THE SIGNAL IN ONE SENTENCE
New Zealand has delivered a useful sentence to every business currently trying to hide a messy decision behind a shiny model. The AI did it is not a legal defense. On November 18, the Commerce Commission published guidance explaining how the Fair Trading Act and Commerce Act apply when businesses use artificial intelligence. The regulator’s message is plain: the same rules apply whether information, decisions or communications come from a person, an algorithm, an AI tool or an AI agent. The business remains accountable for the outcome. That covers more than the chatbot saying something odd. The guidance reaches advertising, product claims, images, recommendations, customer responses, personalized prices, targeted offers, automated pricing, market behavior, access to data and other business decisions. It names fake endorsements, fake reviews, ghost shops, manipulative marketing and exaggerated AI claims as fair-trading risks. It also names cartel conduct, self-preferencing, tying, bundling, predatory pricing and misuse of market power as competition risks. The Commission is not announcing a new AI law. It is explaining how existing duties follow the outcome even when software helped produce it. Boards and senior leaders are expected to know where AI is used, how it influences decisions, what products and outcomes it creates and whether safeguards are working. The plain signal is this: trace every consequential customer outcome to a human owner before the system goes live. An outcome trace is a short operating record. It identifies the use case, the affected person, the inputs, the AI contribution, the final action, the legal duty, the evidence supporting the result, the human owner, the review route, the correction offered and the records retained. It is not a 90-page policy that everybody signs and nobody can find when the price doubles. Start with the outcome, not the model. A business may use five models, three vendors and one alarming spreadsheet to produce a personalized offer. The customer experiences one price. That price is the thing to trace. Who received it? What data changed it? Was the price generated, recommended or merely formatted by AI? Which person or team approved the business rule? What evidence shows the offer was accurate and not misleading? Could a customer understand why it differed from another offer? Who can pause it? How does someone challenge an error? If the company cannot answer those questions, saying that a human was somewhere in the process does not add much comfort. The regulator’s Fair Trading Act guidance makes the first ownership rule straightforward. AI output alone is not evidence that a claim is true. A product description, advertisement, chatbot response or synthetic image still needs reasonable grounds. The business remains responsible for the overall impression, including what was left out. So give every public claim an evidence owner. The outcome trace should link the claim to the source record, the date it was checked, the limitations that must appear beside it and the person authorized to publish it. If a model rewrites the claim, compare the final copy with the evidence again. If the evidence expires, the claim should stop automatically or return to review. That catches the ordinary failures before they become legal philosophy. A travel chatbot invents a refund policy. A retailer’s generated image makes a product appear larger than it is. A health supplement ad creates a fake endorsement. A marketplace fills empty inventory with fake reviews. A company calls a basic rules engine AI-powered because the phrase tested well. The model may have produced the sentence. The business chose to put it in front of a customer. The correction path matters as much as the approval path. Every customer-facing AI system should provide a route to a person who can inspect the actual record, correct the source data, reverse the outcome and prevent the same error from repeating. A generic contact form that returns the customer to the same bot is not review. It is a loop with branding. Record the original output, the action taken, the complaint, the human decision, the correction and the system change. That creates evidence for the customer, the business and, if necessary, the regulator. Personalized pricing needs an especially visible trace. The Commerce Commission says AI can tailor prices, discounts and offers using customer data, and that practices must be clear, accurate and not misleading. In some circumstances, a business should consider whether it needs to disclose its use of AI to avoid a misleading impression. The operating question is not merely whether the algorithm found a profitable price. It is whether the business can explain the rule, defend the data, test for misleading patterns and stop unfair pressure. Trace the price inputs. Separate features that affect cost from features that merely predict willingness to pay. Test whether protected or vulnerable groups receive systematically worse offers through proxy variables. Compare the displayed discount with the real reference price. Check whether urgency messages are true. Give the customer a way to correct stale location, loyalty or account information. The owner should see distribution, not just average revenue. A pricing system can increase sales overall while quietly assigning the most aggressive offers to people who appear stressed, isolated or unlikely to complain. Manipulation can be statistically effective. That does not make it fair. Competition law adds a second owner card. The Commission says businesses remain responsible when AI influences pricing, customer targeting, output or other competitive decisions. Using an independent technology provider does not automatically protect a business from liability. Algorithms may coordinate market behavior, use competitors’ sensitive information or produce outcomes that reduce competition. This is where vendor procurement needs to become less ceremonial. Before adopting a pricing or decision tool, the buyer should know which data the system uses, whether competitor information contributes to outputs, whether data from different customers is separated, how recommendations are generated and what safeguards prevent anti-competitive outcomes. The Commission explicitly recommends meaningful human review rather than simply accepting a recommendation. Put a competition owner on the trace. That person should review the objective given to the system, data sources, information-sharing arrangements, market outcomes and interactions with competitors. They should be able to identify whether a platform favors its own products, whether a bundle forecloses a rival, whether automated prices move in suspicious lockstep or whether a model is using confidential market information. Do not wait for the model to announce that it has discovered cartel conduct. Software rarely sends that calendar invite. The trace should include a baseline and a stop rule. How did prices, rankings or access behave before the system? What change would trigger investigation? Who can disable the tool without waiting for a vendor’s next quarterly release? Which logs will be preserved? Human oversight needs similar specificity. The phrase often means that a person can technically look at an output. Meaningful review means the person has time, information, authority and an alternative. Can the reviewer see the evidence behind the recommendation? Do they understand the legal duty? Can they reject the result without being penalized for slowing the queue? Is disagreement recorded? Are overrides analyzed? Does the owner have authority to suspend the system? If a person approves 400 recommendations an hour, the human may be part of the interface rather than the safeguard. Boards need an outcome map, not a model inventory alone. Knowing that the company uses a language model, pricing algorithm and recommendation engine is useful. It does not show which customer outcomes they influence. Map each system to the claims, prices, rankings, refusals, approvals, communications and market decisions it can affect. Then attach an owner to each outcome class. One model may have several owners because its uses carry different duties. The marketing team may own advertising claims. Customer operations may own chatbot remedies. Revenue management may own personalized pricing. A competition lead may own automated pricing and marketplace rankings. Security may own access and incident response. The owner should be a role with authority, not the name of the nearest data scientist. Technical teams can explain how a system behaves. They should not be left to decide the company’s legal position, customer remedy or competitive strategy by accident. Smaller businesses need a lighter version, not no version. A local retailer using generated product copy can keep a claim sheet with source records and review dates. A service business using a scheduling chatbot can log promises, cancellations and complaints. A marketplace can sample recommendations and prices weekly. A supplier using automated quotes can retain the inputs, rule version and approving person. The trace can fit on one screen: Outcome. Customer or market affected. Data used. AI role. Evidence. Legal duty. Owner. Approval. Disclosure. Review route. Correction. Stop rule. Record location. That is small enough to use and specific enough to matter. The Commission’s guidance also needs careful reading. It describes the regulator’s current view of existing law. It is not a court decision, a definitive legal opinion or a technical certification. The regulator says policymakers may still consider additional tools or safeguards as AI evolves. Businesses should get legal advice where risks are significant. They should also resist using that caveat as a reason to wait. The basic operating duties are already visible. Do not mislead customers. Keep claims substantiated. Treat consumers fairly. Make competitive decisions independently. Do not use a vendor as a liability curtain. Monitor systems. Keep records. Give humans meaningful authority. The outcome trace turns those principles into something a team can inspect before and after launch. New Zealand’s intervention matters beyond New Zealand because it refuses a convenient category error. AI does not create a separate universe where ordinary business conduct becomes nobody’s fault. It changes the mechanism. It may change speed, scale and opacity. It does not erase the company standing behind the result. That is the useful bit. When an AI system writes, prices, recommends, ranks, refuses or persuades, follow the line to the business decision and the person authorized to own it. If the line disappears into a vendor dashboard, the system is not ready. Give every outcome an owner. Give every owner evidence. Give every affected person a route back. Accountability should be easier to find than the chatbot.
01
WHAT ACTUALLY CHANGED
New Zealand’s Commerce Commission released AI guidance for businesses on November 18, 2026.
The guidance explains how the Fair Trading Act and Commerce Act apply to business use of AI.
The Commission says businesses remain accountable whether an outcome comes from a person, algorithm, AI tool or AI agent.
Boards and senior leaders are expected to know where AI is used and monitor its effects, safeguards and risks.
Fair-trading risks named by the regulator include misleading content, unsubstantiated claims, fake endorsements, fake reviews, ghost shops, personalized pricing, manipulative marketing and AI washing.
Competition risks include cartel conduct, self-preferencing, tying, bundling, predatory pricing and misuse of market power.
The guidance says an independent technology provider does not automatically protect a business from liability.
The Commission recommends meaningful human review, regular system audits and records of data sources, design, governance, risk assessments and decisions.
The Commission is applying existing law rather than announcing a new AI-specific statute.
The regulator says additional tools or safeguards may still be considered as AI evolves.
02
WHY THIS MATTERS
Customers experience the final claim, price, recommendation or refusal, not the internal division of labor between people and software.
AI can increase the speed and scale of misleading conduct before ordinary review catches it.
A generated statement still needs evidence and reasonable grounds.
Personalized offers can become misleading or manipulative when data use and pricing logic are hidden.
Automated pricing and targeting can create competition risks even without an explicit human agreement.
A vendor contract does not replace the buyer’s responsibility for outcomes.
Human oversight is weak when reviewers lack evidence, time, authority or a genuine alternative.
A model inventory does not show which customer and market outcomes each system can change.
A named owner makes approval, monitoring, correction and shutdown operational rather than rhetorical.
Clear records help affected people obtain corrections and help businesses investigate failures.
03
WHERE IT COULD HELP
- Map every AI system to the claims, prices, recommendations, rankings, refusals and communications it can affect.
- Create an outcome trace for each consequential use before launch.
- Name the affected customer or market, data used, AI role, final action and accountable owner.
- Link every public claim to evidence, a review date and an authorized publisher.
- Recheck evidence after AI rewrites or reformats a claim.
- Expire claims automatically when their supporting evidence becomes stale.
- Give customers a human review route that can inspect records, correct data and reverse outcomes.
- Record original outputs, complaints, human decisions, remedies and resulting system changes.
- Document personalized-pricing inputs and separate cost factors from willingness-to-pay proxies.
- Test prices and offers across relevant customer groups and vulnerable circumstances.
- Verify reference prices, discount claims, urgency messages and AI disclosures.
- Review vendor data sources, competitor information, data separation, objectives and safeguards.
- Assign a competition owner for automated pricing, rankings, targeting and market decisions.
- Compare market outcomes with a pre-deployment baseline and define investigation thresholds.
- Give reviewers evidence, time, authority and a protected way to reject recommendations.
- Analyze overrides and disagreement instead of treating them as reviewer noise.
- Define a stop rule and preserve the logs needed for investigation.
- Keep the outcome record concise enough for routine use by small businesses.
- Train staff on both consumer and competition duties attached to their AI use.
- Seek qualified legal advice when the business or competition risk is significant.
KEEP A HAND ON THE WHEEL
The guidance and examples come from New Zealand’s Commerce Commission. This article is general editorial analysis and is not legal advice. The Commission is explaining its current view of the Fair Trading Act and Commerce Act. The guidance is not a new AI-specific law, a court judgment or a guarantee that any particular conduct will be investigated or found unlawful. The Fair Trading Act and Commerce Act pages state that the guidance should not replace legal advice. Whether a disclosure, pricing practice, recommendation, agreement or market outcome complies with the law depends on the specific facts. An outcome trace does not itself establish compliance. It is an operating record that can improve review, monitoring, correction and accountability. The regulator does not certify particular AI systems or supervise AI development generally. Its focus is conduct affecting consumers, competition and markets. Human approval is not necessarily meaningful oversight. Reviewers need relevant evidence, time, skill, authority and an alternative to accepting the output. Personalized pricing can have legitimate uses. Risk depends on the data, representation, context, customer effect and surrounding conduct. Algorithmic price similarity does not by itself prove illegal coordination. Competition analysis requires evidence about conduct, design, information sharing, market effects and applicable law. Vendor documentation and model evaluations do not replace testing on the buyer’s real use, data and market. Watch for enforcement activity, court interpretation, sector-specific guidance, changes to penalties, new regulatory tools and evidence from businesses implementing the guidance.
04
TERMS WORTH KEEPING
SOURCES AND VERIFICATION STATUS
This article was written from the materials below. Product claims and dates were checked against those sources on November 18, 2026.
THE PUBLICATION ENGINE
WANT A SIGNAL OF YOUR OWN?
We build source-grounded publications, private briefings, and editorial systems for organizations with something useful to say.
WORK WITH US