Frequently Asked Questions

These are the questions we hear most from teams at banks and insurers. If yours is not here, the troubleshooting guide covers problems you may hit, and Key terms explains the vocabulary.

Accuracy and trust

How accurate are the answers?

For questions about your customers' data, such as a policy status or a balance, the agent fetches the value from your system and only phrases it, so the facts are as accurate as your system. For procedure questions, accuracy depends on how clearly your written procedures are captured. Test your agent with real questions before launch to measure it.

Can it make things up?

The platform forbids the agent from inventing amounts, dates, names, or policy numbers, and figures always come from your systems. Like any AI, it can occasionally phrase something imperfectly, which is why you test, review sessions in Monitor, and use approvals for anything that changes a record.

What stops it from showing one customer another customer's data?

Data limits are enforced by the platform on every lookup, based on the signed identity of the person chatting. They are not instructions to the AI, so they cannot be talked around. See the safety overview.

What happens when it does not know the answer?

It says so and, if you have enabled handoff, offers to pass the conversation to your team. It is designed to decline rather than guess. See human handoff.

Privacy and compliance

Where does our data go?

The agent reads your systems live and does not copy them into Achilles. The conversation and the results needed to answer are sent to the AI model provider you choose. For strict residency, use a self-hosted model. See AI models and self-hosting.

How is personal data protected?

Sensitive fields can be masked or hidden per role before the AI ever sees them. Only roles you designate see full values. Details are in personal data.

Can we show an auditor what the agent did?

Yes. Every conversation, permission decision, action, and configuration change is recorded and can be reviewed. See the audit trail.

Capabilities

Which languages does it support?

It follows the customer's language and works well in English and Hindi, with specialist models available for other Indian languages. Test with real customer phrasing in each language you plan to offer, including mixed-language messages.

Which channels can we use?

Your website through the widget, voice, and WhatsApp and Instagram through messaging channels. Email is outbound only, for notifications and handoff.

Can it take actions, like cancelling or updating something?

Yes, if you grant that ability and choose the roles allowed to use it. By default any action that changes a record asks the person to confirm first. See approvals and agent abilities.

Can it hand a customer to a person?

Yes. Handoff can trigger on phrases like "I want to speak to someone", on sensitive results such as a high-value claim, or after repeated failures to resolve. Your team is notified with a reference number and summary. See human handoff.

Does it remember returning customers?

Within a conversation, yes. Across conversations, no: it looks up current information again each time. That keeps answers fresh and avoids storing extra personal data. See how it works.

Cost and control

How do we control cost?

Set a cheaper model for simple lookups, cap the reply length, and keep conversations focused with a clear scope. The console overview shows usage and estimated cost so you can spot changes early. See AI models.

Can different staff see different things?

Yes. Roles decide which abilities a person can use, which records they can see, and which fields are masked. Learn more in roles and access and users and teams.

Getting live

How long does go-live take?

A first working agent often takes a day or two: upload procedures, connect a read-only data source, and review the draft with Agent Builder. Allow a further week or two for testing with real questions, a compliance review, and the website or channel setup. Follow the quickstart.

What do we need from IT?

Four things: a read-only account for the systems the agent should read, a small code change to place the widget and issue signed user tokens, network access between the platform and your systems, and, for messaging, a webhook connection. Give them the integration overview. Building, testing, and monitoring need no IT help.

Can we run it in our own environment?

Yes. See self-hosting.

How do we change the agent after launch?

Edit it in the console, test, and publish. Each change is versioned, so you can compare and roll back. Use Monitor to find conversations that need improvement.

Who can change the agent?

Workspace members according to their permissions, and every change is attributed in the audit trail. Manage members in users and teams.