From days to minutes: investigating deadheads with an AI-enabled planning workflow
VeeOne used an independently developed AI-enabled workflow within a real Jeppesen crew-planning environment to investigate deadheads, test multiple scenarios and produce a decision-ready report in roughly 30 minutes.
Could we reduce deadheads — and what would it cost?
It is a familiar question for airline planning teams.
Deadheading is sometimes unavoidable, but when the number appears high, planners need to understand why. Could the optimizer produce fewer deadheads? Would that increase duty days or weaken roster quality? Could overly restrictive rules be forcing more deadheads than necessary?
A reliable answer requires more than changing one cost and running the optimizer again. An experienced analyst must examine the current plan, study the cost model and ruleset, form hypotheses, configure several scenarios, wait for the runs to complete, compare the results, explain the trade-offs, create visualisations and document their findings. For an investigation like this, the turnaround for an experienced analyst can be two to three days, which limits how many investigations airlines can undertake.
VeeOne tested whether an AI-enabled planning agent could make that investigation faster and more practical — while keeping an experienced analyst in control.
A real planning environment, not a theoretical exercise
VeeOne applied an independently developed AI-enabled workflow within a Jeppesen crew-planning environment used as part of its planning-as-a-service delivery for an airline customer. The objective was to test the approach against a real planning problem, with real plans, rules, costs and operational trade-offs.
The workflow covered four stages: diagnosis, hypothesis generation, scenario execution and reporting.
1. Diagnosing the plan
The agent began by examining recent crew plans and their key performance indicators. It then reviewed the optimizer’s cost model, relevant cost parameters, the planning ruleset and the rule buffers that could be contributing to deadheads. From this material, it produced an initial assessment of the factors most likely to be influencing the optimizer’s choices.
This was the first human approval gate. An experienced analyst reviewed the assessment, checked that it made operational sense and approved it before the investigation continued.
2. Turning hypotheses into scenarios
Based on the approved assessment, the agent proposed a set of optimization experiments.
In this case, the proposals included testing a different balance between deadhead and duty-day costs, relaxing selected planning rules and adjusting associated rule buffers.
The purpose was not simply to find the scenario with the lowest deadhead count. It was to expose the trade-offs: what improved, what deteriorated and which constraints and rules were shaping the result.
The analyst reviewed the proposed experiments and decided which scenarios were appropriate to run. Only the approved scenarios proceeded.
3. Running the experiments
The agent submitted the approved optimization scenarios with their corresponding parameters. Once the jobs were running, the analyst did not need to monitor each one manually. The workflow tracked their completion and moved to analysis when the results became available.
4. Comparing and explaining the results
When the optimization jobs finished, the agent collected the outputs and compared deadheads with related planning KPIs. It identified the trade-offs between scenarios, produced suitable visualisations and generated an HTML report containing its conclusions and recommendations. The report required no manual editing, although the analyst remained responsible for sense-checking the findings before using them in customer or stakeholder discussions.
What the investigation revealed
The analysis produced two findings with direct operational relevance.
First, it identified a particular planning rule that was materially driving deadheads. This gave VeeOne concrete evidence with which to raise the rule and its buffer settings in discussions with the airline customer. In the end, we decided to reduce the buffer setting slightly because the tested scenarios showed a significant reduction in deadheads and improved productivity.
Second, the investigation identified a better balance between the optimizer’s deadhead and duty-day cost settings. The resulting recommendation was applied when configuring the next production run.
The value was not that AI found something an experienced analyst could never have discovered. The value was that it made a thorough, multi-scenario investigation fast and practical enough to perform within normal planning work.
From days to minutes
Roughly 30 minutes from initial analysis to generated report
Around 20 minutes of that time was spent waiting for the optimizer.
The agent compressed an investigation that would typically require days of analyst work and turnaround into a workflow completed in roughly 30 minutes, with most of that time spent waiting for the optimizer. More importantly, it reduced the manual workload between the question and the answer.
From our experience, planning teams rarely suffer from a shortage of useful questions. They suffer from limited time to investigate them properly. Faster analysis means more questions can be examined despite constrained analyst bandwidth. With the help of AI agents, VeeOne can respond more quickly when its customers ask about deadheads, productivity, roster quality or unexpected characteristics in a plan.
It also becomes easier to react when the planning environment changes: a new network structure, revised operational requirements, different contractual constraints or an updated planning rule.
Instead of relying on assumptions or one isolated test run, planners can compare evidence from a structured set of scenarios. That evidence can help explain why trips or rosters look a particular way to operations leaders, chief pilots, unions and other stakeholders.
The same approach could eventually support proactive checks before a plan is published. Unusual results could be identified, investigated and explained before they generate questions from crew or operational teams.
Human judgement remains central
The workflow was not an exercise in autonomous production planning. The analyst approved both the initial assessment and the proposed hypotheses and scenarios. The agent then handled the repetitive work of execution, comparison, visualisation and report generation. The analyst retained responsibility for validating the conclusions, deciding whether recommendations should be applied and determining how findings should be communicated.
This division of work is central to the approach. AI increases the analyst’s investigation capacity; it does not remove the need for planning expertise or accountability.
What comes next
VeeOne is now investigating a more iterative form of agent-led experimentation.
In a future closed-loop workflow, the agent could examine the first batch of results, formulate follow-up hypotheses and propose or launch additional scenarios. The process could continue until it reaches a defined conclusion, target, stopping condition or decision requiring human input.
That capability remains future work. The completed investigation described here used an analyst-approved batch of scenarios with clear human review points.
Explore what this could mean for your planning team
If your planning team has important questions that remain unexplored because investigations take too long, contact VeeOne to discuss where AI-enabled analysis could improve your planning process.