Your AI Pilots Won't Move the P&L. Here's What Will - Adam Nahari, Apollo

When an aircraft at a large cargo airline experiences a failure, a crew scheduler has 15 minutes to work out who can fly. The scheduler has to check where each pilot is, what they’re qualified to fly, and whether they’ve had enough rest. Finding one legal assignment by hand can consume the entire window. Even then, the scheduler has little basis for knowing whether another assignment would cost less.

Often, the scheduler finds no legal option in time and turns to premium pay to cover the flight. That decision recurs every day across the network, in a business where crew is the second-largest cost after fuel. For Adam Nahari of Apollo Global Management, the opportunity was in the cost of the assignment the scheduler could find before the deadline. The minutes spent searching were only part of the story.

Nahari puts the airline’s experience alongside a gap in reported AI adoption and financial results. He cites figures showing that 55% of U.S. adults had used generative AI within three years, compared with 20% for the PC and 30% for the internet at comparable points in adoption. Almost 90% of businesses reported using AI in at least one function, yet only 6% reported a significant impact on earnings before interest and taxes (EBIT). Reported time savings added up to about 2% of all U.S. working hours. Little of that saving was visible on income statements.

Nahari distinguishes three places a business can use AI. General-purpose copilots and chatbots can make people more productive, but competitors have access to the same tools. In his view, competition tends to pass those gains on to customers through prices. A second layer draws on company knowledge. A function-specific copilot or a retrieval system over internal documents helps employees find and use information that belongs to the business.

The third layer reaches into core operations, such as deciding which pilots to assign to a flight. A competitor can buy access to a model, but it still has to connect its own operating data and change how its schedulers work. Nahari says businesses reporting material EBIT impact were three times as likely to redesign a workflow end to end. His own rule of thumb, drawn from transformations he has led or observed, is five units of change management for every unit of engineering. Much of that effort goes into choosing the problem, getting people to use the system, and measuring what changes.

Nahari’s first move is to choose one or two problems with a plausible path to a financial result. Starting with EBITDA, he breaks the airline into four operating drivers. Routes represent demand, aircraft represent supply, maintenance constrains what can fly, and crew drives cost. The size of the crew expense justified concentrating on a single scheduling initiative.

He screens a use case with five questions. Which financial line would the CFO expect to move? Can the team agree on a baseline before work begins? Does the company own the data, and is its infrastructure ready? Is there a specific person who wants the tool and will champion it? Finally, does the problem rank among the CEO’s top five priorities? At the airline, those questions connected a recurring decision at one desk to a cost the finance team could track.

The second move was to redesign the search itself. A faster version of the manual workflow might help a scheduler check a candidate more quickly, while still leaving them to decide whom to check next and when to stop looking. Nahari calls that the “10% play” and the redesign the “10x play.” He presents those numbers as illustrative.

The team consolidated flight, crew, and maintenance data and built an optimization engine that applied the constraints across possible assignments. It generated the full set of feasible solutions, ranked them by cost, and presented them in a decision-support app. Nahari reports that a scheduler who previously took 15 minutes to find one answer could now see the ranked options in under a minute. The measured comparison he gives is search time. He supplies no figure for realized payroll savings or margin improvement.

Nahari’s third move starts before funding. He funds a project only if the CEO has publicly committed to it as a company priority. Without it, an initiative can stall without anyone explicitly rejecting it. Meetings slip, data requests take weeks, and pilot dates move. The P&L owner has to assign people, the CFO has to fund the work and track its effect, and the CTO has to reserve engineering capacity. Those commitments need renewing as other demands compete for attention.

The fourth move gives the business a hand in building the system. The airline’s VP of crew scheduling co-sponsored the project. Working schedulers joined the design effort from day one, met with engineers every other day, helped design the interface, and participated in A/B tests. When their colleagues began using the app, the people introducing it could explain decisions they had helped make.

The airline also hired a chief information officer, a VP of data and AI, and a senior director of data science. The recruits brought airline experience alongside their technical expertise, so they understood the conditions under which a scheduler had to make a decision. In Nahari’s account, the organizational investment exceeded the technology investment.

At a workshop Nahari attended, about 30 executives were asked how they might redesign their businesses with AI. Three out of four had never used an AI tool before that week. His fifth move is to put tools into the hands of the people expected to make those choices. Trying them on familiar work gives decision-makers a basis for judging which steps a system could take over and where an opportunity might justify redesigning a whole process.

The sixth move was to write down the conditions for continuing the pilot and let the CFO keep score. The airline gave each stage 90 days to demonstrate value. At each stage, the team used A/B testing against a baseline chosen in advance. It also set a separate adoption threshold requiring more than 50% of crew schedulers to use the tool. That denominator matters. A few schedulers using the app heavily could generate plenty of activity while most of the desk continued to work manually.

The CFO and the project team also estimated the cost of operating at scale, including compute, tokens, and the people doing the work. That total had to be significantly lower than the financial impact they expected. The ranked search had already answered an engineering question: the desk could compare feasible assignments before its deadline. The adoption test and the CFO’s calculation addressed the remaining business question of whether enough schedulers would act on those options to justify the cost.