Product

An AI workforce you can configure, test and trust

Orkesta combines specialised AI agents with the controls a real support operation needs: knowledge management, permissioned tools, approvals, human takeover, audit trails and analytics.

How a conversation is handled

Step 1
Receive
A message arrives on WhatsApp or website chat and is stored against the right business and customer.
Step 2
Understand
The coordinator classifies intent, urgency and risk, and decides which specialist should handle it — or a person.
Step 3
Investigate
The specialist searches your published knowledge or calls approved tools to look up the relevant record.
Step 4
Verify
Code — not the model — checks every tool call and every draft reply against what the tools actually returned.
Step 5
Resolve or escalate
The customer gets a grounded answer, a ticket reference or a request filed for approval — or a handover to your team with a summary.

The agents

Logical roles with distinct responsibilities, tools and permissions — configured per business.

Coordinator

Classifies each request (general question, order, payment, delivery, return, refund, complaint, request for a person) and its risk. Anything sensitive — threats, legal issues, fraud, abuse — goes straight to your team.

Support agent

Answers from your published FAQs and policies and cites the sources it used. If nothing published answers the question, it asks or escalates instead of improvising.

Records agent

Retrieves order, payment and delivery status from connected systems after verifying the customer. It never reveals whether a record exists to someone who cannot verify it.

Resolution agent

Opens support tickets with the facts gathered so far and files refund requests for human approval. It only confirms a ticket or request after the system has created it.

Everything around the agents

Shared inbox

Conversation history, internal notes, assignment, status and an AI activity panel showing the route, tools, sources and outcome of every automated reply.

Human takeover

One click hands a conversation to a person. Any automated reply still in progress is withheld — guaranteed at the moment of sending.

Knowledge management

FAQs, policies and uploaded PDF or text documents with versioning. Nothing is used until it is published; edits never change the live version until you publish them.

Approval queue

Sensitive actions such as refunds wait for a decision. Each request is re-checked against your records at the moment you approve it.

Integrations

Connect your store or records system through a standard connector, the website chat widget, and the official WhatsApp Business Platform.

Analytics

Resolution and escalation rates, first-response time, tool success rates, model usage and estimated cost, and the questions your knowledge base cannot yet answer.

Technology

Large language models handle language; deterministic application code owns identity, permissions, tool execution, state and audit. The model proposes — the platform decides.

Multi-model by design

Each agent can run on a different provider — OpenAI, Anthropic, DeepSeek or xAI — through one gateway with timeouts, bounded retries, explicit fallbacks and per-call usage and cost tracking.

Structured, verifiable outputs

Agents finish by calling structured tools. Replies are checked against the evidence gathered in that conversation before anything is sent.

Retrieval with provenance

Answers come from versioned knowledge with full-text retrieval; every answer records the exact document and version it relied on.

Evaluation before release

A versioned set of realistic scenarios — including prompt-injection, refund pressure and system outages — runs against every configuration change.

Versioning and rollback

Agent configurations are drafted, tested, published and rolled back like code, with a complete audit trail.

Inference roadmap

As volumes grow we are evaluating optimised inference and open-weight models to lower latency and cost, adopted only where they measurably improve quality, speed or cost.

See it on your own scenarios

Bring your FAQs and a few real customer questions — we will show you how the workforce handles them.

Book a demo Start free