Skip to content
NewPazau Agent 3 is here: Procedures that test themselves before they go live
Changelog

What we shipped, every week.

New capabilities, improvements and fixes across Pazau Agent, Inbox, Voice, Insights and the platform. Every entry is live in production for every region on the date shown.

LatestAgent 3 follow-ups: simulation exports and approval SLAs· September 23, 2026
Get the changelog by email

One email on Thursdays with everything that shipped that week. No marketing, and you can unsubscribe in one click.

Agent9Platform5Voice4Inbox4Insights3API3Security3
  1. AgentInsightsAPI

    Agent 3 follow-ups: simulation exports and approval SLAs

    Thanks for a busy launch week. More than 600 workspaces ran a simulation in the first seven days. This release covers the most-requested follow-ups and a handful of fixes.

    New

    • Agent: export simulation results as CSV or JSON, including per-conversation diffs and scorecards.
    • Actions: approval SLAs. Pending approvals escalate to a backup approver after a timeout you choose, and they're flagged in Inbox.
    • API: POST /v2/simulations and GET /v2/simulations/{id}, so you can run simulations from CI before promoting a Procedure.
    • Insights: a resolution-rate dashboard that uses the Agent 3 definition, with 7-day repeat-contact tracking.

    Improved

    • Agent: simulation queue times cut from a median of 90 seconds to 20 during peak hours.
    • Insights: resolution rate can be split by Procedure version, so you can see the before and after of every publish.

    Fixed

    • Agent: simulations for Procedures created before Agent 3 failed if they referenced an archived Action.
    • Voice: some calls migrated to Agent 3 Procedures mid-call used the old greeting.
    • Actions: the approval card in Slack didn't render the amount for currencies with no minor unit (JPY, ISK).
  2. AgentVoicePlatform

    Pazau Agent 3

    Pazau Agent 3 is generally available. Procedures now test themselves before they publish, Actions can span multiple systems with approvals built in, and the same Procedures run on voice as on chat. Read the full announcement: Introducing Pazau Agent 3.

    New

    • Agent: simulation gates on publish. Every Procedure change is replayed against up to 2,000 of your historical conversations, and publishing is blocked if the resolution rate or policy adherence drops below your threshold.
    • Actions: multi-step Actions with approvals are generally available, with compensating Actions, per-step spend limits and approver routing to Slack, Teams or Inbox.
    • Voice: voice parity. The same Procedures and Actions run on voice, chat, email and messaging, with no separate voice flows to maintain.
    • Agent: a published resolution rate definition. A conversation counts as resolved only if no human touched it, the customer didn't come back about the same issue within 7 days, and the customer didn't abandon or ask for a person.
    • Platform: Agent 3 is available in the US and EU regions (Dublin and Frankfurt) on day one.

    Improved

    • Agent: reasoning traces are 40% shorter to read, with collapsed tool calls and a one-line summary per step.
    • Voice: p95 voice-to-voice latency under 800ms with multi-step Actions, using filler speech only when an Action is slow.
    • Agent: multilingual Procedures. Write once in English and run in 45+ languages, with per-language overrides where policy differs.

    Fixed

    • Agent: a Procedure that asked for confirmation twice when an Action returned a warning now asks once.
    • Actions: approval timeouts now respect the approver's business hours when configured.
  3. AgentInbox

    Simulation at scale and handoff summaries

    Simulation gets bigger and faster ahead of next week's launch. The handoff to your human team gets tighter too.

    New

    • Agent: simulation now runs against up to 2,000 historical conversations, up from 500, and finishes in under 6 minutes at p95.
    • Agent: simulation scorecards grade every replayed conversation on resolution, policy adherence and tone, with a pass threshold you set.
    • Inbox: structured handoff summaries: what the customer asked, what the agent already did, which Actions ran, and what's left.

    Improved

    • Agent: simulations sample conversations stratified by topic and channel by default, so rare intents aren't drowned out.
    • Inbox: human agents can rate a handoff summary, and the rating feeds Insights.

    Fixed

    • Agent: simulations of Procedures that call knowledge search now use the knowledge base as it was at the time of the original conversation.
    • Inbox: the "AI attempted" badge was missing on conversations handed off from Voice.
  4. AgentPlatform

    Procedure folders, owners and review

    As teams run hundreds of Procedures, they need the same controls they use for code. Procedures now have owners, folders and a required-review workflow.

    New

    • Agent: Procedure owners and required reviewers. Publishing can require approval from one or two reviewers, set per folder.
    • Agent: folders with inherited permissions, so a billing team can own every billing Procedure.
    • Platform: sandbox-to-production promotion for Procedures and Actions, with a diff before you promote.

    Improved

    • Agent: the Procedure list shows each Procedure's 7-day resolution rate and volume.
    • Agent: @-mentions in Procedure comments notify reviewers in Slack and Teams.

    Fixed

    • Platform: sandbox workspaces could inherit production webhook URLs when cloned. They now start with webhooks disabled.
    • Agent: deleting a knowledge article used by a Procedure now warns you and lists the affected steps.
  5. SecurityAgent

    Pazau Guard: custom redaction entities

    Every industry has identifiers that aren't in any standard PII list. Guard now lets you define your own, and redacts them before any model sees the conversation.

    New

    • Security: custom redaction entities. Define them with a regex, a checksum rule (Luhn, IBAN mod-97) or examples, such as member numbers, policy IDs or loyalty numbers.
    • Security: a redaction test bench. Paste a sample conversation and see exactly what Guard masks before you turn a rule on.
    • Agent: Procedures can reference redacted values by token, so the agent can pass a card's last four digits to an Action without ever seeing the full number.

    Improved

    • Security: built-in detection of national ID formats extended to 31 countries, including Sweden's personnummer and Ireland's PPSN.
    • Security: redaction adds a median of 9ms per message, down from 22ms.

    Fixed

    • Security: phone numbers written with spaces between each digit were not always redacted in chat transcripts.
    • Agent: redacted tokens were occasionally shown to the customer as "[REDACTED_1]" in a reply. The agent now rephrases.
  6. VoicePlatform

    Voice p95 under 800ms

    Voice-to-voice latency is now under 800ms at p95 and around 620ms at p50, measured from the caller's last word to our first audio byte. How we got there is in Voice agents under 800ms.

    New

    • Voice: speculative response generation. The agent starts drafting while the caller is still finishing a sentence, and discards drafts that don't match.
    • Voice: a latency breakdown in every voice trace: ASR, reasoning, Action calls and TTS, each in milliseconds.

    Improved

    • Voice: p95 voice-to-voice latency down from 1,140ms to 790ms. The p50 is now 620ms.
    • Voice: filler phrases ("Let me pull that up") only play when an Action is expected to take longer than 700ms.
    • Platform: media servers added in us-west-2 and eu-central-1 to shorten the network path.

    Fixed

    • Voice: long numbers such as confirmation codes are read back in groups of three, not as a single number.
    • Voice: the TTS voice could change mid-call after a warm transfer back to the AI agent.
  7. AgentAPI

    Multi-step Actions with approvals (beta)

    Real resolutions often take more than one call: look up the booking, check the fare rules, rebook, then issue the difference as a voucher. Multi-step Actions let Pazau Agent chain them, with a human approval at any step you choose.

    New

    • Actions: multi-step Actions (beta). Chain up to 12 Actions, pass outputs between steps, and define compensating Actions that run if a later step fails.
    • Actions: approval steps. Route to a person or a Slack channel, set a timeout, and set a default outcome if no one responds.
    • API: POST /v2/approvals/{id}/decide for approving or rejecting from your own tools.

    Improved

    • Actions: input schemas now support enums and regex validation, so malformed values are caught before they reach your API.
    • Agent: the customer is told "I'm checking this with a colleague" while an approval is pending, instead of a silent wait.

    Fixed

    • Actions: OAuth tokens for Salesforce connections refresh 5 minutes before expiry instead of on the first failed call.
    • API: action.completed webhooks now include the conversation_id.
  8. InsightsInbox

    Contact driver alerts

    Insights now tells you when something new is driving contacts, before the queue tells you. Spikes are detected hourly and sent to Slack, Teams or email.

    New

    • Insights: contact driver alerts. When a topic's volume rises more than 3 standard deviations above its 28-day baseline, Insights posts an alert with example conversations.
    • Insights: "Why did this escalate?" summaries on every human handoff, grouped by root cause.
    • Inbox: one-click "Create Procedure from cluster" from any Insights topic.

    Improved

    • Insights: dashboards support custom date comparisons, such as this peak week against the same week last year.
    • Insights: QA scorecards can now be shared with agents in Inbox, with comments.

    Fixed

    • Insights: time zones on scheduled reports now follow the report owner's setting, not UTC.
    • Inbox: tags applied by automation rules now appear in Insights filters within 1 minute instead of up to 1 hour.
  9. VoiceSecurityPlatform

    EU data residency for Voice

    Pazau Voice now runs fully in the EU. Call audio, transcripts and traces for EU workspaces are processed and stored in Dublin and Frankfurt, like the rest of the platform. See the Trust Center for the full data map.

    New

    • Voice: EU-resident speech recognition and synthesis. Audio for EU workspaces never leaves the EU region.
    • Security: a per-workspace data residency report that lists every storage and processing location, exportable as PDF for your DPO.
    • Platform: EU-region phone numbers in 14 countries, provisioned from the admin console.

    Improved

    • Voice: first-audio latency for EU callers down 90ms on average, now that calls no longer route through US media servers.
    • Security: call recordings are encrypted with per-workspace keys, and customer-managed keys (AWS KMS) are available on Enterprise.

    Fixed

    • Voice: call recording retention settings were ignored for calls transferred between two Pazau numbers.
    • Platform: the region badge in workspace settings showed "US" for some EU workspaces created through the API.
  10. Agent

    Procedure simulation (preview)

    You can now replay a Procedure change against your real historical conversations before it goes live. This is an early preview of the evaluation system we describe in Evaluating agents with simulation.

    New

    • Agent: Procedure simulation (preview). Run a draft Procedure against up to 500 past conversations and see which outcomes changed, step by step.
    • Agent: simulated Actions run against recorded responses from your systems, so nothing is refunded or rebooked for real.
    • Agent: a diff view that highlights conversations where the draft took a different Action than the live version.

    Improved

    • Agent: the Procedure editor autosaves drafts every 5 seconds and keeps 50 versions of history.
    • Agent: knowledge retrieval now prefers articles updated in the last 90 days when two articles conflict.

    Fixed

    • Agent: Procedures with more than 40 steps rendered slowly in the editor. They now load in under a second.
    • Agent: a Procedure renamed while a conversation was in progress showed the old name in that conversation's trace.
  11. Voice

    Voice: faster barge-in and better endpointing

    Callers interrupt. A voice agent that keeps talking over them feels broken. This release cuts the time it takes Pazau Voice to notice and yield.

    New

    • Voice: configurable endpointing sensitivity per Procedure, so "read me your policy number" waits longer for digits than "anything else?" does.
    • Voice: DTMF fallback for any spoken input step. Callers can type digits instead when speech recognition confidence drops below your threshold.

    Improved

    • Voice: barge-in detection p95 down to 180ms from 290ms. The agent stops speaking and discards the rest of its planned utterance.
    • Voice: fewer false barge-ins from background noise. We now require 120ms of voiced speech before yielding, which cut false interruptions by 64% on our street-noise test set.
    • Voice: live transcription word error rate on accented English improved 11% relative.

    Fixed

    • Voice: warm handoff could drop the last two seconds of caller audio from the transcript passed to the human agent.
    • Voice: hold music restarts from the beginning after a transfer instead of mid-track.
  12. Inbox

    Collision detection for WhatsApp and SMS

    Two agents replying to the same customer on WhatsApp is a bad look. Collision detection, which already covered email and chat, now covers every messaging channel.

    New

    • Inbox: collision detection for WhatsApp and SMS threads. You'll see who else is viewing or typing, and the second reply is held with a one-click "send anyway."
    • Inbox: WhatsApp template messages can now be sent from macros, with variable validation before send.
    • Inbox: SLA pause rules. Pause the clock while waiting on the customer, and resume automatically on their reply.

    Improved

    • Inbox: the conversation list loads 2.3x faster for queues with more than 5,000 open conversations.
    • Inbox: the AI-drafted reply now shows which knowledge article or Procedure step it drew from.

    Fixed

    • Inbox: emoji reactions from WhatsApp no longer reopen resolved conversations.
    • Inbox: SMS messages over 160 characters sent from Inbox are no longer split into out-of-order segments on some carriers.
  13. APIAgent

    Idempotency keys and a faster Actions API

    If your integration retries, it should never refund twice. The Actions API now accepts idempotency keys, and we have tightened latency on the endpoints you call most. Full reference on the developer docs.

    New

    • API: Idempotency-Key header on POST /v2/actions/execute. Keys are kept for 24 hours, and a repeat request returns the original response with idempotent-replay: true.
    • API: GET /v2/traces/{id} now returns the full step list, including Procedure step IDs and Action inputs and outputs (PII-redacted).
    • API: webhook event action.approval_requested for building approvals into your own tools.

    Improved

    • API: p95 latency on POST /v2/conversations/{id}/messages down from 310ms to 140ms.
    • Agent: Actions that time out now report the downstream system's name and timeout in the trace instead of a generic error.

    Fixed

    • API: cursor pagination on GET /v2/conversations could skip a record when two conversations shared a created_at timestamp.
    • API: rate-limit headers now reflect the organization limit rather than the per-key limit.
  14. Insights

    QA rubric versioning in Insights

    Your QA rubric changes. Your trend lines shouldn't break when it does. Insights now versions every rubric and scores each conversation against the version that was live at the time.

    New

    • Insights: QA rubric versioning. Each edit creates a new version with a changelog, and you can re-score up to 90 days of conversations against a draft before publishing it.
    • Insights: side-by-side score comparison between two rubric versions, broken down by criterion.
    • Insights: a "Compliance" criterion type that marks a conversation as failed regardless of the total score, for required disclosures such as call-recording notices.

    Improved

    • Insights: 100% QA scoring now completes within 4 minutes of a conversation closing at p95, down from 11 minutes.
    • Insights: topic clustering handles mixed-language conversations and no longer creates duplicate clusters for the same driver in English and Spanish.

    Fixed

    • Insights: CSAT prediction showed "—" for conversations shorter than three messages. These now get a prediction with a low-confidence badge.
    • Insights: CSV exports with more than 100,000 rows no longer time out.
  15. SecurityPlatform

    SCIM group-to-role mapping

    Identity teams can now manage who does what in Pazau entirely from Okta, Microsoft Entra ID or any SCIM 2.0 provider. No more reassigning roles by hand after a reorg.

    New

    • Security: SCIM group-to-role mapping. Map an IdP group to any Pazau role (Admin, Supervisor, Agent, Analyst, or a custom role), and membership changes apply within 60 seconds.
    • Security: custom roles can now restrict access to Procedures by folder, so a team can edit only its own SOPs.
    • Platform: audit log export to Amazon S3 and Google Cloud Storage on an hourly schedule.

    Improved

    • Security: deprovisioned users have their sessions revoked within 30 seconds, down from the next token refresh (up to 15 minutes).
    • Platform: the admin audit log now records the IdP group that granted each permission.

    Fixed

    • Security: SAML logins with an email address in mixed case no longer create a duplicate user.
    • Platform: the audit log filter "Actor type: API key" returned no results for keys created before March. It now returns them.
  16. AgentInbox

    Spend limits for Actions, and faster macro search

    Refunds, credits and goodwill gestures now have hard ceilings you set per Action. You can let Pazau Agent issue money back without worrying about a bad day.

    New

    • Actions: per-action spend limits with daily and per-conversation caps. When a cap is hit, the Action pauses and routes to a human with the full trace attached.
    • Actions: a currency-aware limit editor. Caps can be set in USD, EUR, GBP or SEK, and conversion uses the day's ECB reference rate.
    • Inbox: saved views can be pinned to the sidebar per team, not just per agent.

    Improved

    • Inbox: macro search is now fuzzy and ranks by usage in the last 30 days. Median time to insert a macro dropped from 6.1s to 2.4s in internal testing.
    • Agent: Traces now show which Procedure version handled the conversation, next to the step that ran.

    Fixed

    • Inbox: long email threads with more than 200 messages no longer truncate the oldest quoted replies.
    • Actions: a retry after a 429 from the downstream API could be logged twice in the trace. It is now logged once, with the retry count.
Looking for API changes?

Versioned API and webhook changes, with deprecation timelines, are published in the developer docs.

docs.pazau.com

See what your team could resolve.

Bring 1,000 of your real conversations. We'll show you, in a 30-minute call, how many Pazau Agent resolves end-to-end.

Book a demoSee pricing