Dr Philip D. Mann, PMP, PMI-RMP, is Principal Consultant at Vector Strategic Consulting LLC and author of The SCAR Framework: A Systematic Approach to AI Decision-Making in Critical Systems shares his view.
Sometime in September, the FAA intends to switch on SMART in live airspace for the first time. The Strategic Management of Airspace, Routes, and Trajectories system will begin as what FAA senior certification adviser Steve Fulton called a “bounded introduction”, limited to en route traffic at 24,000ft (7,315m) and above. The trade press has asked whether it will work. The better question is who answers when it does not.
SMART is one half of a 12-year, $875m contract awarded on 22 June to Air Space Intelligence, a Boston software company that beat Palantir and Thales. The other half is Flow Management Data and Services, or FMDS, which will replace the Traffic Flow Management System that has run the FAA’s Air Traffic Control System Command Center in Warrenton, Virginia, for two decades. FMDS is the backbone. SMART is the predictive layer above it, ingesting airline schedules, weather, airport capacity, and airspace constraints to flag congestion before aircraft depart. It operates on two timescales that should never be confused. The strategic function looks days, weeks, and months ahead to shape schedules and capacity; the tactical function works the same day, deconflicting projected trajectories and updating arrival times while aircraft are en route.
The FAA’s assurance on authority has been consistent. Fulton told the FAA and EASA safety conference in June that people, not software, will continue to separate aircraft. Nick Daniels, president of the US controllers’ union NATCA, said in April that “humans will separate airplanes”. That is the correct design intent, and it is also where the governance problem begins.
A prediction is never the last link in the chain. At the Command Center, a forecast bottleneck becomes a traffic management initiative (a ground delay programme, a reroute, a miles-in-trail restriction). That initiative passes to the traffic management unit at an en route centre, then to the sector controller who works the aircraft arriving under it. Suppose a strategic forecast overstates Thursday’s demand into Chicago Center: a ground delay programme goes out, aircraft hold at gates from Boston to Atlanta, and the sector controller who inherits the compressed push never sees the forecast that caused it. The machine’s judgement has shaped the picture, and no one has transferred any authority.
The seam nobody owns
In 1997 Raja Parasuraman and Victor Riley showed that operators calibrate trust to automation imperfectly, and that miscalibrated trust produces overreliance as readily as neglect. Nothing since has overturned it. A recommendation that is wrong, late, or built on stale schedule data will not announce itself. It becomes the plan, and the plan flows downstream.
Two questions follow, and the SCAR framework (Safety, Complexity, Accountability, Resilience), this author’s structure for AI decisions in critical systems, frames them. Under Accountability: when a prediction and a controller’s judgement diverge, whose call stands, where is that written down, and who answers when the outcome is a delay cascade or worse? Under Resilience: what does the system do when its inputs degrade? A predictive layer that fails gracefully hands the traffic manager a clean picture and says so. One that fails brittly keeps recommending, with confidence. The FAA has not yet published which behaviour SMART will exhibit, or what evidence beyond vendor demonstrations shows the answer holds on a July afternoon with convective weather stacked across the Ohio Valley.
Fast timetables defer hard questions
The timing matters because of the money. On 17 July, Administrator Bryan Bedford told Reuters the agency needs a further $10bn from Congress beyond the $12.5 bn already appropriated, with traffic expected to double within two decades, and he promised a fast timetable. Fast timetables earn the next appropriation. They are also the condition under which accountability questions get labelled “later”. NextGen spent two decades and tens of billions of dollars promising much of what SMART now promises; the lesson was that governance lagged deployment.
EASA’s AI Roadmap 2.0 (2023) draws the line differently. It grades applications by the level of human oversight required and demands assurance evidence to match. The FAA’s bounded introduction is compatible with that approach only if the bound is set by demonstrated performance rather than by the calendar.
What good would look like
Three pieces of evidence would close most of the gap. Validation results under representative workload and degraded data, rather than demonstration conditions, should be published before the scope expands below 24,000ft. A written decision-authority document, visible to traffic managers and controllers, should state what happens when SMART and human judgement disagree and who owns the result. Expansion gates belong to measured performance at each phase, not to a funding cycle. None of this is hostile to SMART. It is the difference between a decision-support tool the workforce trusts and one it quietly learns to ignore.
