The policy is signed, filed and quietly ignored, and the organisation's AI risk keeps climbing anyway. The reason is structural, and it sits in what the document is for.
A written AI policy is a liability instrument before it is a safety document. Everything below rests on that, so it is worth stating plainly. A policy converts broad duties into specific, evidenced undertakings. We approve tools before use. We check outputs before release. We keep a register. We log what the model did and which version ran. Every one of those sentences becomes a promise a client, a claimant or a regulator can read back to you. Where practice does not match, the document stops being your defence and becomes an inventory of what you failed to do.
Once you accept that, the news from the other side of the world seems different.
What the New Zealand reporting actually establishes
Adoption and governance have come apart inside the same organisations, and the reporting shows the split in numbers.
hcamag.com reported that the Employers and Manufacturers Association's Workforce 2030 White Paper found only 13% of New Zealand companies have an AI policy in place, and that HR is commonly left out of AI adoption leadership even though workforce policy, training and the employment consequences of automation sit in its remit. The percentage matters less than what sits behind it. Other organisations are using AI now, and most of them have no written policy.
The same account includes Daniel Watson of Vertech, an Auckland cybersecurity and managed IT firm, who explains the practical exposure. Staff paste commercially sensitive information, customer data or intellectual property into systems with no written rules covering approved platforms, permitted inputs or who answers for the result. Watson raises a second problem that HR handles and operations rarely model: if AI absorbs the work juniors used to learn on, "where do the experienced people come from in ten years" (hcamag.com).
Training is where that shows first. hcamag.com's report cites Dayforce's 16th Annual Pulse of Talent, which found that 52% of companies surveyed globally give employees no AI training at all, and that 23% of New Zealand employees say their skills are becoming obsolete.
The Ministry of Business, Innovation and Employment's Responsible AI Guidance for Businesses, published in July 2025 and covered in the same hcamag.com report, recommends governance structures, staff capability building and human oversight of AI outputs. It is voluntary. Voluntary does not mean weightless. The UK Jurisdiction Taskforce reasoning, covered by Bird & Bird, is that the standard of care moves with what competent peers actually do, and published guidance is one of the things a court weighs when it asks what a reasonable organisation should have known.
What the measurement of the enforcement gap reveals
The gap is measurable, and the measurements are not flattering.
trussed.ai's guide to responsible AI policy states that 78% of employees admit to using AI tools their employer never approved, that 11% of what they paste into those tools is confidential, and that companies without enterprise AI governance frameworks lost an average of $4.4 million per incident in 2025.
Read the three together. Most of your AI estate was chosen by people who never filed a form. The most sensitive material is flowing into the tools with the least oversight. The loss, when it lands, runs to millions.
A PDF in a shared drive cannot stop any of that. It can only state an intention that no one is checking.
What the law already reaches in your AI estate
The absence of a UK AI Act does not create an absence of liability, and this is settled enough to plan around.
The NTIA's AI Accountability Policy Report records a joint statement by the Federal Trade Commission, the Department of Justice's Civil Rights Division, the Equal Employment Opportunity Commission and the Consumer Financial Protection Bureau that existing legal authorities apply to automated systems and new technologies just as they apply to other practices. The same report notes how hard it can be for a person harmed by an AI system to bring a claim at all, because proof that the system caused the harm is difficult to find.
That second point is a risk in your risk register, and it cuts in two directions. If a claimant cannot prove causation, your exposure looks smaller. If you are the party who failed to keep the records, the same evidential gap tends to be resolved against you. More on that below.
Bird & Bird's analysis of the UKJT Legal Statement on liability for AI harms explains how funds are distributed. Contract is expected to be the primary mechanism for allocating liability along the AI supply chain, with negligence filling the space where no contract governs. Professional liability runs in both directions: a professional can be negligent for leaning on an AI output without proper scrutiny, and negligent for declining to use a tool that competent practitioners in the field now use as a matter of course.
Where the exposure actually lands
A&O Shearman's analysis of the same Statement is blunt about who bears the risk. Foundation model developers are, in most circumstances, less likely to be liable for harm arising from unforeseeable or insufficiently tested uses of their general-purpose models. Application developers and deployers carry that risk instead. A&O Shearman calls the position the squeezed middle: you can disclaim liability downstream while being unable to hedge it upstream.
Translate that into a procurement decision. The organisation that selected the tool, integrated it, set its permissions and put it in front of staff is the organisation most likely to answer for what it does. Your AI policy is the record of whether those choices were made deliberately or absorbed by default.
imobisoft makes the commercial consequence explicit: contract is where risk is allocated whether or not anyone negotiated it deliberately, and the position you accepted by not negotiating is your risk position.
Records now carry legal weight
This is the most actionable finding in the whole body of UK analysis, and it is the one most AI policies miss.
A&O Shearman notes that AI disputes will turn on evidence, and that the ordinary but-for test will often still work. It fails where evidence was never gathered, was not retained, or cannot be reconstructed because the system behaved autonomously. English law can respond to that gap. It may draw inferences. It may treat a claimant's evidence more benevolently where the defendant's own failure to record or retain information created the uncertainty.
According to imobisoft's account of the Statement, the records that function as your legal position are:
- decision logs, showing what the system decided and when
- input and output retention, keeping what went in and what came out
- model and prompt versioning, so you can identify which configuration produced the output
- evaluation records, showing how the system performed before anyone relied on it
- the documented reasons the model was selected over the alternatives
- the human oversight record, showing who reviewed a consequential output and what they did about it
- the retention schedule, naming how long each of the above is kept and who approves deletion
Six months from now, the only organisation that can produce that set is the one that decided what it would keep before the first tool went live.
What the policy itself has to contain
Gerrish Legal's guidance treats the AI policy as the foundation for using the technology responsibly, and lists the components that make it functional: scope, aim and objectives; clear definitions; alignment with organisational values; governance and accountability with named AI leads or a defined approval pathway; ethical principles covering fairness, transparency and human oversight; classification of systems into permitted, restricted and prohibited uses; operational obligations covering documentation, incident reporting, data handling and impact assessment; integration with existing data protection, cybersecurity and governance policies; and a general provision covering exceptions, non-compliance and who to contact.
Read that list against most published AI policies. The classification section is usually missing, which means no one can say which tools are banned and which are merely unapproved. The operational obligations section is usually missing too, which means the document is silent about what happens when a failure occurs.
The level of detail should track the level of risk. Gerrish Legal makes that point directly: the more significant the potential impact of a system, the more robust the policy and procedures around it need to be, and responsibility for explainability stays with the organisation using the tool even when the tool is bought in.
Lawfulness is a precondition, and the rules are already published
The ICO's guidance on ensuring lawfulness in AI sets out what a UK organisation has to establish before processing personal data in an AI system. You break the processing down into distinct operations, identify a lawful basis for each one, choose the basis that reflects the true nature of the relationship with the individual, determine it before processing starts, document the decision, and record it in the privacy notice. You cannot swap bases later without good reason, and special category data needs both a lawful basis and a separate condition.
The guidance separates development from deployment and says a different basis may be needed for each. It also warns that consent is hard to make valid for complex processing, that performance of a contract rarely covers training, and that legitimate interests require the three-part test of purpose, necessity and balancing.
An AI policy that is silent on lawful basis leaves that decision to whoever happens to be using the keyboard.
What a Lawtelligence case reveals about AI governance
Lawtelligence's account of AI governance for legal practice provides a case that is useful for any sector. Lawtelligence reports that in April 2025 the High Court in R (Ayinde) v London Borough of Haringey 2025 EWHC 1040 (Admin) dealt with judicial review grounds citing five non-existent cases, and that the judge found the barrister's conduct improper, unreasonable and negligent and ordered the barrister and the instructing law centre to pay £2,000 each in wasted costs.
Lawtelligence notes that the Divisional Court judgment that followed, in Ayinde v London Borough of Haringey; Al-Haroun v Qatar National Bank 2025 EWHC 1383 (Admin), handed down on 6 June 2025, prompted the Bar Standards Board to issue formal guidance, valid from 18 May 2026, treating a barrister's use of AI tools as outsourcing a support service under rule C86 of the BSB Handbook, so responsibility for the work stays with the barrister.
On the solicitor's side, Lawtelligence records that the Solicitors Regulation Authority's Risk Outlook states that, as with any other technology in the firm, the solicitor remains responsible and accountable for the outputs from AI they use.
Neither regulator has banned the tools. Both have said existing duties on competence, confidentiality and supervision already cover them. If your firm has never written an AI policy, you have no record that anyone decided what AI may do, who checks it, or what happens when it gets something wrong.
What to do before the next tool is approved
Work in this order.
- Write the acceptable use policy first. Name the approved platforms, the categories of data that may go in, the categories that may not, and the person who signs off a new tool. Publish it where staff will find it.
- Classify every AI tool already in use, and publish the list against three labels:
: Approved. On the register, with a named owner, a permitted data scope and a documented selection rationale. Conditional: Usable only for defined tasks, with output checked by a named person before release. Prohibited. May not touch client, customer or personal data, with the list circulated and enforced.
- Name the individual who checks consequential AI output before it reaches a client, a court or a customer, and record the check. An unnamed checkpoint does not survive a deadline.
- Switch on decision logs, input and output retention, and model and prompt versioning before go-live. Set a retention period with your legal advisers, and write down who approves deletion.
- Document the model selection: the alternatives considered, the testing run, and the reason this tool won.
- Keep one person able to do the task without the system. hcamag.com's report contains Watson's warning that when a vendor changes its product, adjusts pricing or removes a feature a core process depends on, the business with no retained human capability has a continuity event on its hands.
- Review the policy at least annually, and immediately after a regulatory change, a new tool or an incident. Name the reviewer and log the date.
No part of this requires new law. The duties already exist in contract, in negligence, in professional standards and in data protection. What most organisations lack is the record showing they thought about any of it before the tool went live.
That is the difference between a policy that protects you and one that documents the exposure. Write the document as though it will be read aloud in a room you are not standing in, because one day it will be.
Sources
- AI policy gaps putting NZ businesses at risk, expert warns
- Liability Rules and Standards
- English law is already well positioned to deal with AI Liability and any resulting claims
- AI Liability Under English Law: UKJT Legal Statement
- AI Liability: English Law Already Decides Who Pays
- Responsible AI Policy: A Practical Guide
- Do You Need an AI policy?
- How do we ensure lawfulness in AI?
- How To Build An AI Governance Policy
Source: AI policy gaps putting NZ businesses at risk, expert warns, hcamag.com





