Conversational design¶

Status: Current | Last reviewed: 2026-10-06

Use Design with AI to create and refine a Voice, Chat, Copilot, Translation or Background agent, or a multi-agent demo, through conversation. Your conversation, emerging design and working draft are saved in your workspace. This feature requires authoring access and an enabled Generate with AI connection.

The design entry point must be enabled for your deployment. In the left sidebar, select Design with AI below Agents; in a collapsed sidebar, hover the lightning icon to see its label. If it is absent, your administrator should check deployment enablement and your authoring permissions. If the page opens but cannot start, check the Generate with AI connection and authoring service. Standalone designer exports display the same screens with simulated responses, as described below.

Home and navigation¶

Home is an operational dashboard in the teal Deloitte identity. It shows repository counts, captured recent activity, your generation notices, your agent portfolio and favorite agents. Search matches names, descriptions, IDs, tags and agent types. Select a result to open its editor, or submit the search to see the full repository.

Select Design with AI to start or resume a design conversation. Insights is deferred and removed from navigation; old links return to Home. The sidebar expands on browse screens and collapses in editors.

In Vanguard, counts and activity come from the running app. In a standalone designer export, they use captured data and local changes. Notifications contain workspace events, not a rollup of every agent's drafts or errors. Recent activity starts with five entries; use Show more to reveal the remaining loaded entries and Show less to return to the compact feed.

Agent repository and notifications¶

Agents, Analytics and Design with AI share the same page header, with a teal title and the global notification bell on the right. The sidebar has no duplicate notification entry. Home keeps its greeting and uses its Notifications widget without a separate header bell.

Agent cards show their type, status, version and last-modified date. Unpublished Changes means a published agent also has a draft. View Details opens the side drawer; use its menu for Duplicate, Export and Delete. Edit tags in the agent’s canvas Settings with Save tags. Duplicate lets you choose an agent source version and change its name and description. Workflow code export is available when enabled. Workflow activation is in the workflow editor’s Settings.

Client logos appear in the drawer only and come from active Branding linked to voice/chat agents. Workflows do not have Branding or client logos. Agents without a linked logo show no client logo. Changing an agent name never assigns a client automatically.

Home notification actions are saved for your user and tenant and synchronize across devices. A failed sync displays a Retry action while keeping the local changes. See Agent tests for the new testing screens.

Start a design¶

Open Design with AI, choose Voice, Chat, Copilot, Translation, Background or Multi-agent demo, and describe what it should help with. Select the arrow to start. For example: “Help customers find their orders and ask a person to review delivery exceptions.”

The designer saves a starting draft and asks only about missing decisions that matter, such as the outcome and allowed actions. A clear brief can go straight to a proposed design. It builds on what you already told it and chooses routine technical defaults.

For Voice, a demo request creates one voice agent. Existing context, secure verification links and requested in-call visuals can support that agent. The designer does not expand the request into a whole demo application.

The designer can research public information, inspect existing agents and available capabilities, and—when the Library connection is configured—learn from previous demo videos and transcripts. It should explain what it reused and link to sources it relied on. Examples guide the experience; your requirements determine the business rules.

Describe the outcome in your own words. The designer uses supported technical defaults and asks about choices that change behavior or required connections. For normal call ending and escalation it uses native session ending and a summary of the AI interaction. These are instructions in the agent's prompt, not separate use cases or questions about how your phone system works. Dialling and briefing another participant is a separate requirement.

When the agent handles protected member information, such as health claims, the designer selects Customer Authentication and supported secure defaults. With built-in identity, this uses device passkey verification and permits native first-time enrollment. The caller still consents to the link and must finish verification before receiving protected details. Your configured identity requirements take precedence. Authentication appears as a capability, rather than a separate business procedure or a configuration questionnaire.

New research features apply to new design conversations. Existing OpenAI conversations retain the tools they started with; you can also open the draft's editor and select New chat in the Assistant to use the latest guidance. Existing conversations receive updated scope and capability defaults when they read the saved design state.

Answer the designer's questions¶

While the designer is asking questions, a question form replaces the message box. Select a question by name or use Back and Next to move between them. Each question has its own answer box. Some also offer choices; you can add detail or write your own answer. Choose I’m not sure yet when you need help deciding. Nothing is selected for you.

The answered count shows your progress. You can revisit any answer before choosing Submit answers, which sends the whole group together. The normal message box returns once your answers are accepted. You can still attach images or documents while answering.

Unsubmitted answers survive a refresh in the same browser tab. They are not shared with another tab or device until submitted. If submission fails, your answers remain in the form so you can try again. Questions must be answered or marked undecided before you can build the draft.

You can name a preferred AI or voice platform, describe your organisation's requirements, or ask the designer to recommend one. Explain what matters to you: quick replies, a reassuring voice, cost or privacy. You do not need technical knowledge.

For Voice agents, the standard technology choices include:

No choice is selected for you. Final costs depend on the models and how the agent is used. The designer should check your information-handling requirements before recommending a setup; choosing one does not make it compliant by itself.

The designer summarizes the agreed brief and prepares the design. Questions belong in the conversation; the right panel records your answers and progress. If you want a quick starting point, say “Use sensible defaults for a first design,” then review its assumptions and choose Build draft.

When continuing an earlier draft, the designer may ask about choices that were never agreed, such as language, technology or how to reach a person.

Ending calls and handing off to a person¶

The built-in End Session action can finish a completed call or hand it back to your connected phone system for routing to a person. It can pass details such as why help is needed, a short summary and an order reference. Tell the designer if your phone system already handles this; you do not need to create a separate transfer tool just because it is absent from your saved tools.

The designer uses the normal handoff summary unless you specify a custom mapping. It asks about routing only when your request requires custom behavior. A callback request is a separate action. A warm transfer that dials and briefs someone before joining the calls is also a different option.

The agent should explain the handoff before ending its part of the conversation. Test the routing on your connected phone channel before relying on it; ending the AI conversation alone does not confirm that a person answered.

Continue a design reopens a saved conversation. Background agents use the existing background-work editor and start inactive.

A design reply or Build draft can run for up to 30 minutes by default. If it reaches the limit, the agreed design and any saved draft changes remain available. Review the saved progress before choosing Build draft again.

The design page shows Working on your design, Ready to continue, or Needs your attention beside saved conversations. Generation in progress also lists your unfinished Generate with AI requests. Select one to continue in its editor, including answering any waiting question. Statuses refresh while the page is visible.

Share a reference¶

In the conversation, select Add a reference to share a screenshot, diagram or written guide. Use PNG, JPEG or WebP images up to 5 MB, or plain-text and Markdown files up to 80 KB. Each design can hold eight files.

After uploading, tell the designer what to use from the reference. Ready to share means it is waiting to be read; Sent to designer means its contents were supplied to the assistant. Check that the resulting description matches your source. Uploading a guide does not connect it as the finished agent's knowledge source.

Select a filename to download it or its remove button to delete the attachment. Wait for a reply to finish before changing references. Removing a file does not erase information already included in the conversation or design. PDF, audio and video files are not supported here yet.

Browse agents¶

Choose Agents in the sidebar. Voice, chat, copilot, translation and background agents appear together with their type and status. Background agents are the agents previously called workflows; they automate tasks using the existing graph editor. The former Workflows list opens this same page.

• Use All, Voice, Chat, Copilot, Translation, or Background to filter the list. Search names, descriptions, IDs or tags, and sort by last modification or name.
• Select an item's heart to favorite it; select My favorites to show saved entries.
• Choose New agent, then select Voice agent, Chat agent, Copilot, Translation agent, or Background agent. Each opens the appropriate creation wizard. Import distinguishes conversational-agent and background-agent JSON or ZIP exports, with reference mapping when needed.
• Select an item to open its editor. Use its Manage button to edit tags, clone, export or delete it, or activate/deactivate a background agent.

Cards are the default view. Use Card view or List view to switch presentation. Filters adds status, creation-date and modification-date controls. Sorting supports name, type, status, creation, modification and recently opened order.

My favorites uses your saved favorites. My recents shows saved visits in the current tenant. My agents matches recorded creator metadata; older entries without that metadata are excluded.

The dashboard reference's Archive and Unpublish actions are not included. Subagent editing continues to use the real app's prompt scopes and configuration; there is no linked-agent canvas model.

The simulator and emulator sidebar tabs have been removed from this preview. The agent editor still includes its phone preview. If one list cannot load, the other stays available; choose Try again to retry.

Try the prompt editor¶

Open an agent and select Edit on its conversational node, then Behavior. If you are viewing a published version, choose Open draft, or Save As Draft if there is no draft yet. Published versions are read-only.

In the draft you can type in the prompt, use @ or {{ for the variable/tool dropdown, and use \ to browse snippets. Actions → Add opens the tool library, including the inline-action creation dialogs. Each action's settings button opens its configuration dialog. Draft edits save in your workspace. In standalone prototype mode they save in this browser instead.

Copy full prompt still requires the live prompt renderer and is unavailable in the standalone prototype. That limitation does not prevent editing, dropdowns or action dialogs.

Shape the design¶

The conversation is where you supply information and request changes. The right panel shows the result. Its title and description update as the assistant refines the design.

For example, ask “Keep the tone calm and reassuring” or “Never change an order without checking the customer's identity.” The assistant refines the proposed design and updates the overview. Behavior and test cases evolve throughout the conversation as your needs become clearer. When a decision needs your input, the designer opens a question form in the conversation.

Each reply includes a collapsed Thinking section. Expand it to follow live progress, such as checking templates, reviewing your reference or saving a change. It shows brief updates and design decisions, and remains available after reloading.

The designer can look for relevant templates, existing agents, available capabilities, and existing forms or visual components. It uses these as references while discussing what fits your needs.

When you have not specified a real system connection, the designer creates working sample tools so you can try the experience. It gives you example inputs and expected results in the conversation, including any sample names or verification details you need. Ordinary valid inputs should work too; special examples let you try failures and policy limits.

Sample actions do not cancel real orders, issue refunds or schedule calls. They also do not retain changes: a sample cancellation does not change a later order lookup. The designer checks sample-tool examples during preparation and building. Full conversation evaluations in Performance are separate.

Behavior and Validate focus on the content without repeating default Proposed or Not run labels. After a successful build, created tools show their connection status. A tool marked Created · connection needed has an editable starting definition that reports the missing connection when called. This deliberately disconnected option is used only when requested. Created · not verified means the real connection and full conversation have not been verified; the build reply separately reports sample checks. Set up credentials through the usual connection settings, rather than entering them in the conversation.

Validate prepares examples and expected outcomes—the “goldens”—as the design develops. Filter by Conversations, Procedures or Boundaries. Cases are refined as you discuss requirements. Test execution happens separately after building the draft.

On smaller screens, use Conversation and Design to switch views.

Build and open the draft¶

A starting draft is saved in the background. Stay in Design while agreeing the work, systems and examples. When ready:

1. Select Build draft. The designer creates the agreed instructions, procedures, specialists, tools and settings, and saves the test cases.
2. Review the build result. Agent examples appear in Performance → Goldens and their expected outcomes in Performance → Goals. Background examples become saved inputs in the workflow tester; their expected outcomes stay in Design.
3. Select Open draft for detailed configuration and testing. This becomes available only after the build checks confirm the required parts exist.

The page stays in Design throughout the build. If it stops before finishing, Continue build resumes from saved work. An older draft with only proposed procedures or tests needs this build step too.

Use the supplied sample inputs to test the draft. Connect real systems and review safeguards before publishing or activating it for real work.

If saved editor changes make the overview out of date, the designer shows a notice. Ask it to review those changes and build the revised design. Manually edited tests are preserved rather than overwritten. Avoid editing the same draft in another window while a design reply is making changes.

You can reload during a reply and return to saved progress. A lost browser connection does not stop work already accepted by the assistant. If a reply fails, review the saved design and ask the assistant to continue. Conversations that have ended remain viewable; their draft is still available in the editor.

This version prepares test cases but does not run them. Voice input to the designer and generating new customer-facing forms are not available yet.

Standalone design prototype¶

From an extracted designer export, run npm install and npm run dev at its root, then open port 3100. Use Node.js 22+ and npm 10.5.1+.

The export shows the same Design with AI screens as the real app, including all four agent types, Thinking progress, reference controls, output tabs, Build draft and Open draft. Its purple PROTOTYPE badge identifies the simulated backend. Replies are scripted examples; no model runs. The live service's replies and results will depend on your actual conversation and configured capabilities.

Each prototype reply normally takes about 30 seconds. Expand Thinking to see activity updates while you wait; text starts appearing before the turn finishes. You can keep typing your next message, but sending, changing references and opening the editor wait for completion. Reloading resumes the saved progress. This timing is simulated and does not predict the live service's response time.

Designs and drafts save in this browser. Reference files can be added, downloaded and removed, but are not interpreted by AI. Sent to designer is a simulated state here. Example design replies update the overview; Build draft updates the sample draft's description and conversational prompt. It does not compile the proposed specialists/procedures, connect systems or create Performance records. Translation and Background drafts use starting configurations to review in the editor. Background drafts remain inactive; validation examples are not executed.

Reset data clears current mock designs, drafts and uploaded references. Older browser-local briefs remain stored separately and are not automatically converted into current sessions or live conversations. Import dialogs still demonstrate the flow without creating imported agents. The older studio/real tree is a historical snapshot.

See also¶

• Creating an agent
• Pipeline editor
• Testing with the emulator