// analysis
Airfares: can AI agents distort the signals behind the price?
Open the documentsThe source documents
The documents cited in the analysis and, where identified, their associated passages. An automatically detected association is not a verified claim.
Consult the references in the article below. No document has been identified for this shortcut.
Sources and reading context
Topics
Contents9 sections
A traveller needs a return flight, a checked bag and an arrival before dinner. Their AI agent tries different departure times, moves the trip by a day, checks a nearby airport and repeats the exercise across several distributors. At the end of the search there is still one person, one trip and, perhaps, one booking. In the systems processing those requests, the activity may look much larger.
This is a hypothetical scene. The economic problem is already worth examining: flight-search data are used to inform some commercial decisions and demand forecasts. When the way people search changes, the meaning of the data can change too. An airline that misreads that shift could protect too many seats for future customers who never arrive. A distributor could face a much larger computing bill for each ticket it sells.
Several channels need to be separated. Shopping requests consume computing resources and can inform demand forecasts. More requests from the same traveller can change the meaning of those signals. The causal effect of AI agents on fares remains to be measured in the public sources reviewed through 5 November 2026. The mechanisms below distinguish computing costs, possible forecasting errors and better comparison for the customer.
The interactive simulator follows one traveller and an assistant making a hundred searches. A fictional seller mistakes those searches for additional customers, then recognises their common origin.
Pricing a seat against tomorrow’s passenger
Machines were comparing fares long before conversational assistants. France’s civil aviation authority, the DGAC, describes an automated price-collection system that obtains fares through Travelport or, with permission, directly from certain airlines. Automated shopping can serve an entirely legitimate measurement purpose. 1
To understand what agents might disrupt, start with the seller’s calculation. An empty seat loses its commercial value when the aircraft departs. Selling it cheaply today, however, may prevent the airline from selling it to someone willing to pay more tomorrow. Revenue management organises that trade-off. In an article published in March 2026 and updated in April, Amadeus describes the choice between accepting a booking now and preserving capacity for future demand as the enduring economic foundation of these systems. 2
Consider a deliberately stripped-down example. One seat remains. A customer will pay €180 today. Another might pay €300 later. With a 40% probability of making that later sale, waiting produces expected revenue of €120. Taking €180 now is preferable. If the model instead estimates an 80% probability of the later sale, expected revenue rises to €240. Within this simple model, protecting the seat becomes rational.
The amounts and probabilities are illustrative assumptions, with no measured relationship to bot traffic. The example allows only one potential future sale, assumes an unsold seat has no residual value and uses a risk-neutral seller. Connections, cancellations and differences in cost are left out. Its purpose is to expose the transmission mechanism: an error in expected demand can change an availability decision even when no additional seat has been sold.
Method and limitations
An immediate 180 euro offer and a possible 300 euro future sale. At 40%, expected future revenue is 120 euros; at 80%,240 euros. The calculated threshold is 180/300=60%. One possible future sale, zero residual value for an unsold seat and a risk-neutral seller. Probabilities are assumed, with no estimated relationship to request volumes.
In a system using booking classes, the decision may close access to a cheaper offer. A booking class here denotes selling conditions and a fare level, not necessarily a separate cabin. Continuous pricing allows other adjustments without sticking to the same predetermined fare steps. Architectures differ, but the underlying question remains the value of capacity retained for later. 2
Kevin R. Williams’s research on airline markets distinguishes responses to demand shocks from adjustments associated with changes in travellers’ willingness to pay over time. This research on algorithmic pricing provides the economic setting: several forces can move a fare between two searches. 3
Between page refreshes, customers may book, availability may change or a different offer may appear. Isolating a search effect would require comparing the same flight at the same instant with and without the additional request. That counterfactual is central to identifying a causal effect.
Search activity already enters the forecasting process
A demonstration by software vendor PROS dated 18 September 2025 describes agents using search and booking data to adjust forecasts, integrated with revenue management and pricing tools. This is a supplier’s claimed functionality; robustness in individual deployments would require an audit. 4
FareNet documentation from PROS combines searches, conversions and load factors to inform fare adjustments. It concerns data collected in the relevant booking engines. The path an agent follows determines which requests actually reach that collection point. 5
That distinction matters. Depending on the route through the system, an agent’s request might end at a cache, reach a distributor or trigger several downstream calls. A single response can contain many offers. Fare options, technical requests and travellers are different units. Even “a search” can mean different events to different organisations.
Research by Hortaçsu, Natan, Parsley, Schwieg and Williams gives a concrete example of searches informing demand measurement. In their airline application, the authors measure consumer arrivals from people initiating a search on an airline website, using clickstream records. Section 5.1 of the paper published in 2023 describes this measurement choice. It explains the value of shopping signals for estimation; the study concerns a different question from disruption by AI agents. 6
The attraction for a forecaster is understandable: searches provide information before a purchase takes place. The challenge is interpreting that information. A person may be choosing among several departure dates. An agent may inspect every one of those dates systematically. Treating each alternative as an independent purchase intention would turn a choice set into additional demand.
Method and limitations
1000 travel intentions and 100 bookings over the same period and at the same measurement point. Ten requests per intention yield 10000 requests;1000 yield one million. Conversion per request changes from 1% to 0.01%, while bookings per intention remain at 10%. l 0 g calculation, not an observed industry statistic.
Imagine a fixed cohort of 1,000 travel intentions, of which 100 become bookings within the observation period. At ten requests per intention, there are 10,000 requests and one booking per hundred requests. At a thousand requests per intention, there are one million requests and one booking per ten thousand. Request-based conversion falls from 1% to 0.01%. Conversion per intention remains 10%.
The scenario keeps purchases fixed and changes only search intensity. It exposes a possible measurement failure: the denominator evolves with the way offers are explored. Dates, destinations and prices raise similar issues. An assistant can systematically test options its user will rarely buy.
The fare outcome depends on processing. A model placing substantial weight on raw volume could overestimate demand. Another might read weaker conversion as a lack of appeal. A recalibrated system could group repetitions and identify meaningful alternatives more accurately. The direction can therefore change with the data and decision rules.
The scenario below illustrates a possible error: counting every search from the same traveller as another customer. The numbers, fares and decision rule are entirely fictional. They do not reproduce any airline’s pricing system.
l0g / interactive simulation
One traveller, 100 searches
Your assistant compares flights. Does the seller know it is searching for just one person?
The game’s assumption: the seller mistakes searches for new customers.
You
“I need a ticket to Lisbon.”
Ready to compare flights.
Searches sent
You look up a flight.
The seller’s website
Normal reading“A traveller is looking for a ticket.”
You look up a flight. A ticket is offered at €180.
Fictional example: the prices and numbers are illustrative. No real airline is being reproduced.
All l0g tools →The software-security foundation OWASP calls intentional automated manipulation of metrics skewing. That threat supplies a technical comparison. Unintentional noise from a legitimate shopping assistant remains a separate analytical hypothesis in this investigation. 7
The first bill is for computing
On 7 October 2025, the International Air Transport Association, IATA, described pressure on systems and costs from rising searches per booking. Its public introduction to a white paper mentions 14 airlines, 11 IT providers and eight sellers contributing to the work. This is an industry assessment. Any look-to-book ratio needs explicit search and booking units. 8
The mechanism is straightforward. When exploration grows faster than sales, the work associated with each completed order can increase. The expense also depends on the complexity of the answer, the extent of shared processing and how often a response can be delivered without a complete recalculation. A broad search and a targeted verification need not impose the same cost, even if each counts as one request.
Caching addresses that problem by keeping a result temporarily so that the work need not be repeated. On 5 February 2026, Sabre announced a product that chooses between live airline polling and cached results. Its release also presents the architecture as a way for agencies to manage traffic constraints and exposure to airline penalties. This is a commercial description, without a public tariff schedule from which a universal cost per search could be derived. 9
Method and limitations
A cache shares temporary responses. The agent then shortlists candidates, rechecks price and availability live and requests separate approval to buy. The stages are conceptual, drawing on the cache/live trade-off discussed by IATA and Sabre. Benefit depends on reuse and freshness; cached results guarantee neither price nor availability when buying.
An agent could begin a broad exploration with indicative information and reserve live checks for a shortlist. The savings would depend on how widely responses can be reused, how long they remain useful and how quickly offers change. Final verification is a separate step: a remembered price does not guarantee that the same seat will remain available on the same terms when the user buys.
The trade-off raises a distribution question. A smaller intermediary may need frequent access to supply a useful comparison. Expensive or restricted access could leave it showing fewer options or less current information. A larger service able to pool searches may have an advantage. That is a possible economic consequence of access conditions, to be investigated contract by contract, rather than evidence that exclusion has already occurred.
Who pays this expense depends on contracts and competition. An airline could absorb it, charge a distributor, use caching to reduce repeated computation or restrict some uses. The share eventually reaching travellers depends on those choices and competitive alternatives.
Holding a seat and finding a better alternative
A price request generally does not reserve a seat. Other actions go further, creating a temporary booking or holding availability. When holds accumulate without purchases, they can deprive other customers of inventory that would otherwise be for sale. OWASP’s denial of inventory category explicitly includes flight seats. 10
A hold first affects availability: a customer may lose access to an offer, buy a more expensive alternative or abandon the purchase. Following this channel requires records of temporary reservations, quantities, duration and release. Price searches and actions that hold seats therefore need separate observation.
Agents could also strengthen the passenger’s position. Consider an assistant that finds an acceptable connection, checks baggage charges and discovers an airline its user would otherwise have missed. By genuinely expanding the alternatives, it may make some price differences harder to sustain. This is a material countervailing economic mechanism, whose size remains to be measured.
Method and limitations
Suppliers and the trade body describe the computing-cost channel. A biased forecast may change an offer depending on models and corrections. Inventory holding is a generic OWASP threat, without a measured airline frequency here. Better comparison may strengthen competition. Arrows illustrate conditional mechanisms, with no observed net effect or quantified fare impact.
The benefit depends on visible offers and the assistant’s objective. A low base fare can lose its advantage after baggage charges. Partial market coverage limits alternatives. Commissions or paid placement can also influence rankings. These features determine the quality of comparison.
Another scenario for controlled testing is assistants following similar rules and buying together when a fare crosses a shared threshold. Real bookings would then cluster in time, leaving the airline with more correlated demand to manage. This remains a working hypothesis.
Revision frequency can amplify or dampen such interactions. A slow model, supervised by analysts and anchored in sales, behaves differently from a tool updated rapidly from recent signals. Assessing an unstable loop requires following that timing, its transmission to fares and buyers’ responses.
Personalised pricing requires a different investigation
A price increase offered to everyone and a price tailored to a particular person are distinct phenomena. The former may reflect availability or anticipated demand. The latter involves using individual or group characteristics to determine the pricing offer. The presence of AI does not, by itself, establish which mechanism is operating.
The French precedent is specific and historical. In its 25 January 2016 account, France’s data-protection authority CNIL reported that joint inspections with consumer-protection authority DGCCRF observed price variations without establishing IP-address-based modulation on the sites examined. The period and scope bound that finding. 11
On 28 October 2026, the US Federal Trade Commission ordered nine companies, including Delta, United and JetBlue, to provide information about personal data in pricing and, in other sectors, compensation. The Section 6(b) exercise is a study; its launch establishes an information demand, a different status from a finding of wrongdoing. The 45-day response period starts when the orders are received. 12
The published model order makes the inquiry concrete. It asks about data inputs, pricing systems, conditions for modifying or overriding their outputs and evaluations of the tools. Those records could allow company statements to be tested against procedures and assessment results. 12
Delta set out its position in a 31 July 2025 letter published on 7 August. It described route- and flight-level aggregate data and analyst supervision. It said personal information was excluded from the data shared with Fetcherr. This supplies the company’s account, with its date and evidentiary status. 13
Even an entirely aggregate system could, in principle, misinterpret automated searches. Conversely, a system that filters repetition perfectly could raise a personalisation question if it uses other individual information. The two investigations require different evidence.
The passenger’s assistant creates another boundary. A spending limit, family emergency or fixed departure date helps it choose. Passing those constraints to a seller could also reveal the customer’s bargaining position. The flow of private constraints therefore warrants an audit, without attributing that transmission here to an identified service.
Method and limitations
The agent keeps budget, urgency and flexibility within its private decision space. It sends the information necessary for searching, receives offers and obtains explicit approval for a binding action. Mandate and remuneration should be clear. Returned data remain content to examine. The figure does not describe a supplier implementation or attribute budget disclosure or budget-based pricing to any company.
A protective design can keep private constraints on the user’s side, disclose only the information required for the search and require separate confirmation before purchase. It should also explain who pays the intermediary and how results are ranked. The software’s mandate matters as much as its search capability: an assistant effective at maximising commission would be pursuing a different objective from one instructed to minimise the total cost of a trip.
To measure the effect, follow the decision all the way through
The first task is to give every counter a precise unit. What constitutes a request? At which layer is it counted? Does one response contain one offer or hundreds? Is the denominator a reservation, an order, an issued ticket or a passenger? Can the same travel intention appear in several sessions? Without answers, two apparently precise search-to-purchase ratios may not be comparable.
The paths through the system must then be separated. A cached exploration, a live request and a seat hold should leave distinguishable traces. That need not require distributing the traveller’s civil identity throughout the supply chain. Temporary grouping of requests, limited in purpose and duration, could be examined. Its tracking risks and attribution errors would need assessment. A travel intention inferred from a collection of events remains an estimate.
The inquiry must follow the signal through the decision: the input reaching the model, revised forecast, recommendation and fare applied. Cancellations, available capacity, competitors and human interventions can change the result. Records should identify both a vulnerable stage and filters that absorb noise.
A persuasive experiment would begin in an isolated copy of the system, using authorised data and explicit scenarios. Researchers could vary repeated searches while holding underlying intentions and purchases constant, then examine forecast and decision differences. Findings would apply to the configurations tested. An observation of real operations would require a consenting partner and a design that accounts for interactions across flights.
Nearly simultaneous quotes, matched for services across channels, can illuminate differences in access, freshness and total price. A causal search effect needs an additional protocol. This documentary review involved no test bookings or traffic generation against airline systems.
The decisive test would increase an agent’s searches while holding travel intentions and purchases fixed, then observe the estimated number of buyers and resulting decisions. Effective filtering could contain part of the statistical risk. Computing costs would still depend on the volume served and its processing.
For travellers, the promise should be judged by a concrete outcome: access to suitable offers, protection of private information and the expected total cost. Request volume alone says little about service quality. Agents could broaden choices and shift bargaining power; access rules and contracts will also determine who benefits from this way of buying.
Further reading
Pricing algorithms and competition raise another issue: sellers reacting to rivals’ decisions. Our investigation into connected-car data and insurance follows personal information through to commercial use.
Method and limitations
Documentary review closed on 5 November 2026. Sources include public authorities, original academic work, supplier documentation and a dated airline statement from 2025. Commercial capabilities are attributed to their providers. The review used public documents, with no interviews, internal logs, test bookings or traffic generation.
Both numerical scenarios are hypothetical. The possible future sale gives €300 × 40% = €120 or €300 × 80% = €240; indifference occurs at 180/300 = 60%. For the cohort, 100 bookings divided by 10,000 or 1,000,000 requests yield 1% and 0.01%, while 100/1,000 intentions stays at 10%. These calculations isolate mechanisms using entirely illustrative inputs.
The full demand-estimation paper was consulted for its arrival-measurement method. Williams’s dynamic-pricing study is used for a qualitative point supported by its institutional notice and abstract. The IATA public introduction was reviewed separately from the full white paper, which remains outside the corpus. The FTC model order is accessible; company responses and actual AI-agent effects remain to be documented. The opening artwork is conceptual.
Editorial licence: CC BY 4.0.
Sources and references
- Indice des prix du transport aérien de passagers
Passages and scope
Methodology, questions 10 and 11. First published 21 April 2022, updated 15 October 2026. Automated collection through Travelport or authorised direct access. This describes price observation.
- Impact of Artificial Intelligence on Airline Revenue Management
Passages and scope
Published 26 March 2026, updated 29 April. The trade-off between a booking now and future demand. A supplier’s account, identified as such.
- The Welfare Effects of Dynamic Pricing: Evidence from Airline Markets
Passages and scope
Institutional notice and abstract of the 2022 article. Demand shocks and changes in willingness to pay. No numerical welfare estimate is reproduced.
- Smarter Market Moves with PROS Revenue Management Agents for Airlines
Passages and scope
Demonstration dated 18 September 2025. Search and booking data in forecasts; PROS’s claimed functionality, with no deployment audit here.
- FareNet Reporting Suite
Passages and scope
Undated product documentation. Searches, conversion and load factors in the relevant booking engines. Supplier-defined scope.
- Demand Estimation with Infrequent Purchases and Small Market Sizes
Passages and scope
Full paper consulted, section 5.1: clickstream records and consumers initiating a search provide the arrival measure in the airline application. Published research from 2023, addressing a different question from AI-agent effects.
- OAT-016 Skewing
Passages and scope
OAT-016 taxonomy: intentional automated manipulation of metrics. Unintentional noise from legitimate assistants is a conditional hypothesis in this article.
- Look-to-Book: managing search volumes while creating customer-centric Offers
Passages and scope
Public introduction dated 7 October 2025. Participants: 14 airlines, 11 IT providers and eight sellers. The full white paper remains outside the reviewed corpus; an industry perspective.
- Cache-powered Intelligent Shopping : annonce produit
Passages and scope
Release dated 5 February 2026 on Cache-powered Intelligent Shopping. Cached/live architecture and search constraints. Advertised performance remains a commercial claim.
- OAT-021 Denial of Inventory
Passages and scope
OAT-021 taxonomy. Accumulating holds without purchases, including flight seats. This requires an inventory-holding action distinct from a price query.
- IP Tracking : conclusions de l’enquête conjointe
Passages and scope
Historical account dated 25 January 2016. CNIL/DGCCRF inspections and observed price variations, limited to the inspected sites and period.
- FTC Launches Personalized Pricing and Compensation InquiryFTC official model order
Passages and scope
Release dated 28 October 2026 and official model order consulted. Nine recipients, with 45 days from receipt to respond. Section 6(b) study. The order requests data, models, modifications or overrides and assessments, including specifications 7-10 and 18-23.
- Delta responds to misinformation around AI pricing
Passages and scope
Letter dated 31 July 2025, published 7 August. Delta’s account of aggregate data and analyst oversight; an interested-party statement.
This analysis is not investment advice.
// cite this analysis
l0g, “Airfares: can AI agents distort the signals behind the price?”, l0g.fr, published November 05, 2026, updated November 05, 2026, https://l0g.fr/en/analysis/airfares-ai-agents-demand-signals/
$ cd ../analysis