When Algorithms Collude but Firms Do Not: The Dilemma under Section 2(b) of Indian Competition Act, 2002

In January 2025, a woman in Bengaluru posted a video that unsettled India’s competition law community. She had placed the same groceries on online order in her cart on Zepto, once on her iPhone, and once on her partner’s android. The prices were different, not by a rounding error but that groceries cost measurably more on the Apple device. Nobody at Zepto had called a competitor,  nor had anybody agreed to anything in concert, but it was an algorithm that had simply learned that iPhone users are statistically wealthier, less price sensitive and could be charged more.

Around the same time, the Commission had separately begun examining unfair pricing and discounting in the quick-commerce sector. Speaking at the 10^th National conference on Economics of Competition Law in March 2025, CCI Chairperson Ravneet Kaur cautioned that algorithms can facilitate “cartels without human communication” and “price coordination without explicit agreement.” While this captures the emerging challenge, the CCI’s own landmark Market Study on AI and Competition (October 2025) acknowledges a deeper unresolved issue that India’s Competition Act, 2002, may be structurally ill-equipped to address autonomous algorithmic collusion under its existing definition of “agreement”.

Section 2(b) of the Act, 2002 Trap

Definition of “Aagreement” under Section 2(b) of the Competition Act, 2002 is expansive and includes any arrangement or understanding or “action in concert”, whether or not formal or in writing, and whether or not intended to be enforceable by legal proceedings. The CCI and the Supreme Court have interpreted this broadly. In Excel Crop Care v. CCI (2017), the Apex Court affirmed that circumstantial evidence such as price increase and identical market behaviour could establish the Section 3 violation.

The phrase “action in concert” still requires, at its irreducible minimum, some concert, some coordination between entities capable of forming an understanding.

Even the most liberal purposive interpretation of Section 2(b) has never been tested against the proposition that a machine learning system, operating without any human instruction to coordinate, could itself constitute or facilitate an “action in concert.”

The CCI confronted this question in Samir Agrawal v. ANI Technologies Pvt. Ltd, 2018, that Ola and Uber drivers, though independent operators and competitors, all used the platform's algorithm to set fares. Was that a hub-and-spoke cartel? The CCI said no. A hub-and-spoke arrangement requires an agreement between the spokes, the competing drivers, to delegate pricing to the hub. Since no such agreement existed between the drivers themselves, the ingredients of the Section 3(3) violation were absent. The Supreme Court upheld this finding in December 2020.

The CCI’s reasoning was doctrinally defensible, which licensed the very problem it refused to engage with. What the Commission left unanswered is the harder issue underneath: What happens when multiple competing platforms, not their users, each independently deploy self-learning pricing algorithms that, without any communication between them, converge on coordinated pricing outcomes?

The Parallel Machines Problem

The existing Indian literature on algorithmic collusion has done valuable work mapping the Ezrachi- Stucke taxonomy onto Indian doctrine. But the gap lies at the intersection of two India-specific features that have not been analysed together: the concentrated structure of the quick-commerce market, and the evidentiary standard required to establish coordination under Section 3(3).

The first is the structure of India's quick commerce market. Blinkit, Instamart, and Zepto together held roughly 88% of the sector in March 2025 and individual shares variously reported around 35%, 31% and 22%. They operate in the same cities, serve overlapping consumer bases, and price identical SKUs against each other in real time. Later brokerage reports put Blinkit at roughly 46% and Zepto, Instamart at about 29% and 25%, respectively, though these figures vary with time and by sources.

Each has a sophisticated AI pricing engine trained on publicly available competitor pricing, demand patterns, and consumer behaviour with powerful incentives to converge on the same strategy. Independent tracking of grocery prices across Blinkit, Zepto and Instamart in an analysis from December 2024 to January 2025 recorded exactly same pattern as introduced earlier: prices for same items varied by devices, user login state, time of week, and frequently of app used.

The second feature is that none of this requires any communication between Blinkit, Instamart and Zepto. The algorithms do not communicate with each other. There is no shared software vendor functioning as a hub. There is no exchange of non-public, competitively sensitive information of the kind the 2023 (Amendment) has explicitly recognised in hub-and-spoke cartels. Rather, coordinated behaviour is simply the equilibrium maximising strategy ins an oligopolistic market with high price transparency and instant observability.

In effect, this form of algorithmic tacit collusion remains beyond the reach of currently drafted Ssection 2(b), which assumes that concert necessarily requires consent.

The loophole the 2023 Amendment left open 

The 2023 Amendment was the most significant revision to India’s competition framework in over a decade. It codified hub-and-spoke cartel liability under the amended Section 3(3) by adding a proviso, where there is presumption of appreciable adverse effect on competition where a facilitator that is not itself a competitor in relevant market is treated to be part of an anti-competitive agreement if it merely participates, or intends to participate, in furthering an agreement already reached between rival firms.  

But the amendment’s hub-and-spoke provision still requires identifiable spokes that are aware of the arrangement and derive benefit from or contribute to it. It captures a scenario where Blinkit and Zepto both subscribe to a third-party pricing platform that aggregates each other’s non-public data. It doesn’t capture two separate AI systems that have never shared a data point, have never been configured by the same vendor, and have never received the same instruction but that have nonetheless, through iterative reinforcement learning on public market signals, converged on pricing strategies indistinguishable in their effects from a cartel. This is the Amendment’s blind spot.

The CCI’s own AI Market Study for 2025 acknowledges that self-learning algorithms can independently engage in collusive behaviour by observing which pricing actions yield higher profits over time. As the CCI itself states, “algorithmic coordinated conduct absent human intervention” is a key concern yet the Study recommends only “collusion stress-testing” and internal AI governance by firms.

Proposed Insufficient Solutions

There are two proposed responses: either stretch Section 2(b) through purposive interpretation, or wait for legislative reform. Both seems to be inadequate.

Purposive interpretation of Section 2 (b) without any judicial decision creates a rule impossible to apply consistently. If the Supreme Court holds that self-learning algorithm convergence, without human communication, amounts to “action in concert”, it will be drawing an inference for which the statutory text provides no guidance. Waiting for legislative reform ignores that the harm is happening now.

There is a third path, grounded in existing Indian doctrine. Section 3(1) prohibits any agreement that “causes or is likely to cause an appreciable adverse effect on competition.”, read alongside Section 2(b)’s “action in concert” language; there is a credible doctrinal argument for what might be called a “foreseeable coordination” standard on the basis of circumstantial evidence, grounded in Supreme Court’s decision in Rajasthan Cylinders..

The argument is that a firm that deploys an AI pricing system in a concentrated market optimised for revenue maximisation, capable of observing and responding to competitor pricing in real time without any safeguard against supra-competitive equilibria, has made a series of design choices that forcibly produce coordination. This is not a negligence standard; rather, it is constructive intent of a firm’s deliberate choice to build and retain a system it can reasonably foresee will converge with rival’s pricing, satisfying Section 2(b)’s “action in concert”, imputed through design rather than express communication. Liability attached is not because the algorithm agreed with its rival, but because its designer created a system whose expected output in the relevant market conditions was functionally indistinguishable from agreement.

In Rajasthan Cylinders, the Supreme Court cautioned against price-eating parallel pricing as evidence of collusion but did not foreclose circumstantial inference where the only rational explanation for the pricing pattern was coordination. An AI system that consistently produces iOS price premium across Bengaluru, Mumbai, and Delhi simultaneously, mirroring its two main competitors, may satisfy that threshold, not because of the algorithm, but because the design made coordination the inevitable outcome.

CCI’s Institutional Challenge

There is a further dimension that the 2025 Market Study points at without fully confronting:, is the CCI’s ability to investigate these cases at all. Proving traditional cartel coordination requires emails, meeting records, and testimony. Proving algorithm coordination requires reverse- engineering, a machine learning model to show that it’s training parameters and objective function foreseeably produced collusive outputs.  The Director General’s Office currently has no published capability to do this. India’s Competition Regulator has begun building AI-related capacity, but its public materials do not show a specialised technical unit on the scale of the UK Competition and Markets Authority whose DaTA function was built for algorithmic cases, or a parallel embedded in-house team of data scientists comparable to the CMA’s.

As the AI market study observes, the commission has adopted a “light touch” approach centred on “self-audit and proactive compliance”. Audit reports cannot replace independent forensic scrutiny. A “foreseeable coordination” standard is only as strong as the evidence available to reconstruct it after the fact. Since design-stage intent is really documented, regulators must infer foreseeability from market outcomes, the absence of competing design, safeguards, and expert analysis of model’s objective. This evidentiary burden may limit the standard’s effectiveness, to cases with observable outcomes.

Accordingly, independent algorithmic design audits should form a distinct category of evidence. The DG’s powers under Section 41 already provide the statutory basis, what remains is the technical expertise to deploy them effectively.

Conclusion

India’s competition framework has recognised the challenge of algorithmic collusion but has yet to develop a doctrine and framework capable of addressing autonomous algorithmic coordination under Section 2(b).

Filling this gap does not require waiting for the Digital Competition Bill. It requires the CCI to do what it has always done: the statute broadly, reason carefully from first principles, and let market reality drive doctrine forward.

---

Authored by Shivam Singh and Priyanshi Jain, Final-Year students at Dharamshastra National Law University, Jabalpur.