How Can AI Optimize the Front End of Airlines Operations?

Any article we read about the use of AI in the airlines industry primarily focuses on chatbots, virtual assistants, robot kiosks, digital ticketing systems or the potential use of facial recognition for check-ins. However, there is a lot more that AI has to offer the airlines in as far as the front end is concerned.

In an adjoining article, I wrote about how AI can help the airlines depart and arrive on time. In this article I discuss opportunities for AI to substantively leapfrog front end improvements in the airlines, considerably reduce capacity risk, maximize revenue aggregation and optimize revenue per unit of capacity (seat-trip and per unit available payload). AI can principally alleviate the front-end issues of distribution, market segmentation by traveler type, the willingness to trade off the risk of spoilage for low price, product and service differentiation by physical cabin, virtual cabin, and by terms of sale, network design from point-to-point and line-hauls to connecting networks, capacity, inventory, revenue management and pricing systems design, issues of dynamic pricing, optimizing revenue on payload-limited operations, and finally, optimizing the economics of unbundling and ancillary fees.

Distribution – Journey from Airline Reservations Systems to the Web

Airlines automated the manual, card-based reservations systems in the 1960s, in American’s case re-utilizing Cold War era IBM technology originally designed to link remote, ballistic missile/bomber radar and missile/interceptor launch sites with central command and control sites.

The reservations systems were complex, but also simplex, based on TTY teletype circuits – a one-way display available inventory (seats) and pricing in response to an itinerary request, followed by “buy or ignore” – with no capture and retention of information about offered itineraries and price points that did not attract sales.

Airlines rolled out improved versions of their internal systems to selected, large travel agencies/groups starting in the late 1970s, just as the CAB and Congress deregulated the Airline industry. Still simplex. Airlines begin to design and operate ‘hub and spoke’ networks that facilitate aggregation of revenue to and from many origin and destination spokes on a single flight to or from a hub. Revenue management systems evolve to consider the need to optimize use of available inventory on connecting itineraries, not just single flight segments.

In the mid-1980s, US DOT required airlines to eliminate differential pricing for sales, then later to spin off their internal reservations systems as stand-alone Computer Reservations Systems (CRS, later Global Distribution Systems – GDS) firms, which provided inventory hosting, itinerary pricing and airport passenger services systems for relatively high ‘segment booking fees’ to airlines, while paying travel agencies inducements to retain and use their systems to book.

Airline hubs grow in size and complexity, hundreds of flights to scores of destinations, multiple times daily, typically up to eight directional ‘connecting banks’ of flights per business day. Airlines start to capture flight booking data for use in setting capacity and price/inventory levels by flight. Airline and GDS communication networks evolved from pure TTY to incorporate some full-duplex realtime systems, such as for ‘last seat availability’ and certain passenger service functions.

By the mid-1990s, in a bid to control distribution costs, airlines began to eliminate travel agency commissions, to ‘disintermediate’ what they desire to be direct relationships with customers. Travel agencies still handled more than 70 percent of bookings.

Airlines had evolved revenue management systems to consider much of the historical booking data they held in silos in the form of ‘profiles’ – how rapidly in advance of day of departure (flights are typically loaded 330 days in advance of departure or DIAD) flights book, by class of inventory and fare level. Fast booking flights typically attract attempts to improve pricing, while slow booking flights might attract more discount inventory.

By the late 1990s, with widely available realtime fiber communications capacity, server, TCP/IP and html technology, airlines stood up their own e-commerce websites for direct sales to consumers.

With the flexibility that their own e-commerce sites provided, Airlines were also able to use XML representations of ancillary services such as seat selection, pre-booked meal selections, and advance bag check fees to sell direct to consumers what GDS firms still struggled to incorporate into their seat-sales-only oriented, legacy systems.

By 2018, more than half of total network airline revenue (and a far larger percentage of transactions) is direct from consumers. Credit Card discount fees are now on the radar for avoidance, eliciting alternate settlement proposals including eChecks, bitcoin and the like. By 2018, the two largest GDS firms had launched XML ancillary services sales capabilities, to compete with airline e-commerce sites.

For an airline, implementing a robust AI powered know your customer (“KYC”) solution that is able to determine the best possible personalized ticket price for a particular customer that is unique to that customer based on a variety of data points, will prove to be enormously valuable.

Market Segmentation – by traveler type, willingness to trade off the risk of spoilage for low price

Airlines have long understood that different passenger-trip types have different sensitivities and elasticities, with business travelers (less than 30% of unique customers at network airlines) preferring schedule times, flight frequency, service and recognition, would pay-up for these attributes, while leisure travelers (more than 70% or unique customers at network airlines) were in most cases purely price seekers.

Airlines began in the late 1970s to design fares with associated rules and conditions of that would ‘fence’ discounts from use by business travelers, chiefly with non-refundability and length of stay restrictions.

By the 1990s, airlines typically offered inventory on itineraries under up to 26 ‘classes’ (corresponding to a letter of the alphabet), ‘nested’ within and under physical cabin designators, so, for example (F/C/Y cabins, with booking and fare designators P/F/A/I/R, J/C/E/G/S, Y/K/M/V/L/X/Z. International itineraries might carry different sub-classes than wholly domestic itineraries.

Airlines engage in very granular segmentation because they realize that to harvest the maximum revenue from under the demand/price frontier, they must have as many equilibrium points as possible, one per segment.

It is however important for an airline to re-visit its market segmentation strategy with additional data points, and deploy deep learning and advanced analytics to gain a deeper understanding of its market and the customer.

Product/Service Differentiation – by physical cabin, virtual cabin, and by terms of sale

Most airlines, like railroads and maritime transport before them, choose to offer a differentiated product, with First and Standard classes of service and amenities. Distribution capabilities were designed around F (First Class) and Y (Economy) cabins, then expanded in the 1980s to incorporate a C (Business Class) cabin, with more ‘working room’ than economy but less caviar and Champagne than First. More recently, virtual cabins have been opened up to expand the number and attributes of offers and price points airlines can match with customer preferences, to harvest more of the potential under a demand/service attribute/price curve. The use of AI and data science here will also prove to be of value.

Network Design – from point-to-point and line-hauls to connecting networks

From the CAB-era line-hauls that mirrored railroad routings (LGA-BNA-TUL-DFW-PHX-SAN), airlines evolved in the deregulated era that began in 1978 to embrace network dynamics. LGA-DFW-SAN, plus LGA-DFW-PHX and LGA-DFW to forty other cities to the West, including many in Central and South America, and Asia. 

Embracing networks had other advantages, as well. By aggregating local revenue on a spoke, say, LGA-DFW, with onward revenue to forty other cities, any point-to-point competition on any single spoke was less impactful to the network, and allowed network airlines to retain revenue and customers much more easily. However, here as well, AI and deeper data analysis can prove valuable in optimizing network design and performance.

Capacity/Inventory/Revenue Management/Pricing Systems Design

Beginning in the late 1980s, eschewing the use of large aircraft with attractive low seat costs, but which were difficult to fill, year-round, at economic fares, airlines began ‘right-sizing’ scheduled seat inventory by down-gauging smaller market and off-peak time of day domestic flying to contract and affiliated regional operators. This capped capacity risk by eliminating low occupancy factor, low daily frequency, 120 seat and larger mainline aircraft, replacing them with 37-76 seat regional-operated aircraft, with the added benefit of additional daily economic flight frequencies.

Airlines did likewise throughout the mainline fleet plan spectrum, eliminating the largest wide body aircraft such as 747, DC10, L1011, A340 in favor of more economic twin-engine equipment such as 777, 767, 757, A300, and A330. More recently, even the 777 is judged too large by many network carriers and is being replaced by the 787 and A350. The A380 is a unique business case that never sold in economic production quantities. The goal was to cap capacity risk by eliminating low occupancy factor, low daily frequency, very large aircraft operations that were uneconomic except in peak season of the year.

As airline networks evolved in scale, airlines invested in academic and commercial operations research techniques to evolve from manual to mechanized to automated techniques by which to monitor and manage inventory, and ‘open and close’ seat availability display at various fare-price points. As airline network topology evolved during deregulation from CAB-dictated legacy line-haul and point-to-point, toward hub-and-spoke connecting networks, the mathematics of revenue management had to change.

Originally designed as manual inventory controls on individual flight legs (e.g. LGA-BNA) or line-hauls (LGA-BNA-TUL-DFW, much like a railroad would), evolving to mechanized point-to-point revenue management (LGA-DFW, no matter intermediate service points), then to automated origin-destination revenue management (LGA-SAN via connecting flight legs, LGA-DFW/DFW-SAN), to fully optimized, realtime calculation of ‘Expected Marginal Seat Revenue’ (EMSR)-based ‘bid-ask’ pricing. Each serial improvement in the mathematical logic of revenue management was judged to deliver from a 6 to 9 percent improvement in network revenue. Outlier cases were identified and pitched to teams of market analysts who approved or over-rode and fine-tuned automated results. There is still human capital in the equation, but this is a great AI opportunity to explore and mine.

Moving beyond mass market forecasting, and based on voluminous, depersonalized, historical data, airlines had the ability to determine what past prices on what past terms various passenger types had willingly paid (‘bid prices’), and thus could set thresholds for prices they were thus willing to accept, displaying these as ‘ask prices’. Airlines now engage in price discovery experimentation, set prices, and judge success by the ability to improve on historical EMSR.

As airlines develop historical bases of add-on, ancillary service and amenity sales data, they are creating the ability to estimate an ‘EMAR’ Expected Marginal Ancillary Value -- what revenue beyond the seat sale, a customer-type, or eventually an individual customer, is likely to spend. Like fare-pricing, airlines are engaging in price discovery to determine what they can successfully charge for ancillary services, for example, checked baggage and overhead bin space in peak season and during holiday travel periods, and move that threshold higher.

Similarly, airlines have developed portfolios of non-air products including affinity credit cards that are approaching $2 billion annually for the large network carriers, for the sale of mileage to credit card issuers, and shares in acquisition, interest charges and annual fees paid by consumers. All these areas can materially benefit from data analysis and AI based predictive modeling exercises.

Dynamic Pricing

In order the maximize revenue under the demand curve, airlines are experimenting with replacements for traditional published pricing, envisioning a time when they can set an EMSR seat price for each unique individual at the moment they express interest, with the airline understanding this to be worth a final value of EMTR Expected Marginal Trip Revenue, including the EMAR ancillary services sales value. Consumers will be lulled in by a low-price offer, with the airline fully knowing the customer’s out of pocket spend with the airline will be far higher. Just like at a car dealership!

‘Bleeding edge’ dynamic seat availability displays and pricing have been in beta test by one of the leading airlines since 2014, and in a few documented cases got loose into the commercial environment, creating immediate push-back. Loyal customers reported trip pricing quoted higher when booked under their airmiles number, as opposed to when booked anonymously. An airline that knows that a customer has regularly and willingly paid up in the fare structure, may utilize that information to set the threshold fares displayed.

In 2017, the airline went a step further, publicly eliminating the historically fixed, but periodically devalued, airmiles travel award redemption chart. The airline now offers 100% dynamically priced loyalty reward redemptions – mirroring availability and pricing displays and linked to the values thereof for revenue travel. With no award chart to compare, and with trip-specific terms and conditions, it is difficult for customers to determine whether redemption is at a premium or discount to traditional chart values.

What this points to is that airlines hold ‘big data’ – more than a 30-year purchase and travel history for customers who joined loyalty programs in the early years -- on individuals’ travel propensities and willingness to pay, for both airfare and more recently, for ancillary services. Setting offer prices based on an individual’s demonstrated willingness to pay is effectively an individualized EMSR, an individualized ‘ask’ price.

There remain real AI opportunities on the airline front-end. Not only for harvesting additional revenue from an airline’s known, loyal customers, but also on the ‘prospecting’ and ‘switching’ side – ‘find me more customers just like that in the wild’ and ‘find me customers just like that who are presently loyal to a competing airline’. AI based marketing and conversion campaigns are something the airlines should actively be looking at.

Optimizing Revenue on Payload-Limited Operations

In attempting to optimize the efficiency and economics of aircraft operation, airlines routinely ‘stretch the legs’ and utilize payload/range performance capabilities of aircraft to the limit.

On routes that utilize all of an aircraft’s available payload, there is an emerging opportunity to evaluate the competitiveness of various customer and revenue types that may consume mission-limited payload, including what payload to offload when additional fuel mass is required for longer duration seasonal flight routings, such as westbound in winter in the northern hemisphere. 

Should the last 10,000 pounds of available payload or container position loaded (or first container position or 10,000 lbs off-loaded to accommodate an additional 10,000 lbs of trip fuel) be passengers at 250 pounds each with bags, or should it be 10,000 pounds of high-value standby cargo?

There is an opportunity for the airlines to utilize big data and AI to optimize onboard revenue during the booking process, continually, during the 330 days a flight is resident in the inventory and reservations system, right up to realtime on flight-date.

An airline can forecast both passenger and cargo demand, comparing each to historical booking profiles, and integrating the mass, volumetric requirements (aircraft operated at the fringe of payload/range charts are often both ‘space and weight’ limited) and revenue value of cargo to the revenue management equation. This effectively expands the scope of revenue management to incorporate and maximize what is today the 6-11 percent (as much as 20 percent on wide-body international operations, much less on bulk-loaded domestic narrow-body operations) of overall combination airline revenue that cargo represents.

Mass and volume overall are new dimensions in which to maximize revenue aggregation and optimize revenue per unit of capacity (beyond seat-trip, to per pound/cubic foot of available payload). This may eventually drive fares to incorporate and be based on passenger mass, all-up, bags included.

The Economics of Unbundling and Ancillary Fees

Discriminant pricing or product-price differentiation is the underpinning of revenue management. Unbundling ancillary services takes this to a new level and dimension. Unbundling enables discriminant pricing on service attributes.

As background, the maximum amount that a person is willing to spend on a product or service, for example, travel from A to B on terms quoted on date C purchased D days in advance of departure, is called that person's reservation price. That person's reservation price can also be represented as a probability distribution (the probability of a customer purchasing travel at a price and set of terms).

If travel is sold at a single price P, all potential customers with a reservation price less than P will not buy the travel (resulting in demand destruction) and the airline also fails to fully realize the revenue opportunity from all customers with a reservation price higher than P, who would have willingly paid more.

Thinking out loud about the demand curve for travel A-B on terms C and D, the total area under the curve is the maximum amount of revenue that can be realized if each unique, potential customer is fully exploited at respective reservation prices.

Achieving this would require a nearly infinite number of prices, and is impractical. By contrast, if a single price, P was chosen, with the associated number of units sold being N, the area within a rectangular area (P, N) represents the revenue that would be realized. For any given single price, the area within the single rectangle is significantly less than the total area under the demand curve. 

In revenue management, numerous price points are offered to sell the same service, differentiated on terms of sale and use. These different prices are rationalized by differences in associated terms and services that the customer receives and purchases under, for example, assumption of capacity risk – non-refundability and non-transferability. This is called multiple discriminant pricing or product-price differentiation. 

By summing up proceeds from demand (N1, N2, N3, ...) at multiple prices (P1, P2, P3, ...) on different terms and with different features, filling in more boxes, so to speak, more of the area underneath the demand curve is captured, resulting in greater revenue realization. Customers are differentiated by the price they will pay for the service. 

The same idea applies to unbundling. Unbundling theoretically, at least lowers costs by removing variable service costs, can economically shift the demand curve to a lower price point yielding higher total quantity demanded, and also creates via an added services dimension the potential to realize yet further revenue from follow-on sales of ancillary services desired by travelers. 

Unbundling creates an additional dimension of sales opportunities, checks more boxes under the demand curve and realizes more of the total potential revenue, via an added dimension of product-price differentiation.

In 2018, U.S. airlines earned $17 billion and airlines globally earned $65 billion in ancillary service revenue not including the sale of frequent flyer miles to credit card issuers and others, consisting of the amenities consumers added to their ticket purchases, including fees paid for checked and carry-on baggage, assigned seats, early boarding, buy-on-board meals, inflight connectivity and entertainment.

The airlines should start using big data historical traveler purchase information and AI tools to merchandise improved seating availability such as in the ‘Premium Economy’ virtual cabin and other ancillary services in the days, weeks or months between booking and boarding, attempting to harvest more and maximize non-fare revenue.

Others are utilizing their flight booking profiles across markets to promote, offer, and in some cases even pay to move low-fare travelers who purchased non-refundable tickets, whose plans have changed since they booked their trip months prior, from their now heavily-booked flight to an underperforming flight, earlier or later in the day.

By so doing, the airlines can open-up sold-out inventory to accommodate higher-fare, close-in business traveler bookings. This improves network revenue and both travelers’ satisfaction. Likewise, many of the travel waivers offered during stormy weather fall into this category, with airlines realizing that many non-refundable fare purchasers as well as must travel business customers may wish to change flights.