When pricing bots bid on public contracts
Pricing algorithms used to bid on public contracts can keep prices high without communicating or punishing rivals, which is the evidence competition law looks for. Public buyers can protect public money with stress tests of bidders' pricing systems, published bid records and audits of shared AI models.
Competition watchdogs look for two signs of bid-rigging: secret communication and punishment of a rival who undercuts. Recent research shows that pricing algorithms can sustain high prices with neither (see the research notes below). That leaves public procurement, which spends public money on road salt, school buses and paving, exposed in a way current enforcement isn't built to catch.
The myth of the necessary punishment
Traditional antitrust law assumed that keeping prices artificially high requires a visible stick. If two companies agree to inflate bids, each firm feels an urge to cheat. Dropping the bid by two percent steals the entire public contract.
To prevent cheating, human cartels rely on threats. If a rival breaks rank, the cartel floods the market and undercuts them into insolvency. Economists called this a punishment phase. Without a credible threat of retaliation, price coordination collapses.
This assumption shaped every audit tool in public procurement. If watchdogs saw no cross-vendor emails and no retaliatory price drops, they concluded that vendors were competing independently. High bids were dismissed as ordinary supply chain friction or inflation.
Recent empirical studies have overturned that belief. When modern pricing algorithms learn in simulated auctions, they settle on supracompetitive prices without establishing retaliatory threats.1 Instead of policing each other with punishment phases, the algorithms converge onto dynamical attractors where high bids become stable resting states.2
The algorithms never threaten each other. They simply find mutual resting points where undercutting never appears profitable. When public officials search only for angry retaliation, unpunished algorithmic coordination passes right through unnoticed.
How shared model roots align bids
When multiple competing vendors buy software from third parties, another form of tacit coordination takes hold. Contractors rarely build complex pricing engines from scratch. Instead, three asphalt firms might all subscribe to the same cloud bidding service or fine-tune models from the same foundation provider.
Competition agencies have warned that using a single pricing intermediary violates antitrust rules even if the competitors never communicate directly.3 When competing vendors delegate pricing to a shared intermediary, the software pools market data and optimizes collective returns across all clients. The Federal Trade Commission affirmed that price-fixing by algorithm remains illegal price-fixing.3
Even without an explicit intermediary platform, shared foundation models create correlated bidding behaviour. If two distinct pricing bots rely on models trained on identical industry data, they share similar inductive biases. They react to changes in fuel costs or labor supply in parallel lockstep.
To an outside observer, the two bids look like independent responses to market pressures. In reality, the algorithms share an invisible intellectual heritage that dampens price competition. Competition watchdogs have begun advising public buyers to map shared software dependencies across bidding pools.4
Three concrete safeguards for public tenders
Public procurement directors do not need to rewrite contract law to protect public tenders. They can add three practical safeguards to their bidding guidelines today.
First, mandate counterfactual market testing for automated bidding software. Before a vendor deploys an automated pricing system on public tenders, procurement authorities should require simulated stress tests. Regulators introduce synthetic demand and cost shocks to observe how the software responds. If an algorithm ignores cost reductions and maintains elevated prices, the software displays uncompetitive coordination.
Second, require verified bidding receipts. Every automated bid should carry a signed, tamper-evident log documenting the input data and prompt configuration that produced the price. If a vendor claims that high bids reflect sudden freight spikes, public buyers can review the verified execution trace. Tamper-evident receipts eliminate post-hoc excuses and make automated bidding auditable.
Third, audit foundation model dependencies across bidder pools. Standard tender questionnaires should ask vendors to disclose whether their pricing software relies on common third-party intermediaries, shared data feeds, or shared proprietary fine-tunes. If three major bidders in a municipal tender use the same commercial pricing engine, the procurement officer can split the tender or require independent manual pricing.
Keeping public tenders truly public
Sealed public bids were invented to ensure that tax dollars buy quality work at honest prices. When automated systems turn that process into a silent, coordinated dance, the public pays the difference in every repaved road and every school bus route.
Protecting the public interest does not require banning software from commerce. It requires updating our inspection tools. Just as bridge inspectors test steel under load rather than trusting a mill certificate, public buyers must test algorithms under pressure.
By coupling counterfactual testing with transparent execution receipts, municipal governments can keep automated bidding accountable. When the next set of tender envelopes lands on the council table, public officials can open them knowing that competition remains real.
To inspect verified procurement action cards and tested audit protocols, visit our Shared Horizon action library or explore our briefings for policy makers.
Related essays: Five questions to ask before buying an AI agent · Why pricing bots learn not to compete. Further reading: When the middleman pads the meter · AI agents should hold permissions that expire · What governments are already doing about AI agents.
1. Eshwar Ram Arunachaleswaran, Natalie Collina, Sampath Kannan et al., Algorithmic Collusion Without Threats, arXiv:2409.03956 (September 2024).
2. Kaue Lopes de Moraes and Rodrigo De-Losso, Algorithmic Collusion Without Deterrence: Auditing the Punishment Mechanism in Q-Learning Pricing Agents, arXiv:2610.21867 (October 2026).
3. Federal Trade Commission, Price Fixing by Algorithm is Still Price Fixing, FTC Business Guidance (March 2024).
4. Competition Bureau Canada, Algorithmic pricing and competition in public procurement, Government of Canada (October 2024).
