Unpacking Strategic Partnerships in AI: Competition Law Risks for Australian Businesses

A procurement team approves a joint AI development deal with a key supplier, sharing anonymised spend data to train a forecasting model. The contract looks standard — mutual NDAs, clear IP clauses — but the data pooling clause effectively locks both parties into using each other’s outputs, making independent bidding on future tenders commercially unviable.

This isn’t a cartel meeting in a back room. It’s a routine innovation partnership, structured to accelerate product cycles, yet it creates a structural impediment to competition that the ACCC can challenge under section 45 of the Competition and Consumer Act 2010 as a contract, arrangement or understanding that substantially lessens competition.

The pace of AI integration in procurement and product development is moving faster than the compliance controls designed to govern it. Teams are under pressure to innovate, to share data, to co-develop — yet each of these necessary steps can quietly shift a collaborative venture into territory scrutinised under the Competition and Consumer Act 2010. The risk isn’t always in secret meetings or fixed prices; it’s embedded in the architecture of the partnership itself, where efficiency gains come bundled with unintended coordination effects.

Old partnership risks lived in the contract — ambiguous SLAs, unclear IP ownership, payment delays. AI partnership risks live in the system: in how data flows are structured, how models are trained on shared inputs, and how outputs constrain future commercial freedom. What looks like innovation can function as a de facto exclusionary mechanism, triggering ACCC attention not because of bad intent, but because the effect on competition is substantial and real.

Navigating the ACCC Lens: Beyond Price Fixing

The ACCC no longer waits for the smoking gun of a cartel agreement. Its focus has shifted to the structural effects of strategic alliances, particularly where data sharing or joint development creates barriers to market access that mimic anti-competitive outcomes. An alliance need not contain a price-fixing clause to fall under scrutiny; the mere design of the collaboration can raise concerns under section 45 of the Competition and Consumer Act 2010 if it substantially lessens competition.

Modern partnerships increasingly involve pooled datasets, co-trained AI models, or jointly developed platforms that lock in participants and exclude rivals. The ACCC assesses these arrangements not by the labels parties attach to them, but by their actual effect on market dynamics — treating functional coordination as equivalent to explicit collusion when it reshapes competitive conditions. This evolution means compliance teams must evaluate alliances for their competitive impact, not just their contractual form.

The ACCC’s enforcement approach now treats information exchange as a primary risk vector, especially when shared data enables tacit coordination on output levels, customer allocation, or pricing algorithms. For instance, a joint venture between two competitors to pool supply chain logistics data might not set prices directly, but if the shared dataset allows each party to anticipate the other’s inventory movements and adjust their own supply accordingly, the effect can mirror market allocation. The ACCC has signalled that such outcomes, even absent communication about competitors’ intentions, may still constitute a contravention of section 45 if they substantially lessen competition.

This shift demands that procurement and legal teams assess alliances not only for what is agreed upon, but for what the collaboration makes possible — particularly where AI-driven analytics turn shared inputs into predictive insights that reduce uncertainty about rivals’ behaviour. The focus is on whether the arrangement creates a structural impediment that hinders new entrants or limits existing competitors’ ability to compete independently, regardless of whether the parties subjectively intended to restrict competition.

The ACCC’s guidance makes clear that the substance of an agreement, not its label, determines risk. A memorandum of understanding framed as “innovation collaboration” or a joint research project can still fall within the ambit of section 45 if it facilitates coordination that lessens competition. Regulators now examine whether the alliance creates a framework — through shared platforms, pooled datasets, or jointly developed algorithms — that enables participants to act in concert without explicit negotiation.

This evolution reflects a broader enforcement priority: targeting structural features of markets that sustain coordinated outcomes. Where once the ACCC pursued hard-core cartels with overt price-fixing agreements, it now investigates whether partnerships, particularly those involving AI-driven analytics, create conditions where independent decision-making is eroded. The test is objective — does the arrangement, in practice, diminish competitive pressure — not whether the parties admitted to a conspiracy.

Agentic AI: Where Collaboration Becomes Coordination Risk

When an AI agent autonomously negotiates pricing terms with a partner’s system, adjusting bids in real-time based on shared market data, the line between efficiency and collusion blurs. Unlike static algorithms that follow pre-set rules, agentic AI can learn, adapt, and coordinate outcomes without human intervention — raising concerns under section 45 of the Competition and Consumer Act 2010, which catches agreements, arrangements or understandings that substantially lessen competition, regardless of form.

The risk intensifies when these agents operate within a partnership that pools sensitive commercial information — such as inventory levels or customer willingness to pay — creating a feedback loop where independent decision-making is effectively outsourced. Even without explicit communication between firms, the synchronized behavior of their AI agents can mimic the effect of a cartel, triggering ACCC scrutiny not for what was agreed, but for what the system enables.

The danger extends beyond pricing. Consider an agentic AI managing inventory replenishment across a retail-supplier partnership. By autonomously adjusting order quantities based on pooled sales forecasts and warehouse data, it can inadvertently create coordinated market exclusion — say, by consistently deprioritising deliveries to a rival retailer not in the data-sharing pool. This effect mirrors concerns under section 46, which prohibits taking advantage of substantial market power to lessen competition, even if the outcome stems from autonomous system behaviour rather than a formal agreement.

The ACCC’s focus has shifted to whether the partnership structure, enabled by AI, creates or amplifies market power that facilitates coordinated effects. Regulators now assess whether the technical design of the agentic system — its access to shared data, its decision-making autonomy, and its interaction with partner systems — creates a practical ability to coordinate, regardless of human intent. This shifts liability from proving an agreement to demonstrating that the partnership, as operated, functions as a coordination mechanism.

An increasingly common scenario involves agentic AI systems embedded in logistics partnerships that dynamically reroute shipments based on real-time traffic, fuel prices, and partner warehouse capacity. When multiple competitors independently adopt the same third-party optimisation platform — each feeding their operational data into a shared algorithm — the system may begin to favour certain geographic corridors or delivery windows that collectively advantage a subset of market participants. This emergent behaviour, arising not from explicit coordination but from parallel reliance on a common decision-making engine, can replicate the market effects of a hub-and-spoke conspiracy under section 45, where the AI platform acts as the conduit facilitating parallel conduct.

The ACCC has signalled interest in such cases through its 2023 Digital Platforms Inquiry follow-up, noting that liability may attach not only to the firms using the AI but potentially to the provider if the system’s design enables or encourages anti-competitive outcomes. Crucially, the assessment hinges on whether the partnership structure — defined by data flows, algorithmic parameters, and system interoperability — creates a sustainable capacity to influence market conditions in a way that lessens competition, irrespective of whether any human executive intended that result.

Agentic AI introduces a layer of autonomy that blurs the line between independent action and coordinated conduct, making traditional market power assessments inadequate. When an AI agent negotiates terms, allocates resources, or excludes counterparties based on learned objectives — without real-time human oversight — its behaviour may constitute a tacit understanding under section 45, even if no communication occurred between rival firms. The risk intensifies if the agent’s optimisation logic converges on outcomes that replicate cartel-like effects, such as price stability or output restriction, across a network of participants using the same system.

Under section 46, the concern shifts to whether a dominant firm’s deployment of agentic AI constitutes a misuse of market power by leveraging algorithmic superiority to exclude competitors or entrench position. For example, an AI-driven procurement platform that autonomously prioritises suppliers meeting undisclosed efficiency thresholds could foreclose market access for smaller players, not through overt refusal to deal, but by rendering their offers non-competitive in real time. The ACCC’s focus is increasingly on whether the AI’s design or training data creates a structural advantage that lessens competition, regardless of intent.

Deconstructing Digital Dark Patterns in AI Contracts

Mandatory data pooling clauses, increasingly embedded in AI partnership agreements, function as dark patterns by design. They compel participants to feed operational, pricing, or customer data into a shared model under the guise of optimisation, yet the aggregated output often enables the dominant partner to infer competitors' cost structures or capacity constraints. This isn't mere information sharing; it creates a feedback loop where the AI's recommendations systematically advantage the data aggregator, mirroring the effect of a hub-and-spoke cartel without explicit communication.

Similarly, non-compete restrictions enforced through AI-driven contract execution — such as algorithms that automatically deprioritise or penalise suppliers engaging with rivals — act as behavioural nudges with coercive effect. These mechanisms exploit information asymmetry and processing speed to lock in compliance, undermining the voluntary nature of contractual terms. While framed as efficiency tools, they raise concerns under Schedule 2 of the Competition and Consumer Act 2010, particularly regarding misleading or deceptive conduct in business dealings, as the true competitive restraint is obscured by technical jargon and automated enforcement.

Another prevalent dark pattern involves AI-driven price recommendation engines embedded within joint venture platforms. These systems analyse pooled transaction data to suggest optimal pricing tiers, presenting the output as neutral market intelligence. However, when the dominant partner controls both the data inputs and the algorithm's weighting factors, the recommendations consistently nudge smaller participants toward price levels that maximise the aggregator's margin while suppressing genuine competition. This creates a de facto price leadership structure where adherence to the AI's guidance becomes commercially necessary, yet the mechanism avoids the explicit communication traditionally required to establish a cartel under section 45 of the Competition and Consumer Act 2010.

The consumer protection angle emerges when these pricing algorithms, optimised for platform profitability, are passed through to end consumers as supposedly competitive offers. If the AI's output systematically inflates prices above what a competitive market would bear—while being marketed as the result of independent, efficient pricing—this raises concerns under Schedule 2 regarding misleading or deceptive conduct. The technical complexity obscures the anti-competitive effect, making it difficult for counterparties, let alone consumers, to discern that the apparent market outcome is the product of coordinated design rather than organic market forces.

Mandatory data pooling clauses, often buried deep in partnership agreements, require participants to feed operational, pricing, or customer data into a central repository controlled by the dominant partner or a jointly governed entity. While framed as enabling better AI training or supply chain optimisation, these clauses can create structural dependencies that prevent smaller partners from operating independently or developing competing capabilities, effectively locking them into the aggregator's ecosystem.

Non-compete restrictions enforced through AI model access—where continued use of the algorithmic tool is contingent upon agreeing not to develop or deploy similar capabilities externally—function as operational dark patterns. They subtly constrain market participation by making the cost of non-compliance prohibitively high, leveraging technical integration as a compliance mechanism rather than relying solely on contractual terms that might attract closer scrutiny.

The consequence extends beyond competition law concerns into consumer protection territory under Schedule 2 of the Act. When these structural impediments suppress genuine market rivalry, the resulting prices or service terms presented to end-users lack the competitive discipline that would otherwise prevail. Marketing such outcomes as the product of efficient, independent market processes—when they are instead shaped by coordinated design embedded in contractual and technical architecture—risks constituting misleading or deceptive conduct, as the true nature of the market outcome is obscured from both business customers and consumers.

Uncommon Insights

The ACCC’s recent enforcement priorities reveal a telling misalignment: businesses obsess over contractual clauses while regulators zero in on how systems actually behave in live operation. A partnership agreement may look pristine on paper—no explicit price-fixing, no territorial carve-outs—but if the embedded AI routinely steers volume toward a preferred supplier through opaque scoring logic, that’s where the risk crystallises. Regulators care less about what you promised and more about what the algorithm consistently does when no one is watching.

This shift demands moving beyond pre-signature legal review to continuous operational governance oversight. Traditional due diligence assesses intent; what matters now is detecting effect—whether technical architecture, data flows, or incentive structures produce outcomes that substantially lessen competition, regardless of contractual wording. The real compliance gap isn’t in the signature block; it’s in the feedback loop between deployed code and market behaviour.

Consider a joint venture between two logistics firms that shares real-time freight pricing data through an AI optimisation engine. The contract expressly forbids collusion and mandates competitive bidding for each route. Yet the AI, trained to maximise network-wide efficiency, begins to consistently assign high-volume corridors to one partner while routing low-margin freight to the other—a pattern that emerges not from explicit instruction but from the algorithm’s implicit learning of each partner’s capacity constraints and cost structures. Over time, this creates a de facto market partition that the ACCC could challenge under s 45 as a contract, arrangement or understanding that substantially lessens competition, even though no human ever agreed to it.

The regulator’s focus, therefore, shifts from scrutinising the joint venture agreement to auditing the AI’s decision logs, feedback weights and outcome distributions. What looks like operational efficiency on the surface may mask a coordination effect that alters market incentives. This is why the ACCC’s recent guidance on algorithmic harm stresses testing for “behavioural equivalence”—whether the system’s output mirrors what a human cartel would have produced, regardless of intent or contractual safeguards.

The real regulatory blind spot isn't the contract—it's the feedback loop between the AI and the partners' operational data. When an agentic system continuously ingests pricing, capacity, and cost inputs from each party, then optimises for joint efficiency, it can inadvertently reinforce behavioural patterns that mimic market allocation. The ACCC isn't looking for a smoking-gun email; it's seeking evidence of sustained, reciprocal data exchange that creates a self-reinforcing coordination effect, effectively outsourcing collusion to the algorithm's loss function.

This shifts the compliance burden from pre-signature legal review to ongoing operational governance. Regulators now expect access to model training data, feature importance scores, and outcome parity tests—not just contractual clauses. A partnership that appears arm's-length on paper may still trigger s 45 scrutiny if the AI's decision architecture systematically advantages certain partners in ways that reduce competitive pressure, even if no human ever discussed market shares.

Regulators prioritise behavioural outcomes over contractual labels. An agreement may prohibit price coordination, yet if the AI continuously optimises joint output based on shared cost inputs, the effect mirrors collusion. The ACCC examines whether the system creates a structural advantage that lessens competition—not whether humans intended it.

Operational governance now outweighs pre-signature review. Key metrics include data lineage transparency, model drift monitoring, and outcome parity across competitor segments. A stats row illustrates this shift:

Key Takeaways

Pre-emptive due diligence must extend beyond financial health to include algorithmic transparency. CFOs should require suppliers to disclose data inputs, model versioning, and governance protocols before any partnership scales. This isn't about distrust—it's about ensuring the collaboration cannot be recharacterised as coordinated conduct under s 45 of the Competition and Consumer Act 2010.

Agreements need operational audit trails, not just legal clauses. Structure contracts to mandate real-time monitoring of model outputs for adverse competitive effects, with clear escalation paths if parity deviates across market segments. Transparency here is a control, not a concession.

Embedding supplier intelligence platforms into the vetting workflow transforms due diligence from a point-in-time check into a continuous risk filter. These platforms aggregate regulatory filings, adverse media, and ownership structures, flagging entities with prior competition law engagements or complex offshore layers that obscure control—common precursors to s 46 concerns about market power acquisition through stealth.

When assessing AI-driven partnerships, layer this intelligence with technical due diligence: request model cards detailing training data provenance and fine-tuning logs. If a supplier’s pricing algorithm inadvertently mirrors competitors’ due to shared third-party data feeds—a known risk in logistics and retail tech—the platform’s historical alert on similar sector collusion patterns triggers enhanced scrutiny, turning passive data into an active compliance control.

Map the data lineage of shared AI models before signing. If a joint development agreement pools procurement spend data to train a forecasting tool, trace whether that data could reveal individual bid strategies to competitors via model inversion—a technical pathway to s 45 concerns about information exchange facilitating coordination. Require suppliers to provide differential privacy guarantees or federated learning architectures as contractual safeguards, not just anonymisation promises.

Test exit mechanics upfront. Agreements often lock in AI model usage through irrevocable licences or embedded dependencies, creating de facto exclusivity that may attract s 46 scrutiny if it forecloses market access. Negotiate clauses allowing data retrieval and model retraining on neutral infrastructure, verified by third-party audit, to ensure the partnership remains contestable and does not become a structural barrier to entry.

• Map data flows before model training. Trace whether pooled procurement or pricing data could enable competitors to infer bid strategies through model inversion, triggering s 45 information exchange concerns.
• Build exit rights into AI licences. Negotiate verifiable data retrieval and model retraining clauses to prevent de facto exclusivity that may attract s 46 scrutiny for foreclosing market access.
• Verify supplier tech stack claims. Use intelligence platforms to confirm differential privacy or federated learning architectures are operational, not just contractual promises.