How It Works
Achilles turns your procedures and your systems into an assistant that can answer customers and staff, and take action on their behalf, without you writing software. This page explains what is happening behind the chat window so you can trust it, explain it to colleagues, and tell a compliance reviewer exactly where the controls sit.
Agents are configuration, not code
An Achilles agent is a description, not a program. It says who the agent is (name, tone, languages), which of your systems it may read, what actions it may take, who is allowed to do what, and when it must stop and ask a person. Agent Builder writes that description for you from your procedures and database, and you review it before anything goes live.
| Traditional software project | Achilles agent |
|---|---|
| Requirements handed to developers | Your procedures are uploaded and turned into a draft |
| Weeks of build and release cycles | Draft, review, test, publish in the console |
| Behavior changes need a code release | Behavior changes are an edit, with version history |
| Rules live in scattered code | Rules are readable in one place, and enforced by the platform |
Because the description is data, every change is versioned. You can compare versions, restore an earlier one, and see who changed what in the audit trail.
The journey of one request
Every message, from any channel, follows the same path.
- The customer asks. On your website, in WhatsApp, or by voice. The channel tells Achilles who the person is and what role they have, for example customer, agent, or manager.
- The agent understands. An AI model reads the question in the context of your agent's instructions and works out what is being asked, in the customer's language.
- Permissions are checked. Before anything is looked up, the platform checks that this person's role may use that ability, and limits results to their own records. This check is enforced by the platform itself, not by asking the AI nicely, so a clever question cannot talk its way past it.
- Your systems are consulted. The agent calls your database or service to fetch real values. Numbers such as premiums, balances, and dates always come from your systems. The AI only phrases them, so it cannot make them up.
- The answer is given. Sensitive fields are masked for roles that should not see them, and the reply goes back in plain language.
- Everything is recorded. Each step is logged for audit.
Along the way, three safety valves can interrupt the flow:
| Situation | What happens | Learn more |
|---|---|---|
| The action changes something, like a cancellation | The customer or staff member is asked to confirm first | Approvals |
| The customer asks for a person, or the situation is sensitive | The conversation is passed to your team with a reference number and summary | Human handoff |
| The question is outside the agent's scope | It declines politely and stays on topic | Safety overview |
Simple, common questions such as "what is my order status" take a shortcut that skips the deeper reasoning. They are faster and cheaper, and they follow the same permission checks.
What "memory" means
Achilles remembers the conversation you are in, not the person across conversations.
| Kind of memory | What it is | How long |
|---|---|---|
| Conversation history | The recent messages, so follow-up questions make sense | Short-lived, about half an hour of inactivity |
| Session record | Who, when, which role, how many messages, cost | Kept for your reporting |
| Audit record | Every step the agent took and every permission decision | Kept for review |
Long conversations are automatically condensed by a lower-cost model so they stay fast and affordable. Nothing is lost for audit, because the full record is kept separately.
Achilles does not build a hidden profile of your customers. When a returning customer asks about their policy, the agent looks the answer up again from your system of record, so what they hear is always current. See Personal data for how sensitive fields are protected.
Model providers at a glance
The AI model is a replaceable part. Achilles works with several providers, and you choose per agent, so your agents outlast any single vendor.
| Provider type | Good for |
|---|---|
| Leading commercial models | Best reasoning quality for complex conversations |
| Aggregator services | One account that reaches many models |
| Self-hosted or private models | Keeping data inside your own environment |
| Indian-language specialists | Strong Hindi and regional language support |
You can also set a cheaper model for routine tasks. See AI models for choosing and configuring them.
Where to go next
- Key terms for the vocabulary used across these docs.
- Quickstart to build your first agent.
- Safety overview for how controls are enforced.