Agentic AI for Automated Absence Reporting and Shift Management at Airport Operations

manchester_airports_group · Other · 2024 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.2 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Airport operations

Manchester Airports Group needed to standardize absence reporting and shift-management workflows across airports while reducing operational cost.

Constraints:

Passenger safety and security
Different airport processes
Low latency
HR and rostering integration

Architecture

MAG's absence-reporting system wraps airport HR and rostering workflows in an agentic Bedrock AgentCore architecture. Employees report absences by text or speech; the text agent authenticates the employee, classifies absence type, asks follow-up questions when information is missing, checks HR policies, updates HR and shift records, and notifies managers. Tools are exposed through MCP and adapted for agent readability, while custom context-aware guardrails protect a safety-critical airport workflow. The speech-to-speech path delegates sensitive tool execution to the text agent rather than giving the speech model direct tool access.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  employee-interface[Employee interface] --> text-agent[Text agent]
  employee-interface --> speech-to-speech-model[Speech-to-speech model]
  speech-to-speech-model --> text-agent
  text-agent --> context-aware-guardrail[Context-aware guardrail]
  text-agent --> mcp-tool-server[MCP tool server]
  mcp-tool-server --> hr-system[HR system]
  mcp-tool-server --> rostering-system[Rostering system]
  mcp-tool-server --> policy-knowledge[Policy knowledge]
  mcp-tool-server --> manager-notification[Manager notification]

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Employee interface`	Text or speech channel for reporting absence.	`Text agent`, `Speech-to-speech model`
`Speech-to-speech model`	Voice interface that gathers speech input but delegates tool execution to the text agent.	`Employee interface`, `Text agent`
`Text agent`	Primary agent for authentication, dialogue, classification, and tool orchestration.	`Employee interface`, `Speech-to-speech model`, `MCP tool server`, `Context-aware guardrail`
`Context-aware guardrail`	Custom guardrail with conversation context, built around Lambda and LLM checks.	`Text agent`
`MCP tool server`	Tool exposure layer for HR, rostering, policy, and notification operations.	`Text agent`, `HR system`, `Rostering system`, `Policy knowledge`, `Manager notification`
`HR system`	Records absence and employee information.	`MCP tool server`
`Rostering system`	Updates shift records and supports re-rostering.	`MCP tool server`
`Policy knowledge`	HR policy context for different absence types.	`MCP tool server`
`Manager notification`	Notifies line and rostering managers about the absence.	`MCP tool server`

Data flow

An employee reports an absence through a text or speech interface. The agent authenticates the employee, gathers missing details conversationally, looks up HR policy, classifies the absence, records it in HR and rostering systems, notifies the line and rostering managers, and uses context-aware guardrails to keep the tool chain bounded.

3 supporting quotes

“So for example, we can connect to an API, we can connect to an HR rostering system with an API call.”transcript, stated

“We can look at HR policies with a knowledge base, we can classify absences through an LLM and we can notify a manager through Amazon SNS our messaging solution.”transcript, stated

“In this case, the user is a rostered user, so we need to record both the absence into an HR system and the rostering system and notify both their manager and their rostering manager.”transcript, stated

Why this architecture

MAG needed an agentic conversation because real employees provide incomplete or ambiguous absence reports, and deterministic tools cannot proceed without missing details. The architecture keeps speech convenient while preserving the richer text-agent tool chain, and custom guardrails address airport safety, security, and misrecording risk beyond generic moderation.

3 supporting quotes

“So if an employee comes in and gives incomplete information, then our tool doesn't work already because it can't quite record the full absence.”transcript, stated

“We need a natural conversational pattern that is able to collect all the relevant information from the employee.”transcript, stated

“through and it led us to our second lesson. You have to think broader than standard guardrails. Write custom guardrails for compliant and responsible AI. Current guardrails only look at the last message that's the person has written. When you think about guardrailing, of course the first”transcript, stated

Models used: LLM absence classifier · speech-to-speech model

Scale

The deployed workflow reached 99% consistency in absence reporting.
“Well, we were managed to standardize the way the absences are recorded in the system and we now have a 99% consistency in absence reporting. This is unleashing a lot of potential into analyzing what the patterns in the absences are downstream.”transcript, stated
MAG reduced time for recording absences by 90%.
“And we also reduce the time it takes for recording this absences by 90%. But I'm gonna let Steve tell you more about what's in stock with MAG for the future.”transcript, stated

Alternatives considered (roads not taken)

Deterministic workflow connected to the same tools

Rejected because: Incomplete or ambiguous employee input prevents the deterministic path from recording the absence.

“This is why we really need an Agentic workflow here. We need a natural conversational pattern that is able to collect all the relevant information from the employee.”transcript, stated

Let speech-to-speech call tools directly

Rejected because: MAG kept tool problems in the text agent and used speech primarily as an input/output interface.

“So instead of connecting Amazon Nova Sonic directly to the tools, we connected Amazon Nova Sonic to the text agent that we had already built.”transcript, stated

Patterns

`multi-agent-orchestration` (operations)

The absence workflow coordinates multiple agent-powered steps across airport operations systems.

“So our vision is to create a single unified experience for our colleagues, powered by agents removing a significant”transcript, stated

`safety-policy-enforcement` (operations)

Airport automation is bounded by security, accuracy, and reliability requirements.

“And it's an environment where passenger safety and security is non-negotiable. So MAG had to get started on this journey”transcript, stated

Tradeoffs

Use an agentic conversation for absence capture.

Chose A natural conversational agent that collects missing information over A deterministic workflow wired to the same HR, policy, and notification tools.

Rationale: The deterministic path could not complete when real users gave incomplete or messy absence reports.

“We need a natural conversational pattern that is able to collect all the relevant information from the employee. Now we can connect this agent to all of these tools”transcript, stated

Expose progress feedback while the agent performs tool chains.

Chose UI feedback that explains what the agent is doing behind the scenes over A silent backend workflow that appears frozen during tool latency.

Rationale: Even correct agentic execution felt too slow to stakeholders when the interface did not keep users engaged.

“You can have the most amazing backend that can do all the most amazing tasks. But Agentic still requires an agent to go through a set number of tools, which actually will always take some latency”transcript, stated

How to support voice reporting without weakening controls.

Chose speech-to-speech interface delegated to the text agent over direct speech-model tool execution.

Rationale: The text agent already held the tool instructions, authentication pattern, and guardrail behavior needed for sensitive HR updates.

“the text model that then calls all the tools.”transcript, stated

Gotchas (what bit them)

A deterministic absence workflow failed when employees gave incomplete information.

Root cause: The workflow needed conversational recovery to collect missing absence details before updating HR and rostering systems.

Resolution: MAG used an agentic conversational flow that could ask for the missing information and then continue tool execution.

“So if an employee comes in and gives incomplete information, then our tool doesn't work already because it can't quite record the full absence. This is why we really need an Agentic workflow here.”transcript, stated

Standard guardrails were insufficient for a safety-critical airport workflow.

Root cause: Social engineering or misrecorded absences could affect checkpoints, flight systems, pay, and critical infrastructure.

Resolution: MAG added context-aware guardrails and least-privilege tool access around the multi-turn agent conversation.

“what stops someone from social engineering the agent into doing something malicious? We're like, yeah, that's a great question. Well of course we would add guardrails.”transcript, stated

“If an absence is misrecorded, this means issues with like checkpoints, flight systems, this has like high impact.”transcript, stated

Evaluation

MAG measured consistency, recording-time reduction, security, accuracy, reliability, and latency.

Metrics: consistency · recording time · latency

“that we have the right level of security, accuracy of the solution and reliability. So you can imagine”transcript, stated

Outcomes

The summary reports 99% consistency and 90% recording-time reduction.

The system supports both text-to-text and speech-to-speech absence reporting.

“The solution achieved 99% consistency in absence reporting (standardizing previously variable processes) and reduced recording time by 90%, with measurable cost reductions in overtime payments and third-party service fees.”summary, stated