obligation to explain what your systems do with it.

Singapore's Personal Data Protection Act 2012 imposes parallel obligations around consent, purpose limitation, and data accuracy, with the Personal Data Protection Commission increasingly active in enforcement.  In India, the Digital Personal Data Protection Act 2023 - with full compliance required by mid-2027 - establishes similar foundational requirements:  explicit consent; defined purpose; and rights for individuals to access, correct, and request deletion of their data.

None of these laws is specifically about AI.  All of them apply directly to what AI systems do.

The processing: Familiar machinery, new wrapping

The processing layer of AI is newer in appearance than in substance.

Workflow automation and RPA tools have existed for years - systems that take an input, apply a set of rules or instructions, and produce a result or trigger a downstream action.  Data science models added a layer of stochastic behaviour:  temperature sampling in large language models means the same prompt can produce different outputs at different times.  There are similar characteristics of neural networks trained with random seeds, Monte Carlo methods, and reinforcement learning policies that sample probabilistically from distributions over possible actions.

None of this is entirely new territory for risk management.  Data science model risk management frameworks - particularly in financial services - have long required organisations to understand, validate, and monitor the behaviour of stochastic models.  The discipline exists.

The evolution of customer support is a useful illustration.  Decades ago, many organisations handled customer queries in house, with teams following scripts and escalation procedures.  As cost pressures grew, BPO providers took on that function - offshore teams operating under service agreements and SOPs that prompted the right approach without mechanically enforcing every step.  Customer service agents and managers exercised judgement within boundaries.  Then came IVR systems and early chatbots handling high-volume, low-complexity queries, routing more nuanced cases to humans.  The hybrid model became standard.  Now, LLM-based agents are beginning to handle multi-turn customer conversations, accessing account records, and initiating actions - refunds, bookings, account amendments - without necessarily having a human in the loop.

The governance question at each stage has been the same:  what are the boundaries of each actor's or system's authority, how are those boundaries enforced, and what happens when something goes wrong?  The answer has always involved a combination of policy, monitoring, and human oversight calibrated to the risk level of the action.

Agentic AI does not change the governance framework. It raises the stakes, which means the calibration needs to be more deliberate - not invented from scratch.

What makes this urgent is the pace and informality of adoption.  The FT webinar suggested that more than half of employees are using AI tools that their employer hasn't formally provided or governed.  The BPO analogy breaks down here:  no service agreement was signed by management, no SOP was agreed, no escalation path was defined.  Employees are building informal agentic workflows on top of consumer AI tools, and the processing layer that those workflows represent is entirely outside the governance structures that nominally apply.  That's not an AI problem. It's a process governance problem that happens to involve AI.

Some may argue that the risk from information leakage into an AI system is greater than leakage of the same information to a third party (whether to a software platform or to a human).  Again, the risk of misuse of information once it is outside the boundaries of the organisation is not new, but AI does present a materially higher incidence of misuse given the relatively small fraction of malicious third parties historically versus a currently commonplace practice of refining AI models based on user prompts.  This doesn't fundamentally change the duty of the board to protect the organisation from impact, but it does raise the stakes for failing in this area.

The output: Collateral and actions

The outputs of AI systems are also less novel than the conversation suggests.  AI can produce collateral - data, reports, recommendations, generated content.  And AI can trigger actions - updating a record, sending a communication, executing a transaction.

Organisations have had systems that generate data and take action in downstream applications for years.  Reverse ETL pipelines write data back from warehouses into operational systems.  Action hubs trigger workflows based on rules or model outputs.  The principle that a system can observe a state, make a determination, and act on it is not new.  What's changed is the scale, the speed, and the range of actions that a single system can now plausibly take without human review at each step.

The more significant shift is in surface area.  When a system's scope was bounded - an RPA bot that processed invoices, a rules engine that assessed credit applications - the interfaces requiring governance controls were limited and well-defined.  AI has expanded by an order of magnitude the number of interfaces through which systems are likely to interact with data and external services.  A single LLM-based agent, depending on the tools it's been given access to, might query a database, draft and send a customer email, update a CRM record, and initiate a payment - in a single session, based on a natural language instruction.  Each of those actions touches a control surface that historically had its own governance layer.

The challenge isn't that controls don't exist. It's that they now need to be applied coherently across a far wider and less predictable set of interactions and outputs.

---

Think about three fundamental themes within regulation

Regulatory responses to AI across Asia remain, as the FT panel noted, fragmented and incremental.  Governments have existing data protection and cybersecurity frameworks but there is noise around comprehensive AI legislation.  For boards, this creates genuine uncertainty about long term obligations.

But underneath that uncertainty, there are foundational principles that have been set for years - and that most jurisdictions, regardless of how they approach AI specifically, will continue to anchor to.  The scope of regulatory expectation largely collapses into three questions:

1. Do you know what data you have, where it came from, and how it can be used?

Data provenance, classification, and contractual or regulatory constraints on use.  This is the foundation.  Without it, every subsequent governance step is guesswork.  The Australian, Singaporean, and Indian frameworks given as examples above all start here - and the 2024 Australian amendments make clear that "knowing your data" now includes knowing which data is being fed into automated decisions affecting individuals.  Coming into the 10th anniversary of GDPR, these are not foreign concepts for businesses globally.  While data hygiene is admittedly a mixed bag in the vast majority of organisations, an appreciation for these principles is universal.

2. Is your use of data controlled by appropriate policies and structures?

Role-based and attribute-based access controls.  Data collection notifications and retention policies.  Approval workflows for sensitive use cases.  These are not AI-specific requirements  - they're the infrastructure of responsible data use, and they apply regardless of whether the system consuming the data is a human analyst or an AI agent.  Both humans and agentic systems will have varying levels of compliance with these structures, one through (wilful or otherwise) blindness, the other through systematic randomness.  The governance challenge and requirements here are not new, but the application of the policies will differ at the margins.  We expect regulators will continue to, and hence boards must, adopt a risk-based approach with policy scrutiny and criticality increasing as impacts are closer to the end user.

3. Are the decisions you make using data explainable and in line with your obligations?

This is where AI introduces the most genuine novelty - but even here, the underlying obligation isn't new.  Financial industry regulation, including the Monetary Authority of Singapore's FEAT Principles - Fairness, Ethics, Accountability, Transparency - have required financial institutions to document AI systems used in decision-making and proactively drive board-level governance since 2018.  Concepts like these appear in general data protection regulation from Australia to the EU, suggesting a base level of requirements will converge across industries and countries.  The question of whether a lending decision is defensible - whether you can explain why a customer was approved or declined - predates modern AI entirely.  While AI may make that question harder to answer it is not a new question.  The governance challenge arises, somewhat ironically, from LLMs needing to be used to explain the actions of other LLMs.

---

Build on existing tooling - using AI to help

The main thrust of this is straightforward:  while the technology is new, the components of risk are not.  And critically, the approaches to manage those components are mature.

We have model risk management frameworks developed over years of deploying statistical and ML models in financial services.  We have access governance tools - RBAC, ABAC, dynamic masking and obfuscation - that control who can see and use what data.  We have data quality testing, lineage tracking, and observability pipelines that make data flows auditable.  We have semantic layers that create single sources of truth:  consistent definitions of customers, products, and metrics that prevent the fragmentation that makes both data governance and AI governance impossible.

There is one area where AI genuinely improves the governance picture, and it's worth naming.  Translating regulatory text, contracts, and internal policy documents into machine-enforceable rules has historically required significant manual effort - legal, compliance, and technical teams working together to codify natural language obligations into data access controls, metadata tags, and process gates.  LLMs are now capable of accelerating this materially:  parsing policy documents, extracting structured constraints, mapping those constraints to data fields and access controls, and generating audit-ready explanations in plain language.  The technology that creates governance challenges also provides some of the most practical tools for addressing them.

The surface area over which all of this needs to be applied has expanded significantly. That's a real operational challenge. But it's a challenge of scale and integration, not of inventing entirely new disciplines.

---

What boards should actually do

The FT panel's recommendation - that boards take a more proactive stance on AI governance - is right.  But "proactive" doesn't mean starting from scratch.  It means engaging the professionals who already know this terrain, asking better questions, and resisting the temptation to treat AI as a special category that requires a parallel governance structure.

Data teams understand information lineage, quality, and access control.  Cyber and security teams understand exposure surfaces, monitoring, and incident response.  Legal and compliance teams understand the regulatory frameworks and how existing obligations translate to new contexts.  The board's job is not to understand every technical detail.  It is to enshrine the right questions in policy and structure, ensure those teams are resourced and empowered, and hold the organisation accountable for answers.

The three questions in the regulation section above are a reasonable place to start.  Can you answer them today, for your current systems?  If not, that gap exists independently of AI - and AI is about to make it more visible. 

Next, look at the mechanics of answering these questions on an ongoing basis.  Do mandates for existing sub-committees cover the risk areas or are new forums required?  Does the board have access to the right skillsets to challenge management's assessment of risk?  Is balanced scorecard reporting set up to provide visibility into performance?

It is our view that AI is new technology applied to familiar risk categories.  Boards that approach it as an extension of existing governance - rather than an alien challenge requiring entirely new frameworks - will make better decisions faster and run less risk of either paralysis or addressing symptoms rather than causes. 

---

Cruxdata builds data infrastructure for SMEs and startups across APAC. We help growth-stage companies get the data foundations in place before the governance questions become urgent.