How Chakusa uses AI

Where AI shows up, exactly which actions it can take, what it cannot do, and an honest account of what's still a gap; not a disclosure written to sound resolved.

Engineering draft; not attorney-reviewed Last updated: Draft dated 2026-09-01

Grounded in a full audit of src/lib/ai/* as of 2026-09-01. See the two open engineering gaps this page describes below rather than papering over: no consent gate on the customer-facing assistant, and unredacted admin visibility into AI transcripts.

1. Where AI shows up

Two places. First, a business can turn on an AI assistant that drafts or sends replies to that business's own customers over SMS or WhatsApp. Second, customers can chat directly with an in-app AI assistant inside the Chakusa app.

2. How it works

When either feature is used, the relevant conversation is sent to a third-party AI provider, either OpenAI or Anthropic depending on how the system is configured, to generate a response. The conversation may include context such as the customer's name, notes on file, and recent appointments for the business-facing assistant.

3. What the AI can actually do

The business-facing assistant doesn't have open-ended access to your account; it can only call a fixed, named set of actions, each scoped to the business it's operating for:

Action	What it does
`check_availability`	Looks up open appointment slots for a service between two times; read-only, never changes anything.
`book_appointment`	Books an appointment for the customer at a specific time.
`reschedule_appointment`	Moves an existing appointment to a new time.
`cancel_appointment`	Cancels an existing appointment.
`get_business_info`	Reads the business's own profile, services, and working hours; read-only.
`answer_faq`	Retrieves stored knowledge relevant to a customer's question; read-only.
`request_information`	Asks the customer a clarifying question before proceeding; read-only.
`escalate_to_human`	Hands the conversation to a person on the business's team.

Anything outside this list simply isn't possible for the AI to do; it's not a matter of the AI "choosing" not to.

4. Draft, Approval, and Autonomous modes

A business chooses how much independence the assistant has, from three modes:

Draft (the default for every business). The AI prepares a reply or action; a person on the business's team reviews and sends it. Nothing reaches a customer without a human tap.
Approval. Specific action types (a reply, a quote, an appointment change, a workflow) each require a person's sign-off before they take effect; configurable per business.
Autonomous. A business can opt into letting some actions proceed without per-message approval. Even in this mode, payment-related actions are always blocked by policy; that rule doesn't have a mode that turns it off.

A business can also set quiet hours, block specific topics, and restrict which channels (SMS, WhatsApp, email) the assistant is allowed to use; the assistant respects those settings on every message, in every mode.

5. What it cannot do

Guarantee its answers are accurate
Replace professional, medical, legal, or financial advice
Take a payment, issue a refund, or touch billing; payment-related actions are hard-denied by policy in every mode, not just held for approval
Make a booking, payment, or account change on its own outside what a business has explicitly configured it to do
Act without limits; every response is checked against automated rules before it's sent

6. Automated safety checks

Before a response is generated or sent, an automated system checks for things like attempts to manipulate the AI's instructions, content that looks like it's leaking sensitive information, and language patterns associated with unsafe or inappropriate advice. These are rule-based pattern checks, not a human reviewing every message, and not a second AI judging the first one's safety.

A message can also be automatically held for a human to approve, or escalated to a person entirely, based on confidence and the business's own rules (quiet hours, blocked topics, or a customer who has opted out of contact). A business, or a customer asking to speak with a person, can always take over from the AI. There's also a platform-wide switch that can immediately turn AI features off across all of Chakusa if needed.

8. Who can see a conversation

The business whose customer is messaging (for the business-facing assistant) can see that conversation, the same as any other message in their account.

For the in-app customer assistant, Chakusa staff with admin access can view the full conversation content for support and safety purposes; as of this disclosure, that view is not redacted. Every admin view is intended to be logged.

9. How long it's kept

AI conversation content is currently kept indefinitely rather than deleted on a fixed schedule. Some system-generated "memory" used to personalize responses is deleted automatically after it expires; not all of it currently has an expiry set. A customer can turn off personalization and conversation memory in their app settings (Section 11).

10. Providers

OpenAI and Anthropic. Neither is instructed to use your conversations to train their own models beyond what their own standard API terms provide; we haven't independently audited that claim beyond relying on those providers' published terms.

11. Your controls

As a customer: ask to speak with a person at any time; the AI hands off. Turn off personalization and conversation memory in app settings.
As a business: turn the assistant off entirely, switch modes (Section 4), set quiet hours and blocked topics, and review anything held for approval before it sends.

12. Changes & questions

If how our AI features work changes materially, especially around consent or data handling, we'll update this page and the date at the top.

Questions about how AI is used privacy@chakusarecovery.com

Email us