Algorithmic Pricing and Antitrust: Shared Software, Tacit Collusion, and the New Price-Fixing Cases

Document type: Article Practice area: Antitrust — Price Fixing and Cartels Jurisdiction: United States Last reviewed: 5 September 2026

The problem in one paragraph

Ten firms in a market each subscribe to the same pricing software. Each supplies the vendor with its own transaction data — prices, occupancy, volumes, competitor observations. The vendor pools that data, runs a model, and returns to each subscriber a recommended price. The subscribers adopt the recommendation most of the time. Prices in the market rise.

Nobody in that story met in a hotel room. Nobody exchanged a spreadsheet with a competitor. Each firm made its own decision, using information it paid for, from a vendor it selected. And yet the market looks, from the outside, exactly as it would look if the ten firms had agreed to raise prices.

Is that an agreement under Section 1 of the Sherman Act? The answer courts are converging on is: it depends on facts that have nothing to do with the sophistication of the algorithm, and everything to do with whether the participants understood themselves to be participating in a common scheme.

Section 1 requires an agreement, and always has

Section 1 reaches "every contract, combination in the form of trust or otherwise, or conspiracy, in restraint of trade." The operative word has always been the requirement of concerted action. Unilateral conduct, however anticompetitive in effect, is not reached by Section 1 — it is reached, if at all, by Section 2 or by Section 5 of the FTC Act.

Theatre Enterprises, Inc. v. Paramount Film Distributing Corp., 346 U.S. 537 (1954), states the baseline: "business behavior is admissible circumstantial evidence from which the fact finder may infer agreement," but "this Court has never held that proof of parallel business behavior conclusively establishes agreement." Conscious parallelism — firms independently recognizing their interdependence and pricing accordingly — is not a violation. It is what oligopolists do.

Monsanto Co. v. Spray-Rite Service Corp., 465 U.S. 752 (1984), supplies the evidentiary standard: a plaintiff must present "evidence that tends to exclude the possibility that the alleged conspirators acted independently." Matsushita Electric Industrial Co. v. Zenith Radio Corp., 475 U.S. 574 (1986), added that where the alleged conspiracy makes no economic sense, the inference of agreement is weaker still.

And Bell Atlantic Corp. v. Twombly, 550 U.S. 544 (2007), moved the analysis to the pleadings: a complaint alleging parallel conduct must include "enough factual matter (taken as true) to suggest that an agreement was made." Allegations "merely consistent with" agreement do not survive.

Algorithmic pricing cases live inside this framework. The question is not whether an algorithm can conspire. It is whether the humans who bought and deployed it entered into an agreement — express or tacit, direct or through a hub.

The hub-and-spoke structure

The theory most plaintiffs plead is hub-and-spoke: the vendor is the hub, each subscriber a spoke, and the rim is the agreement among the spokes.

The doctrinal ancestor is Interstate Circuit, Inc. v. United States, 306 U.S. 208 (1939). A theatre chain sent a single letter to eight film distributors, proposing minimum admission prices and restrictions on double features, and — critically — the letter named all eight addressees. Each distributor complied. The Court found a conspiracy without any evidence of communication among the distributors:

"It was enough that, knowing that concerted action was contemplated and invited, the distributors gave their adherence to the scheme and participated in it."

That sentence is the centre of gravity in modern algorithmic pricing litigation. It establishes that a common scheme communicated through a hub, with knowledge that others are being invited to the same scheme, can constitute agreement without spoke-to-spoke contact.

American Tobacco Co. v. United States, 328 U.S. 781 (1946), reinforced the point that a formal agreement is unnecessary: "The essential combination or conspiracy may be found in a course of dealings or other circumstances as well as in an exchange of words."

But a hub-and-spoke claim requires the rim. Where plaintiffs plead a hub and spokes and no rim — no basis to find that the spokes knew of and adhered to a common scheme — courts dismiss. Vertical relationships between a vendor and each of its customers, standing alone, are just contracts.

The features that plaintiffs use to plead the rim in algorithmic cases:

The vendor markets the product as one that raises prices market-wide, and tells prospective subscribers that competitors are using it. This is the Interstate Circuit letter in modern form: an invitation that discloses to each recipient that others are being invited.
Subscribers pool non-public, competitively sensitive data into the model.
The vendor requires or strongly pressures adherence to the recommendation, monitors compliance, and follows up on deviation.
The vendor convenes subscribers at user groups, advisory boards, or conferences where pricing strategy is discussed.
Subscribers know that competitors use the same tool and expect them to follow its output.

The features that defendants use to negate it:

The vendor uses only public data, or the subscriber's own data.
Recommendations are advisory, and adherence rates are low or variable.
The model is individualized — different subscribers receive different recommendations reflecting their own cost, inventory, and demand.
No subscriber knows which competitors use the tool, or how many.
No forum brings subscribers together on pricing.

Per se or rule of reason

If there is an agreement, the next question is what standard applies, and it matters enormously. A per se horizontal price-fixing agreement is unlawful without inquiry into market power, effects, or justification. A rule of reason claim requires the plaintiff to define a market, prove market power, and show anticompetitive effects — and most of them fail.

United States v. Socony-Vacuum Oil Co., 310 U.S. 150 (1940), stated the per se rule broadly: "Under the Sherman Act a combination formed for the purpose and with the effect of raising, depressing, fixing, pegging, or stabilizing the price of a commodity in interstate or foreign commerce is illegal per se." Catalano, Inc. v. Target Sales, Inc., 446 U.S. 643 (1980), applied it to an agreement to eliminate short-term credit — a term of sale, not a price, but "an inherent part of the price."

Against that, information exchange has generally been analysed under the rule of reason. United States v. Container Corp. of America, 393 U.S. 333 (1969), is the closest Supreme Court authority. Container manufacturers exchanged, on request, the most recent price charged to a specific customer. The Court found a violation, but its reasoning was effects-based: the market was concentrated, the product fungible, demand inelastic, and the exchange had "the effect of keeping prices within a fairly narrow ambit." The Court was careful to say that "the exchange of price data tends toward price uniformity" in that setting, not that all exchanges are unlawful.

The distinction that emerges is between agreeing on price and agreeing to share information that affects price. The first is per se unlawful. The second is judged by its effects.

Algorithmic pricing cases test the boundary. Where the allegation is that competitors delegated pricing authority to a common agent, plaintiffs argue per se treatment: the algorithm is the agreement, and the fact that it operates through software rather than a smoke-filled room is irrelevant. Where the allegation is that competitors contributed data to a pool that produced recommendations each was free to accept or reject, defendants argue rule of reason: this is information exchange, and information exchange has never been per se unlawful.

Courts have split, and the split tends to track the facts. Where the complaint alleges that subscribers were required to accept recommendations, that the vendor policed compliance, and that subscribers knew competitors were bound the same way, courts have allowed per se claims to proceed. Where adherence was optional and the pool was thin, courts have applied the rule of reason or dismissed.

Conscious parallelism, plus factors, and the algorithm as a plus factor

Because parallel conduct alone is not enough, plaintiffs plead plus factors — circumstances that make independent action less plausible. The traditional list includes actions against independent self-interest, a motive to conspire, opportunities to conspire, unusual or abrupt price movements, high market concentration, product homogeneity, inelastic demand, barriers to entry, and evidence of inter-firm communications.

Algorithmic pricing supplies some new candidates:

Simultaneous and identical price movements, at a granularity that human coordination could not achieve.
Adoption of a recommendation that departs from the firm's own historical strategy — for example, holding inventory vacant rather than discounting, when the firm had always discounted.
Internal documents acknowledging that the tool works only if competitors also use it. This is the single most damaging category of evidence, and it exists in a surprising number of files.
Adherence rates disclosed by the vendor to subscribers — telling each subscriber that others follow the recommendation 90% of the time is an assurance of reciprocity.
Vendor-imposed limits on deviation, or escalation processes when a subscriber prices below the recommendation.

Note that several of these are about what the participants were told, not about what the software computed. The technology is not the violation. The mutual assurance is.

What the algorithm actually does matters

Not every pricing algorithm raises the same question. It is worth distinguishing four architectures, because their legal profiles differ sharply.

1. Own-data optimization. The firm's own historical sales, inventory, and costs feed a model that recommends prices. No competitor data, no pooling, no vendor aggregation. This is ordinary business analytics and raises no Section 1 issue at all.

2. Public-data monitoring. The model incorporates publicly posted competitor prices — scraped from websites, gathered from shelf surveys. Firms have always watched competitors' posted prices; doing it faster does not create an agreement. The risk here is not Section 1 but the speed of the resulting equilibrium, which is a policy concern rather than a doctrinal one.

3. Pooled non-public data. Subscribers contribute confidential transaction data — actual transaction prices, not list prices; occupancy or utilization; forward bookings. The vendor aggregates and returns recommendations. This is where the legal risk concentrates. The older information-exchange cases apply directly, and the safeguards developed for trade association data exchanges are the relevant compliance model.

4. Delegated pricing. Subscribers give the vendor authority to set prices, or adopt recommendations automatically without human review. This is the structure most vulnerable to a per se characterization, because the firms have in substance handed pricing to a common decision-maker.

The compliance advice follows the taxonomy. Architecture 1 needs nothing. Architecture 2 needs a scraping and data-provenance policy. Architecture 3 needs the full information-exchange safeguard set. Architecture 4 needs a very careful look and, often, a redesign.

The information-exchange safeguards

Antitrust agencies have long recognized that data sharing can be procompetitive — benchmarking, efficiency comparison, capacity planning — and have described conditions under which exchanges are unlikely to be challenged. Those conditions are the compliance template for pooled-data pricing tools:

Historical, not current or forward-looking. Data at least three months old is materially less useful for coordination.
Aggregated, not firm-specific. No participant should be able to identify another participant's data.
Sufficient participants. A commonly cited threshold is at least five participants, with no participant representing more than 25% of the aggregate on a weighted basis.
Managed by an independent third party that receives the raw data and publishes only the aggregate.
Available on the same terms to all industry participants, including non-participants who wish to purchase.
No accompanying discussion of pricing intentions, output plans, or strategy among participants.

A pricing tool built on current transaction data from a handful of large firms in a concentrated market, delivering firm-specific recommendations, meets none of these. A benchmarking product built on aggregated data six months old, sold to anyone who wants it, meets all of them.

The vertical-restraint overlay

Some algorithmic pricing arrangements are not horizontal at all. A manufacturer that supplies a pricing tool to its dealers, or a platform that recommends prices to sellers using it, is engaged in a vertical relationship, and vertical price restraints are analysed under the rule of reason after Leegin Creative Leather Products, Inc. v. PSKS, Inc., 551 U.S. 877 (2007).

Leegin overruled the century-old per se rule against resale price maintenance, holding that minimum RPM must be judged by its effects. The Court identified procompetitive justifications — encouraging retailer investment in services, facilitating entry by new brands, preventing free-riding — and anticompetitive ones, notably that RPM can facilitate a manufacturer cartel or a retailer cartel, and that it is more suspect where it originates with retailers rather than the manufacturer.

Two cautions for anyone relying on Leegin.

First, state law diverges. Several states treat minimum RPM as per se unlawful under their own antitrust statutes, and a national pricing programme cannot rely on federal law alone.

Second, a vertical arrangement can carry a horizontal agreement. If the dealers pressed the manufacturer to impose the pricing tool, or if the manufacturer's adoption was a response to dealer complaints about a discounter, the arrangement may be a dealer cartel operating through a vertical instrument. That is the Interstate Circuit problem again, wearing different clothes.

The platform variant raises related questions. A marketplace that provides sellers with pricing recommendations, or that penalizes sellers who price lower elsewhere, is imposing a vertical restraint. Analysis of such arrangements has become more complex since Ohio v. American Express Co., 585 U.S. 529 (2018), which held that for a two-sided transaction platform the relevant market includes both sides and that a plaintiff must show a net anticompetitive effect across the platform. Whether and how far that holding extends beyond credit card networks remains actively contested, and a defendant relying on it should not assume it travels.

Section 5 and the unilateral-conduct gap

Even where Section 1 fails for want of an agreement, the conduct may not be beyond reach. Section 5 of the FTC Act prohibits "unfair methods of competition," and the Commission has long asserted that it extends beyond the Sherman Act's boundaries — reaching invitations to collude that were never accepted, and facilitating practices that fall short of agreement.

The invitation-to-collude theory is directly relevant here. A vendor that markets a pricing product by telling each prospective subscriber that its competitors are already participating and that the product raises market prices has, on one view, issued an invitation to collude to every recipient. That the recipients never communicated with each other does not defeat a Section 5 theory the way it may defeat a Section 1 claim.

For a subscriber, this matters in a practical way: the vendor's marketing materials are discoverable, they are attributed to the vendor rather than the subscriber, and they will be characterized as the common scheme the subscriber joined. A subscriber that received such a deck, kept it, and signed the contract has a harder story than one that raised the issue with the vendor and obtained written confirmation that the product does not operate that way.

Learned models and the problem of intent

A recurring argument in this area is that machine learning models create liability without culpability: the model learns to price supra-competitively, nobody instructed it to, and the firms deploying it did not know.

The argument overstates the doctrinal problem in one direction and understates it in another.

It overstates it because Section 1 requires an agreement, and an agreement requires a conscious commitment to a common scheme. A firm that deploys a model, does not pool data with competitors, receives no assurance about competitor conduct, and simply prices to maximize its own return has not agreed to anything, however sophisticated the model. Independent parallel optimization is conscious parallelism, and Theatre Enterprises governs.

It understates the problem because in real cases the model is rarely a black box that nobody understands. The objective function was chosen. The training data was selected. The constraints were configured. When a firm configures a model to maximize revenue subject to a floor derived from competitor prices in a pooled dataset, the firm has made a series of deliberate choices, and those choices are documented in configuration files, tickets, and design decisions.

The practical implication is that model governance is antitrust compliance. A firm that can produce a written record of what its model optimizes, what data it uses, who approved the design, and what constraints were imposed is in a far better position than one that says the vendor built it and nobody looked. "We did not know what it did" is not a defence to a claim that the firm agreed with competitors, but it is a terrible answer to every other question in the case.

What "adoption rate" evidence actually shows

Plaintiffs make much of adherence statistics: subscribers accepted the recommendation 85% or 92% or 98% of the time, and that near-uniformity proves delegation of pricing authority.

Defendants respond that a good recommendation should be accepted, that a tool nobody follows is a tool nobody would buy, and that high acceptance shows the model is accurate rather than that the firms conspired.

Both are partly right, and the resolution turns on details that a compliance programme can influence:

Was acceptance measured and reported to anyone outside the firm? A firm's own internal metric is different from a figure the vendor collects and shares.
Was there a target? An internal or contractual acceptance-rate objective converts a recommendation into something closer to an instruction.
What happened on deviation? Nothing, or a call from the account manager?
Was there human review? A documented decision by a named person, with reasons, is the strongest available evidence of independent judgment — and it is cheap to create.
Did the firm deviate when its own circumstances warranted it? A record of the firm pricing below the recommendation to fill inventory, win a contract, or clear stock is powerful.

The compliance conclusion is not "accept fewer recommendations." It is "document why you accepted each one." A firm that follows the recommendation 90% of the time with contemporaneous reasons looks very different from a firm that follows it 90% of the time with no record at all.

Worked example one: the hotel operator

Ines Karlsson is general counsel of Wrenhaven Hospitality, which operates forty hotels in eleven markets. The revenue management team wants to adopt a vendor product that recommends nightly rates.

Ines runs a structured evaluation.

What data goes in? The vendor's sales materials say the model uses "market data." Ines asks specifically: does Wrenhaven's own transaction data leave the building, and is it combined with data from competing hotels? The answer is yes to both. The vendor collects daily rate and occupancy data from participating hotels and uses it to build market-level demand forecasts.

How current is the data? Same day. Forward bookings are included.

How many participants, and what share? In Wrenhaven's largest market, the vendor has six of the eleven comparable properties, representing roughly 70% of rooms.

Are recommendations firm-specific? Yes. Each property receives a recommended rate for each night.

Is adherence required? The contract does not require it. But the vendor's implementation methodology sets an "acceptance rate" target, the vendor's success-fee structure is tied to RevPAR improvement, and the account manager reviews declined recommendations in monthly calls.

What does the vendor say in marketing? The deck includes a slide stating that in markets where the product has high penetration, "the entire market benefits from rational pricing discipline."

Ines's assessment is that the last item is the most dangerous thing in the file, and that the combination — current non-public data, high market penetration, firm-specific recommendations, adherence monitoring, and a marketing claim about market-wide effects — is close to the fact pattern plaintiffs plead.

She does not simply say no. She proposes conditions.

Contract conditions. Wrenhaven will not contribute forward-looking booking data. It will contribute historical occupancy at a lag. The vendor will not disclose to Wrenhaven any competitor-identifiable data, and will not tell Wrenhaven which competitors participate or what their acceptance rates are. There will be no acceptance-rate target, no success fee tied to market pricing outcomes, and no vendor review of declined recommendations.

Internal conditions. Every recommendation is reviewed by a human who documents the basis for acceptance or rejection. Revenue managers are trained that the recommendation is an input, not an instruction, and that the correct question is what maximizes Wrenhaven's return given its own occupancy and cost — not what the market is doing.

Communication conditions. Nobody at Wrenhaven attends a vendor user group where pricing strategy is discussed. Nobody says, in writing or otherwise, that the tool works better when competitors use it.

Documentation. Ines writes a memorandum recording the evaluation, the risks identified, and the conditions imposed. If a plaintiff ever pleads this case, that memorandum is the best evidence Wrenhaven has.

The commercial team is unhappy about losing the forward-booking data, which is where much of the model's value sits. Ines explains that the alternative is a class action, and the CFO agrees. That is often how this conversation ends.

Worked example two: building it internally

Tobias Nwachukwu leads pricing at Calder Industrial, a distributor of fasteners and fittings, and wants to build a dynamic pricing model rather than buy one.

The internal build eliminates the hub. There is no vendor pooling data from competitors, no user group, no marketing deck. What remains is a data question.

Calder's data sources: its own transaction history (fine), its own cost data (fine), published competitor list prices scraped from public websites (fine, subject to terms-of-use and computer-access considerations), and — the item that needs attention — competitor pricing information reported by Calder's sales representatives after customer conversations.

That last category is where most internal models acquire an antitrust problem, and it does not come from the algorithm. It comes from how the information was obtained. Information a customer volunteers about a competitor's bid is ordinary market intelligence. Information a competitor's salesperson gives to Calder's salesperson at an industry event is an inter-firm communication, and feeding it into a model does not launder it.

Tobias's compliance work therefore focuses on provenance:

A data intake policy requiring every competitor data point to record its source.
A prohibited sources list: competitor employees, shared consultants who work for competitors, trade association staff with access to member data, and anyone with a confidentiality obligation to a competitor.
Training for the sales force on what may be collected and what must be reported and not used.
A quarantine process for information that arrives unsolicited from a competitor — logged, not entered into the model, and escalated.

He also documents the model's independence: Calder's pricing decisions are made by Calder personnel, using Calder's model, built on lawfully obtained data, without any assurance to or from any competitor.

The result is a model that produces most of the commercial benefit with a small fraction of the legal risk, which is generally what an internal build achieves.

Worked example three: the investigative demand

Priyanka Raval is outside counsel to Stanhope Storage, which has received a civil investigative demand from a state attorney general concerning its use of a revenue management platform.

The immediate steps are the ordinary ones — legal hold, custodian identification, scope negotiation — but the substantive work is a candid internal assessment, done early, of what the documents will show.

The categories she prioritizes:

Vendor communications. Sales decks, implementation materials, account management correspondence, user group agendas and attendee lists. Anything the vendor said about market-wide effects or about competitor participation is now Stanhope's problem, whether or not Stanhope believed it.

Internal statements about competitors. Emails, board materials, and analyst call scripts. The sentence "the tool only works if everyone uses it" appears, in some form, in a remarkable proportion of these files. Priyanka needs to know whether it appears in this one.

Acceptance rates. How often did Stanhope adopt recommendations? Was that tracked, targeted, or reported to the vendor?

Deviation history. Did Stanhope ever price below the recommendation, and what happened? A history of independent deviation is powerful evidence of independent decision-making. A history of never deviating, or of deviating and being called by the account manager, is the opposite.

The counterfactual. What did Stanhope's pricing look like before adoption, and what changed? An economist will need this.

Priyanka's advice on posture: the defence is not that algorithms cannot conspire. It is that Stanhope made its own decisions, in its own interest, using a tool it selected, without any understanding or assurance that competitors would do anything in particular. That defence is built from documents, and the documents already exist. The job is to find out what they say before the attorney general does.

Pleading and proof: where these cases are won and lost

Because Twombly sits at the front of every one of these cases, a great deal of the doctrine is being made at the motion to dismiss stage, on allegations rather than evidence. That has consequences for how the law is developing and for how a defendant should think about exposure.

At the pleading stage, the question is whether the complaint alleges facts suggesting agreement rather than merely parallel conduct. Complaints that survive tend to plead: the vendor's marketing statements about market-wide pricing effects; the pooling of non-public, real-time data; a high share of the market subscribing; contractual or practical pressure to adhere; and a forum where subscribers met. Complaints that fail tend to plead only that competitors used the same software and prices went up.

At class certification, the fight moves to whether antitrust impact can be shown with common proof. A pooled-data pricing model that produces individualized recommendations creates a genuine problem for plaintiffs: if each subscriber received different advice, based on its own inventory and cost, the impact on each class member may require individualized inquiry. Defendants should preserve this argument early and should resist stipulating to descriptions of the model that make it sound more uniform than it is.

At summary judgment, Monsanto's tends-to-exclude standard governs. Evidence of independent deviation, of internal disagreement about whether to follow recommendations, and of pricing decisions driven by firm-specific factors is what defeats the inference. This is why contemporaneous documentation of pricing decisions is worth more than any policy statement.

At trial, the case is usually about a handful of documents. The email in which a manager wrote that the tool works only if competitors use it. The slide that promised market-wide discipline. The account manager's note about a subscriber who kept discounting. Antitrust trials are won and lost on ten pages, and everyone in this field knows it.

The international dimension

A pricing programme deployed across borders faces regimes that reach further than U.S. law in specific respects.

The European framework treats "concerted practices" as a distinct category alongside agreements, and the concept is broader than the U.S. requirement of agreement. A concerted practice exists where undertakings knowingly substitute practical cooperation for the risks of competition, without concluding an agreement. A firm that receives competitively sensitive information from a competitor is presumed to take it into account unless it publicly distances itself — a presumption with no direct U.S. analogue, and one that makes receipt alone dangerous.

European authorities have also stated that firms remain responsible for the conduct of pricing software they deploy, and that a firm cannot escape liability by pointing to an algorithm.

The United Kingdom has pursued cases involving pricing software used by competing online sellers, treating the shared software as the mechanism of an agreement.

Other jurisdictions vary widely, and several have adopted or proposed provisions addressing algorithmic coordination expressly.

The practical implication for a multinational is that the U.S. agreement requirement is the most permissive standard it will face, and a programme designed to the U.S. line may fail elsewhere. Design to the strictest applicable regime, or accept that the programme must be configured differently by region — which, for a global pricing platform, is often harder than designing conservatively in the first place.

A note on what is not the problem

It is worth being clear about what algorithmic pricing does not, by itself, make unlawful, because overbroad caution has costs of its own.

Dynamic pricing is lawful. Changing prices in response to demand, inventory, time, and competitor behaviour is ordinary competition. Airlines, hotels, ride-hailing platforms, and electricity markets have done it for decades.

Fast pricing is lawful. Responding to a competitor's posted price in seconds rather than weeks is not an agreement. It may produce a market equilibrium closer to the oligopoly outcome, and that is a live policy question, but it is not a Section 1 violation.

Using a vendor is lawful. Buying analytics from a specialist is efficient and universal. The question is what data goes into the pool and what assurances come out.

Sophisticated modelling is lawful. The complexity of the model is legally irrelevant. A firm does not become a conspirator by hiring better data scientists.

What is not lawful is agreeing with competitors — on price, on the method of setting price, or on adherence to a common recommendation — and using software as the medium of the agreement does not change that. The distinction is between a tool that helps a firm compete and a mechanism that helps firms stop competing.

Where this is heading

Three developments are worth watching, and each affects how a programme should be designed today.

Legislative activity. Several proposals would create presumptions or per se rules for pricing algorithms trained on competitor data, and several state legislatures have moved faster than Congress. A programme designed only against current case law may find itself on the wrong side of a statute.

Agency guidance. Enforcement agencies have signalled through statements of interest, speeches, and amicus filings that they regard the pooling of non-public competitor data into a common pricing model as a serious concern, and that they reject the argument that the absence of direct communication defeats a conspiracy claim.

Sector regulation. Housing, healthcare, and insurance are attracting sector-specific attention, and some jurisdictions have adopted or proposed restrictions on the use of competitor data in pricing tools for particular markets.

The design principle that survives all three: build the tool so that it improves the firm's own decisions, and so that it cannot function as a mechanism of mutual assurance. Own data, historical inputs, aggregated benchmarks, human review, documented independent decisions, and nothing that tells the firm what its competitors will do. A pricing programme built that way delivers most of the commercial value and is defensible under any of the standards now in play.

Remedies and exposure

The stakes deserve a paragraph, because they drive how seriously a board takes this.

A per se price-fixing violation exposes a firm to treble damages in private litigation, joint and several liability with no right of contribution, and the aggregation of a nationwide class. It exposes individuals and firms to criminal prosecution, with substantial statutory maxima and the possibility of fines calculated as a multiple of the gain or loss. It exposes the firm to parallel state enforcement, to follow-on litigation by indirect purchasers under state repealer statutes, and to actions abroad.

Even a rule of reason case that the defendant eventually wins is expensive: economists on both sides, extensive discovery of pricing data and model documentation, depositions of data scientists who have never been deposed, and a multi-year timeline.

And there is a category of cost that shows up before any of that: the disclosure and diligence consequences. An open investigation must be disclosed in financing documents, considered in an acquisition, and explained to insurers. Deals have been repriced over less.

Against that, the cost of the compliance architecture described above — vendor diligence, contractual conditions, documented human review, training, and a periodic audit — is very small. The asymmetry is the argument.

The one-page summary a board should receive

Any firm using a third-party pricing tool should be able to answer these questions in writing, on one page, at any time:

What data do we send to the vendor? Is any of it non-public? How current?
What data does the vendor combine ours with? Whose? How many participants? What share of the market?
What do we get back? A market benchmark, or a firm-specific price recommendation?
Can we identify any competitor's data from what we receive? Can they identify ours?
Are we told who else participates, or what they do with the output?
Are we required, incentivized, or pressured to follow the recommendation?
Who reviews recommendations, and is the decision documented?
What did the vendor's sales materials say about market-wide effects?
Do our people attend any forum where competitors discuss pricing?
When did we last audit this, and what did we find?

A firm that can answer all ten has a defensible programme. A firm that cannot answer question 2 or question 8 has work to do, and should do it before someone else asks.