Comparison

Manual crew scheduling vs AI optimisation: what really changes

Manual scheduling searches for a roster that works; algorithmic optimisation compares a very large number of rosters that work and keeps the best one against explicit objectives. The difference lies less in the planner's skill than in how many combinations a method lets you examine, and how quickly you can start over.

The debate is usually framed as human versus machine, which is the wrong frame. The useful one is division of labour: some of this work is combinatorial arithmetic, some of it is judgement. High-performing operators do not choose between them; they stop confusing the two.

What does a human planner do that an algorithm cannot?

Worth putting first, because vendor decks tend to skip it.

  • Context that was never written down. The crew member going through a hard month, the one who genuinely does not mind a last-minute night stop, the client who tolerates an hour's slip and the one who does not. None of it is in the system.
  • Negotiation. A roster is a social object as much as a plan. Persuading, trading, explaining an unpopular assignment: nobody delegates that to software.
  • Arbitration between conflicting objectives. Whether to favour stability or cost this month, whether to protect a base or honour a contract: that is company policy, not a calculation.
  • The exception. The case nobody has seen before, the constraint that appeared this morning, the solution outside the model.

An algorithm does none of these four. A tool that claims otherwise is overselling.

What does an algorithm do that a human cannot?

One thing, and it changes the economics of everything else: exhaustive exploration.

An experienced planner builds a good roster in a few days. What they cannot do is verify that some other arrangement of the same assignments would not have produced two fewer night stops and a fairer spread of weekends. Comparing complete solutions is out of human reach, not for want of skill but for volume: every assignment shifts that person's rolling 7-, 14- and 28-day counters, and therefore the feasibility of every assignment after it.

The FL3XX × SkAI Tech Crew Management Benchmark 2026 (scope: business aviation) puts a number on that difference in scale, using a deliberately tiny case. With 14 flights and 10 crew members, the number of possible rosters already reaches 1.4 × 10²³. A planner evaluating 1,000 rosters a second, a rate no human sustains, would need 321 times the age of the universe to work through them. No real operation is as small as that example (the full demonstration is in our article on the combinatorial explosion of rostering).

This is the figure that makes the comparison honest. What the manual method lacks is less speed than the ability to compare. An experienced planner produces a good solution; what they cannot know is how far it sits from the best one, because the question itself is out of reach without a machine.

Three practical consequences:

  1. Verification is exhaustive rather than sampled. Every assignment is checked against every rule (FTL, qualifications, in-house policy), not only the ones that came to mind at the time.
  2. Recomputation becomes cheap. Once the model exists, regenerating costs minutes, which makes it feasible to test several scenarios before publishing and to recover quickly after a disruption. The stakes are measurable: EUROCONTROL's CODA Digest reports that reactionary delay accounted for 46% of European delay minutes in 2024, or 8.0 minutes per flight. The optimisation literature also shows that integrated recovery, where flight and crew decisions are solved together rather than in sequence, outperforms the sequential approach on recovery cost (Petersen et al., 2012, Transportation Science). Our article on crew disruption management goes into this.
  3. Objectives become explicit. "Spread the weekends fairly" stops being an intention and becomes a measured, weighted, auditable criterion (see crew roster fairness).

How do the two methods compare in practice?

CriterionHand-builtAlgorithmic optimisation
Producing a monthly rosterDaysMinutes (12 on average, observed at SkAI Tech)
Solutions comparedOne, maybe twoA very large number
Regulatory verificationSampled, under time pressureExhaustive, on every assignment
Fairness across crewAn intention, hard to evidenceA measured, auditable objective
Recompute after disruptionHours, restarting the reasoningMinutes
Deviation between published and flown rosterBaseline-20%, observed at SkAI Tech
Traceability of decisionsIn the planner's headIn the configured rules and objectives
Direct flight coverageBaseline+21 percentage points
Duty days per pilot per monthBaseline+0.42 day (capacity already paid for)
Freelance duty daysBaselineup to -40%
Duty spread between crew membersBaseline-19%
Daily time spent on operationsBaseline-1h30 per day
Handling exceptionsExcellentWeak: this is the human's job
Negotiating with crewExcellentNone

The two SkAI Tech figures at the top of the table (12 minutes, -20%) are observed across 2025-2026 deployments, over a scope of more than 70 aircraft and 900 crew members. The five quantified rows after them are the deployment results published in the FL3XX × SkAI Tech Crew Management Benchmark 2026, at business aviation operators. All of them vary with fleet size and operating model: they do not transfer to regional, ACMI or medevac operations, and they describe somebody else's operation rather than yours.

Read honestly, the last two rows do not net off against the others. On their own they are why the planner stays at the centre of the process rather than beside it.

The real risk of staying fully manual: who owns the knowledge?

This is the least discussed and most expensive point. In an organisation that plans entirely by hand, knowledge of the rules, the exceptions, the habits and the tacit balances lives in one or two people. The roster works perfectly while they are there.

One resignation, one long absence, one holiday period, and the operator discovers that none of it was written down. Encoding rules into a tool does not remove the expertise; it writes the expertise down. That is a business continuity benefit, entirely separate from any optimisation gain.

Do you have to choose?

No. What works is a split, not a substitution:

  • To the algorithm: generating the full roster, exhaustive regulatory verification, balancing workload across crew, producing recovery scenarios after a disruption.
  • To the planner: setting objectives and their weights, choosing between scenarios, handling special cases, and owning the relationship with crew and clients.

That is what the crew planning manager at Air Antilles, a SkAI Tech customer, describes: “The time savings are significant. We can now focus on specific adjustments, manage disruption, and better anticipate training and recruitment needs.” The work does not disappear; it moves up a level.

The same pattern shows up outside aviation. At Swiss Federal Railways (SBB), automated disruption qualification relieves dispatchers in the operational centres of up to 15% of their working time (Detecon). And the 2025 AI trial across 90 UK NHS organisations, which covered administrative task automation rather than staff scheduling, returned an average of 43 minutes a day to more than 30,000 staff (Department of Health and Social Care / NHS England). In both cases the experts stayed; their time moved onto the decisions.

Still building every roster by hand? See what SkAI Tech leaves to crew schedulers, and what it takes off their hands.

How do you run a fair evaluation?

The most common mistake is judging optimisation on an empty demo month. A trial worth trusting has four elements:

  1. A real month, already flown. Replay a month you know, including the one that went badly. You already know the answer, which is what makes it a test.
  2. A baseline defined before starting. Generation hours, deviations, positioning legs, reserve call-outs. Defining these afterwards makes any conclusion unfalsifiable.
  3. Your actual rules, not the standard set. The regulatory base is the easy part. The test is whether your labour agreement and in-house policy can be configured, and by whom.
  4. A disruption scenario. Introduce a sick call on day 12 and time the recovery. This is the number your team will live with every day, and it is never the number on the slide.

What actually changes for the team?

Three changes, in this order:

  1. The rules have to be written down. The most demanding phase, and the most useful: putting in writing what the team applies by habit. It routinely surfaces contradictory rules nobody had ever articulated.
  2. The working rhythm shifts. Less time on construction, more on scenario analysis and anticipation: training needs, recruitment, seasonal peaks.
  3. The conversation changes. You stop debating whether a roster is feasible and start debating its objectives: how much cost is stability worth, what fairness are you aiming for. A harder conversation, and a far more productive one.

None of this requires replacing the operations system: an API add-on reads the data already there and pushes rosters back, with no migration and no workflow change.

FAQ

Does optimisation cut planner headcount?

That is not what SkAI Tech customers describe. The load moves: less construction and checking, more disruption handling, analysis and forward planning. In a market where crew planning skills are scarce, the problem is usually retaining those people, not reducing them.

Can you trust a roster you did not build yourself?

Trust comes from traceability, not authorship. A serious tool shows which rules were applied and why an assignment was selected. That is more auditable than a mental process which, by construction, leaves no record.

What if our in-house rules are written nowhere?

That is the norm, and it is the first task of the project: formalising the essentials, the regulatory rules being shipped with the tool. The formalisation has value on its own, whatever you then decide about software.

How long before results show?

Generation time is visible on the first complete roster. Structural gains (deviations, recovery costs) take several months and need a baseline defined before you start. Defining that baseline up front is the only way to be able to conclude anything later.

Sources

  • FL3XX × SkAI Tech (2026, August), Crew Management & Business Aviation Operations — 2026 Benchmark Report: combinatorial explosion and 2025-2026 deployment results, business aviation scope, varying with fleet size and operating model.
  • SkAI Tech indicators (first-party data): average generation time and reduction in roster deviations observed across 2025-2026 deployments, over more than 70 aircraft and 900 crew members.
  • Customer testimonial published on the SkAI Tech site (Air Antilles).
  • EUROCONTROL (2025, 25 July), All-Causes Delays to Air Transport in Europe — Annual 2024 (CODA Digest), eurocontrol.int.
  • Petersen, J. D., Sölveling, G., Clarke, J.-P., Johnson, E. L. & Shebalov, S. (2012), An Optimization Approach to Airline Integrated Recovery, Transportation Science 46(4), 482-500, DOI 10.1287/trsc.1120.0414.
  • Detecon, New data analytics software for SBB (case study), detecon.com.
  • Department of Health and Social Care / NHS England (2025), Major NHS AI trial delivers unprecedented time and cost savings, gov.uk.

Whitepaper

The 2026 Crew Management Benchmark

Transforming crew management with intelligent automation, cost efficiency and interconnected systems for airline rostering and reporting. Free access, in exchange for your details.

Read the report opens in a new tab
Cover of the 2026 Benchmark Report, published by FL3XX and SkAI Tech

See the Future of Crew Management Today

Discover in 30 minutes how SkAI Tech can elevate your crew operations into a strategic advantage.

Book a demo