The AI agent that finishes the job.
Pazau Agent resolves conversations end-to-end. It follows your procedures, rebooks the flight or issues the refund in your systems, and records why it did each step, within limits your team sets.
- GuardPII detected · booking ref, email → redacted in logs
- Procedure matchedInvoluntary schedule change → Rebook (v14)
- Action · get_bookingSOL-88K2Q · 1 pax · fare class V · 200 ms
- Reasoning…
Write the SOP once. Pazau follows it every time.
Procedures are your standard operating procedures written in plain language, the way you'd brief a new hire on day one. No decision trees, no intent libraries, and no flow builder with 400 boxes.
Plain language, real logic
Branches, conditions, and loops written as sentences. Reference Actions with @mentions and approvals inline. CX ops owns them, not engineering.
Simulate before you ship
Every edit is replayed against hundreds of your past conversations. You see what resolves, what escalates, and what changed before a customer does.
Versioned and reviewable
Every Procedure has a version history, a diff view, and optional four-eyes review. Roll back in one click, or roll out to 10% of traffic first.
Actions do the work in your systems, within limits you set.
An Action is a typed API call into your order system, PSS, claims core, or billing platform. Pazau validates every parameter, checks it against your limits, and asks a human when the stakes call for one.
- Approval thresholds. Auto-approve refunds up to $250, route $250 to $1,000 to a CX lead in Slack or Inbox, and block anything above that.
- Spend limits. Caps per conversation, per customer over a rolling window, and per day across all conversations. When a cap is reached, the Action pauses and a human takes over.
- Least-privilege credentials. OAuth scopes per Action, secrets in a KMS-backed vault, and idempotency keys so a retry never refunds twice.
- Import in minutes. Point Pazau at an OpenAPI spec or use one of 120+ prebuilt connectors. Browse integrations
Refund all or part of an order to the original payment method. Use only after the order is verified and the customer has confirmed the amount.
- order_idstringfrom get_order · must belong to verified customerreq
- amountmoney≤ order total − prior refundsreq
- reasonenumdamaged · not_received · late · goodwillreq
- methodenumoriginal_payment (default) · store_credit
get_order, get_booking, get_policy, get_claim_status. No approval needed, logged in the Trace.
issue_refund, rebook_flight, update_address, file_fnol. Thresholds, approvers, and confirmation from the customer.
close_account, change_beneficiary, card_dispute. Step-up verification and a mandatory human approver.
Controls built in, not bolted on.
Pazau Guard runs on every turn, before and after the model. It redacts, checks policy, and decides whether a response ships, gets rewritten, or goes to a person.
Card numbers, SSNs, health data, and passport numbers are masked before inference.
No medical, legal, or investment advice. No promises your terms don't make.
Answers must cite a knowledge source or Action result, or they don't ship.
Legal threats, self-harm, vulnerable customers, and regulator mentions go to a person with a summary.
- RedactedCard number (PAN) removed before model call
- BlockedResponse contained unapproved medical advice → rewritten
- RoutedEU customer → processed in eu-west-1 (Dublin)
- ApprovedRefund $480 · approved by CX lead J. Alvarez
- HeldLegal threat detected → handed to human with summary
Every decision, explained and exportable.
Each conversation produces a Trace: the Procedure version it followed, the knowledge it retrieved, each Action with its inputs and outputs, every Guard check, and the reason for each step. QA teams read Traces. So do auditors.
- Captured per step
- Inputs, outputs, latency, rule matched
- Format
- JSON, CSV, or PDF case file
- Tamper evidence
- Hash-chained, write-once
- Search
- By Procedure, Action, outcome, or customer
{
"conversation": "cnv_8Kq2x",
"procedure": "involuntary_schedule_change@v14",
"steps": [
{ "type": "guard.redact", "fields": ["email", "pnr"] },
{ "type": "action", "name": "search_flights", "ms": 412 },
{ "type": "policy.check",
"rule": "fare_diff <= auto_approve_limit",
"inputs": { "fare_diff": 0, "limit": 150 }, "result": true },
{ "type": "action", "name": "rebook", "status": 200 }
],
"outcome": "resolved", "handoff": false,
"reasoning": "Cancellation is carrier-initiated; same-day
alternative in same fare class, so no charge."
}
// ↳ 7 steps · 1m 48s · model inputs redactedOne agent, every channel.
Write a Procedure once and it runs on chat, email, voice, and messaging, adapted to each channel's format.
45+ languages, one Procedure.
Pazau detects the customer's language and answers in it. You don't maintain translations. Vasa Mobility runs one set of Procedures across 11 languages in 40 cities.
We count resolutions strictly. That's what you pay for.
A resolution means the customer's issue was handled end-to-end with no person involved and stayed handled. A reply that sends the customer to a help article doesn't count.
The 2,100 conversations between the last two bars were excluded because the customer abandoned before confirming, or a Guard escalation fired after close.
Approved Actions still count as resolutions when the approver only clicks approve. Full definition in the Subscription Agreement.
“The year before, peak season meant 110 seasonal hires. This year we handled 486,000 conversations at 3.4× normal volume with zero, and Pazau Agent resolved 71% of them end-to-end.”
Live in weeks, not quarters.
A median Pazau Agent deployment takes 5 weeks from kickoff to 100% of traffic. Your forward-deployed engineer and CX strategist work in your Slack channel throughout.
- 1Week 1
Connect
Plug in your helpdesk, knowledge base, and the first 5–10 Actions. Import 6–12 months of conversation history. - 2Weeks 2–3
Write and simulate
Draft Procedures for your top 20 contact drivers. Simulate against real history until they pass your QA bar. - 3Week 4
Shadow and pilot
Pazau drafts while your team sends. Then it goes live on 5–10% of traffic with human review on escalations. - 4Weeks 5–6
Scale
Ramp to full traffic channel by channel. Weekly reviews of Insights drivers decide which Procedures to add next.
Questions about Pazau Agent.
Which models does Pazau Agent use?+
Do you train on our conversations?+
Do we have to replace our helpdesk?+
What happens when Pazau Agent can't resolve something?+
How do you stop the agent from making things up?+
Can we buy Pazau Agent without the rest of the platform?+
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.