Skip to content
Agent Platform

agpThe governed platform for AI agents

Agents act on your behalf. Every action leaves a receipt.

Put AI agents to work for your customers and teams. They act only with the permission you give them, ask a person when it matters, and keep a record of everything they do that anyone can check.

Book a demoSee how it works

Card Support

ACME Bank · AI agent

Zalo OAReviewed agent

  1. Thẻ của tôi bị trừ tiền hai lần cho đơn 48213.

    Linh · “My card was charged twice for order 48213.”

  2. Card Support wants to issue a refund

    Amount
    6,200,000 VND
    Card
    •••• 4821
    Order
    48213

    Allowed in this conversation only: refunds and transaction lookups, for 15 minutes

    Held: refunds over 5,000,000 VND need card operations to approve

  3. Tôi có thể hoàn lại khoản trừ trùng 6.200.000 VND. Khoản hoàn trên 5.000.000 VND cần bộ phận vận hành thẻ phê duyệt, tôi sẽ báo bạn ngay khi có kết quả.

    “I can refund the duplicate charge. Refunds over 5,000,000 VND need card operations to approve; I'll tell you as soon as they decide.”

    🤖 Tin nhắn này do trợ lý AI soạn.

  4. Đã hoàn tiền. Khoản 6.200.000 VND sẽ về thẻ của bạn trong 1–3 ngày làm việc.

    “Refunded. The 6,200,000 VND reaches your card within 1–3 working days.”

    🤖 Tin nhắn này do trợ lý AI soạn.

What happened

  1. Message received

    From Linh on Zalo OA

  2. Handed to Card Support

    Permission for this conversation

  3. Checked against the bank's rules

    Refund needs card operations

  4. Linh told it's waiting

    Marked as written by AI

  5. Approved by Nhan

    Card operations

  6. Linh told it's done

    Marked as written by AI

Every step signed and on the record

Example conversation with sample names and amounts.

Meet customers and teams on the channels they already use.

  • Zalo OA
  • Slack
  • Microsoft Teams
  • Messenger
  • Telegram
  • Google Docs
  • Google Forms
  • Your website and apps
  • Scheduled tasks

Every action an agent takes should answer three questions: who asked for it, who allowed it, and what exactly happened. Agent Platform answers all three, every time.

How it works

One platform between your channels, your agents and your systems

Every conversation passes through the same core. It finds the right agent, gives it only the permission the task needs, checks your rules before anything happens, asks a person when it matters, and keeps a record of each step. Scroll to follow one refund.

How Agent Platform connects channels, agents, people and your systemsOn the left, customers reach you on channels such as Zalo, Slack, Teams, Messenger, Telegram, Google Docs and your website. In the middle, the platform core has routing, permissions, rules, connections to your systems, approvals and records. On the right, agents such as Card Support and Fraud Screen work together, and people can have delegates. Below, an approver from card operations and your systems such as core banking.CHANNELSAGENTSYOUR SYSTEMSPEOPLE WHO APPROVEAGENT PLATFORMLLinhCustomerSlackMicrosoft TeamsMessengerZalo OATelegramGoogle DocsWebsite and appsRecords6 steps on recordRoutingfinds the right agentPermissionsonly what's neededApprovalspeople decideRuleschecked firstConnectionsto your systemsNNhan's delegateanswers routine questionsFFraud Screenchecks the chargesCCard Supporttalks to customersBuilt by any team,working togetherNNhanCard operationsCore bankingKnowledge baseTicketing1234567
  1. 1

    A customer writes in

    Linh messages ACME Bank on Zalo about a card payment taken twice. The platform confirms the message really came from Zalo and recognises her.

  2. 2

    The right agent picks it up

    The conversation goes to Card Support, with permission for this conversation only, for 15 minutes.

  3. 3

    Agents work together

    Card Support asks Fraud Screen to check the charges. It can share part of its permission, never more than it has.

  4. 4

    Your rules come first

    Before the refund runs, the bank's rules are checked. Refunds over 5,000,000 VND wait for card operations, so Linh is told it's on its way.

  5. 5

    A person approves

    Nhan in card operations approves from the approvals inbox. The approval covers this exact refund and nothing else.

  6. 6

    The action runs

    Only now does the refund reach the bank's core system.

  7. 7

    The reply goes out, on the record

    Card Support tells Linh it's done, marked as written by AI. The message, the rule check, the approval and each reply are kept on a signed record.

Features

Everything an agent needs to be trusted with real work

Each feature solves a problem teams hit the moment an agent does more than chat. Here is what it does, and what it looks like in practice.

Permissions

Each agent gets only what the task needs

Permission is given per conversation, for a short time, and only for what the task needs. When one agent asks another for help, it can pass on less, never more. Take it back once and it's gone everywhere.

In practice

Card Support may make refunds in Linh's conversation for 15 minutes. The fraud check it calls on can only look at her transactions.

Permissions in Linh's conversation

Card SupportActing for Linh15 min

Refunds and transaction lookups

Fraud ScreenAsked for help15 min

Refunds and transaction lookups

Rules

Your rules decide, not the AI

Write your limits once: amounts, who approves what, what must never happen. They're checked before every action, whatever the AI was told.

In practice

At ACME Bank, a transfer between a customer's own accounts goes through at once. A large transfer waits for a ledger officer.

ACME Bank · transfer rules

Transfers of 5,000,000 VND or more need a ledger officer

  • 3,000,000 VND to own savings Checking rules
  • 20,000,000 VND to a recipient Checking rules

Approvals

A person decides when it matters

Urgent, costly or unusual actions wait in an approvals inbox. The approver sees exactly what will happen and approves only that. If the details change, it needs approving again.

In practice

An urgent IT ticket waits for the duty manager. A transfer that's rejected never moves any money.

Waiting for you · expires in 3:58:12

Open urgent ticket · IT Helpdesk

Issue
VPN down for the whole finance team
Priority
Urgent
Asked by
Bob, finance

You approve exactly this ticket. A changed request needs a new approval.

Records

Know why every message was sent

Messages, rule checks, approvals and replies are kept on a signed record that nobody can quietly change. Pick any message and see the steps that led to it.

In practice

A compliance officer opens a reply sent to a customer and sees who asked, which rule applied and who approved.

Reply to Linh · 10:43

Why was this reply sent?

  1. Message received from Linh on Zalo
  2. Handed to Card Support
  3. Held by the bank's refund rule
  4. Linh told it's waiting for approval
  5. Nhan approved the refund
  6. Linh told the refund is done

Every step signed · nothing changed since

Learning

Agents get better, and you stay in control

Feedback and outcomes turn into suggested improvements: clearer instructions, sharper skills, better routing. Each one is tested on past conversations, reviewed by a person, tried quietly alongside the live version, then rolled out to a few conversations first. Any change can be undone.

In practice

A request the agent always handles the same way becomes a fixed routine that is faster and cheaper. If the routine starts failing, the agent goes back to working it out step by step.

Improvement loop: Feedback and outcomes, then Suggested improvement, then Tested on past conversations, then Reviewed by a person, then Tried quietly alongside, then Rolled out gradually, and back to feedback. A person approves, and any change can be undone.Feedback and outcomesSuggested improvementTested on past conversationsReviewed by a personTried quietly alongsideRolled out graduallyYour agentimproves under your rules
Nothing ships without approval Undo any change

Delegates

Your experts, without the interruptions

Anyone can have a delegate that answers routine questions for them. It drafts an answer when it isn't sure, checks before sharing anything sensitive, and only picks up new habits you approve.

In practice

Minh asks Nhan about response times for urgent card incidents. Nhan's delegate drafts the answer, and Nhan approves it in one click.

card-operationsSlack

  1. Minh 10:42

    @Nhan how fast do we respond to urgent card incidents?

  2. Draft from your delegateOnly you see this

    Urgent card incidents get a first response within 15 minutes, then an update every 30 minutes until they're fixed.

    From the card operations handbook · not sure enough to send on its own

  3. Nhan's delegate 10:43

    Urgent card incidents get a first response within 15 minutes, then an update every 30 minutes until they're fixed.

    ✅ Approved by Nhan

Suggested habit: answer teammates' questions about response timesWaiting for your OK

Data protection

Private data stays private

Sensitive data is labelled, and the label follows it everywhere. The platform stops it reaching the wrong place, even when a message or web page tries to trick the AI.

In practice

Customers only ever see their own accounts. An assistant that reads account data can't email it outside the company.

Analyst assistant · draft email

To: partner@outside.example

Summary of Carol's accounts and recent transactions

Contains confidential account data

Not sent: confidential data can't leave the company.

Customers see only their own accounts.

Channels

One agent on every channel

Put the same agent on Zalo, Slack, Microsoft Teams, Messenger, Telegram and your website. Every message says it was written by AI, in the customer's language.

In practice

Card Support answers in Vietnamese on Zalo and in English on the website, with the right AI notice on each.

Card Support · AI notice on each channel

  • Zalo OA Tin nhắn này do trợ lý AI soạn.
  • Slack AI-generated · Card Support
  • Website This message was written by an AI assistant.

Connections

Your tools and accounts, with consent

Connect your business systems and people's own accounts. Agents use a person's account only after they agree, and disconnecting removes access at once. Add channels, tools and checks as add-ons you can remove any time.

In practice

A workplace assistant briefs Bob from his own Gmail and calendar, only after he connects them.

Gmail and Calendar

Not connected · the assistant will ask first

Add-ons you can install and remove

  • Channels
  • Tools
  • Skills
  • Safety checks
  • Rules
  • Workflows
  • Quality checks
  • Knowledge sources

Getting started

From idea to a live agent, without a governance project

Your team builds the agent. The platform brings the review, the controls and the record.

  1. Describe the job

    Start from a template or bring an agent your team already built. Say what it does, what data it may see, what it may do and when a person must approve.

  2. Review before it runs

    Another admin reviews those permissions before the first conversation. Any later change to what the agent may do is reviewed again.

  3. Go live on your channels

    Once its checks pass, the agent is available to your teams and on the channels you choose.

  4. Improve with feedback

    Feedback and quality checks turn into suggested improvements. Each one is tried safely first, and rolls out only with your approval.

Trust and compliance

Built for the audit, not just the demo

Security and compliance teams get records they can check for themselves, and controls that hold even when the AI gets it wrong.

  • A record you can check

    Every recorded step is signed and linked to what caused it. Anyone with access can check a record for themselves, without taking our word for it.

  • Data stays where it belongs

    Confidential data can't be sent where it shouldn't go, whatever the AI was asked to do.

  • Your region, your keys

    Each organization has a home region. Its data, keys and records stay there.

  • Clear about AI

    Agent messages always say they were written by AI, in the customer's language, to help you meet transparency rules such as the EU AI Act.

What we stand by

  • Rules are enforced, not suggested.
  • Every action has an owner.
  • People stay in charge.
  • Anything you add, you can remove.

What we hold ourselves to

Targets for general availability

The bar every release is measured against. These are our engineering targets, not results from customer deployments.

from a template to a live, reviewed agent
<10 min
of agent actions on the record
100%
added by the platform to each step
<80 ms
to take back an agent's permission everywhere
<60 s

Get started

Put your agents under authority you can prove.

Bring one use case and one channel. We'll set it up with the approvals your team needs and walk you through its first records.

Book a demoSee how it works

Or write to hello@agentplatform.example