Release notes

A simulated run now tells you when your test case left out a value that an API tool set to Mock returns. Before, the AI Agent acted on an empty value, and the run looked like an AI Agent fault.

Simulation analysis lists What External systems leaves out below your text:

• An output that came back empty, when your text names neither the output nor its API tool.
• A record written as JSON that has no key that an output of the API tool reads.

The API usage report now shows how long your API tools take to respond, across all of your conversations.

• Each API tool shows its median response time and its slowest 10% response time.
• A chart shows how response times change over the date range you pick.
• The call log shows the response time of every call, and you can sort the log by it.
• In the Conversation View, the details of an API tool now show its typical response time over the last 30 days. You can see if one slow call is unusual for that tool.

End users waiting for a live agent after a Salesforce Messaging handoff now see a queue banner in the chat widget. The banner shows their queue position or estimated wait time. Previously, end users saw no queue indicator while they waited.

Estimated wait time appears with no extra setup. To show queue position instead, a Salesforce admin enables Queue Wait Time and Position on each Messaging Channel and republishes the linked deployment. Salesforce sends one signal or the other per Messaging Channel, not both. The banner appears in the Messaging and legacy Chat widgets only.

get_test_runs now returns the simulated conversation of a single run fetched by test_run_id. Each conversation item holds one readable line, the same text the success criteria judge reads. An assistant can show you the conversation without the dashboard.

Set include_transcript to false to leave out the structured transcript items. The conversation text then arrives once, so the read takes fewer pages.

When every Zendesk SMTP Connector toggle of an AI Agent is off, Ada now rejects the agent replies that Zendesk sends through the connector. Before, Ada still delivered them. Zendesk shows a send error for each rejected reply.

To turn off the connector, delete the connected address and domain in Zendesk before you turn off the toggles. For the steps, see Disable the connector.

You can now control which phone numbers and email addresses your AI Agent shares from Knowledge. By default, the Agent keeps them private, even when an end user asks for one. Opening hours are always shared.

Enable Share contact details from Knowledge under Manners in the Persona settings. Then choose All contact details in Knowledge or Only the contact details I list. In the list mode, enter the phone numbers, email addresses, and *@domain wildcards that the Agent can share. The list needs at least one value before you can save.

A simulation can now answer an API tool from the test case. You describe the records and rules for each test. You can then test every branch of a flow, with no test accounts to set up or reset. Use it to find Playbook gaps and to test fixes in the improvement loop. Every API tool stays Live by default. Set an API tool to Mock to use it.

What changes

When you click Regenerate Password for the Zendesk SMTP Connector, the previous password now keeps working for 5 minutes. Update the password in Zendesk in that time. Before this change, the previous password stopped working at once, so human agent replies failed until you entered the new password.

See Zendesk SMTP Connector.

get_ada_metric has two new metric types, csat_rate_ai and csat_rate_human. They return the AI Agent Score and the Human Agent Score from the Customer satisfaction score report. Each survey counts for the AI Agent or for the human agent who joined the conversation. The two metric types return null when no survey matches.

Change to the HANDOFF filter

The Messaging SDKs now wait for engagement before creating a new conversation session. A hidden, zero-sized, or transparent native view no longer creates a session or sends the automatic greeting at app launch. A native view also waits while the app runs in the background. An existing session still restores at once.

On native views with visibility reporting, isOpen() and getInfo().isChatOpen follow view visibility and return false before the first visibility report.

A new Offer a human agent only after trying other options toggle is in the Preferences tab. It keeps your AI Agent working on an end user’s issue before it offers a handoff. With the toggle on, the Agent uses every option it has, such as Knowledge, Playbooks, Actions, documented steps and follow-up questions. When no option is left, it offers once to connect the end user to a human agent. End users who ask for a human agent are still handed off. The toggle is off by default.

Five MCP server reads now return their results in pages. A large AI Agent configuration no longer arrives as one response that an assistant cannot read in full.

Which reads are paged

These reads are paged by size:

• get_ada_configuration
• get_test_cases, when it lists test cases
• get_test_runs, for a single run fetched by test_run_id
• list_entities with entity_type='variables'
• list_entities, for a version 2 playbook fetched by entity_id

The Snapshot API now exports and imports the message the AI Agent says when it withholds a sensitive image, in the image_redactions_message section. The section is one value: your sentence, or null when the Agent uses the default sentence.

• Export: the snapshot carries your sentence, or null.
• Import: if the section is present, the import replaces the message on the target Agent. null restores the default sentence. If the section is absent, the message does not change.
• Validate: a mode: validate run reports whether the import changes the message.
• Audit log: if the import changes the message, it writes one Updated row for Image Redactions Message.

The AI Agent now records why it hands a conversation to a human agent. It records a reason_code from a fixed list of ten values, or null when it records no code. It also records a reasoning text when one is available.

The v2 Conversations endpoint of the Data Export API now returns a handoffs array on each conversation. Each entry is a handoff attempt with its outcome_status, reason_code and reasoning. The get_conversation MCP tool returns reason_code and reasoning.text on its handoff entries. The new HANDOFFREASON filter in get_available_filters filters conversations by reason code.

Your AI Agent now reads receipts, invoices, and order confirmations that show only the last 4 digits of a card. Before this change, the Agent withheld these images as payment cards.

What’s changed

The rule that withholds payment cards now tells a receipt apart from a card.

A Voice Proactive Outreach that runs a Playbook can now start every call with an Opening line: the exact words that the end user hears the moment they answer.

The line does not depend on how the Playbook begins, and it plays before the Playbook’s first step is ready. Write it once in the default language of your AI Agent. Include your business’s registered name, along with any AI disclosures or recording notices required by law. Every call then opens with those exact words before the Playbook runs.

The Playbook editor now lists every error in one menu in the page header. AI Managers can go straight to each error and fix it before they activate the Playbook.

What’s new

The N errors to resolve menu sits next to the Active toggle and replaces the error count on the canvas.

A new Keypad entry timeout (seconds) setting is on the Voice Configuration tab. It controls how long your AI Agent waits after the last key press before it uses the digits. Set it between 2 and 7 seconds. Pressing # submits the digits at once.

The timeout applies wherever an end user enters a value on the phone keypad. That includes Ask steps in Playbooks, Action inputs, and keypad capture in Greetings, Handoffs, and Processes. A shorter timeout makes keypad entry feel faster. A longer timeout gives end users more time between keys.

You can now replace the sentence your AI Agent says when it withholds a sensitive image. Type your sentence in Message when an image is withheld under Image redactions. The Agent translates your sentence when the conversation is in another language. It uses your sentence for identity documents and other credentials when it can offer a human agent, and the default sentence otherwise. Leave the field empty to keep the default reply.

Admins can now save the Zendesk Messaging signing key ID and shared secret in the dashboard. Open Config > AI AGENT > Handoffs, then Configure beside Zendesk Messaging. Before this change, the Ada team added the key for you.

The Embed2 web chat now supports the Zendesk JWT. After you save the key, the web chat requests the JWT from zdChatterAuthCallback, as the Messaging SDK does. Ada uses the verified details to find or create the Zendesk user at the Handoff. See Use a Zendesk JWT.