Conversational AI

Designing for AI conversations for voice and text UI.

Designing conversational AI experiences — whether chat or voice — requires careful attention to interaction flows and user expectations. The following guidelines support product teams in creating trustworthy and intuitive conversational experiences that prioritize usability, compliance, and ethical standards.

Use this page as implementation guidance for the experimental AI UX Playbook, particularly its principles and guardrails for disclosure, human involvement, safe failure, escalation, and consistent AI identity.

General considerations

Consider the following aspect when designing and working on conversational AI.

  • Frame the conversation with the context and purpose. Define what problems the AI solves and in which logistics contexts (shipment tracking, inventory updates, customer support calls, etc.). This will help you frame the conversation around the domain.
  • Always provide the user with control over the conversation. Consider how easily users can override, clarify, or escalate conversations from AI to human support.
  • Indicate security and privacy. Show the user that the AI will securely handle sensitive logistics data and communicate its data handling practices transparently.
  • Make sure the AI decisions and output are explainable. Users should be able to understand how and why the AI arrived at specific recommendations or responses. A transcript records the interaction, but an explanation should also communicate the relevant sources, assumptions, limitations, and uncertainty.
  • Gracefully handle error management with a respectful tone. AI should handle misunderstandings, errors, or incomplete user inputs in a graceful and non-offensive tone and voice.

Conversational interactions

Whether chat or voice, conversational interactions must feel natural, clear, and responsive. Conversational interfaces like chat and voice introduce affordance challenges and require strict management of the user expectations. Users must know what they can do and can’t do in order to avoid bad UX and lose of trust in the AI solutions.

Align capability expectations

  • Introduce AI capabilities and scope at the start of the conversation.
  • Keep AI identity visible throughout chat experiences. For voice, disclose AI use at the start of the call and repeat it at every agent handoff or topic change.
  • Indicate the AI limitations to establish what users can and cannot do.

Example:

“Hello! I’m Maersk’s virtual assistant. I can track your shipments, check container status, and connect you with support if needed. How can I assist?”

Ensure contextual awareness and retention

  • AI should retain up to date context, minimising repetitive user inputs.
  • Reference previous user interactions within the conversation.
  • Aim to acquire information from the user by asking questions if context is not enough for a high-accuracy response.

Example:

User asks: “Where’s shipment XYZ123?” and then “When will it arrive?”

AI response: “Shipment XYZ123 is at the Port of Rotterdam. It is expected in Hamburg on June 28, 16:00 CET.”

Design clear communication

  • Avoid technical jargon. Use plain, actionable language.
  • Avoid long responses unless specifically requested by the user (e.g. ‘provide detailed report’)
  • For voice interactions, keep sentences short and simple.
  • Avoid humour and simulating human emotions.
  • Use natural conversational language without implying that the AI is human or experiences emotions. Avoid system commands like “Confirm the new shipment number by saying ‘Confirm’.”

Example:

AI: “Shipment ABC123 arrives tomorrow at 10 AM. Can you confirm that?”

If user says no, then ask follow-up questions

AI: “OK. Can you then please tell me again when it arrives?”

Ensure secure and ethical data handling

  • Request consent where personalised features or sensitive data processing require it, following Maersk privacy and data ethics guidance.
  • If handling users’ information, communicate clearly that it is secured.

Example:

“To verify your account securely, please provide your registered phone number or email.”

Design for edge cases and recovery

Clarification of intent before answering

  • If the AI doesn’t fully understand the user’s request, it should ask for clarification rather than guessing or returning a generic “no result.”

Example:

“I’m not sure how to help with that. Could you try rephrasing your question?”

Unable to answer

  • When a request falls outside the AI’s capabilities or domain, clearly communicate that, and if possible, offer an alternative.

Example:

“I’m sorry, I can’t assist with that as its not in my scope. You might try this (alternative suggestion) instead.”

Guardrail answer

  • When the user input is irrelevant, inappropriate, or clearly trying to trick the AI, respond with a polite redirection to the AI’s intended purpose.

Example:

“I am here to help you with [purpose of AI]. How can I assist you today?”

Conversational UI elements

Conversational UI is enhanced significantly through supporting visual or interactive elements, especially in logistics scenarios.

Text based UI elements (Chatbots, Search, Suggestions)

  • Use AI color for indicators and UI. (Coming soon)
  • Create clear visual separation of user inputs from AI responses.
  • Use typing indicators to signal when the AI is processing user requests.
  • Provide users with quick action buttons to access common tasks or responses (e.g., “Track Shipment,” “Contact Support”).
  • Allow users to explore AI recommendations or decisions further.
  • Show confidence when it affects a user decision. Use meaningful states tied to a defined threshold and next action rather than an uncalibrated percentage.
  • Provide a stop/cancel option in cases where AI response could take too long.

Example:

“Estimated arrival: June 28, 16:00 CET. Low confidence — review the latest port data. [View details]”

Voice-based UI elements

  • For voice interactions, visual UI supports should provide real-time feedback or follow-up interaction:
  • State that the user is speaking with AI at the start of the call and repeat the disclosure at every agent handoff or topic change.
  • Provide visual indicators showing when the AI listens or speaks.
  • Provide buffer pause before responding to a prompt to acount for slow speaking or short pauses.
  • Aim to provide real-time display of voice-to-text transcripts for improved accessibility and understanding.
  • For multimodal interfaces, allow easy confirmation or cancellation for critical actions.
  • Offer easily accessible buttons for escalating to human support during a voice interaction.

Scenarios and recommendations

Below are common scenarios illustrating conversational desired AI behaviors for Maersk logistics products.

ScenarioAI BehaviourExample
Shipment tracking (Chat)Contextually aware, concise.”Your shipment (ID 123456) is currently docked in Rotterdam. Expected in Hamburg by June 30.”
Inventory management alert (Chat)Action-oriented, concise, explanatory.”AI recommends restocking item #245 (current stock below minimum). [Approve] [Modify Order]“
Customer support call (Voice)Clear, short prompts; explicit confirmation for critical tasks.”You requested shipment cancellation for shipment 123456. Can you confirm this?.”
Misunderstanding (Voice)Graceful, patient error recovery.”I didn’t catch the shipment number. Could you please repeat it clearly?”
Data privacy prompt (Chat)Clear, consent-driven approach.”To personalize your shipment updates, I need your permission. [Allow] [Decline]”

See also