How we use AI
Last updated 2 November 2026
Myla uses AI to help businesses answer their customers. This page explains what it does, where people stay in control, and what we do to keep it honest and safe. It is written for the businesses that use Myla and for anyone who messages them.
What the AI does
| Task | What happens |
|---|---|
| Drafting replies | Reads the conversation, the business's knowledge sources and, where connected, the customer's order or booking, and suggests a reply. |
| Writing help | When a team member asks, rewrites the reply they wrote (for example friendlier, shorter or with spelling fixed), follows a rewrite instruction an admin has saved for the business, or translates it. The person still decides what is sent. |
| Checking reply drafts | Checks details against the sources used for the draft. A second AI provider, chosen from Anthropic and OpenAI, checks selected finished drafts for unsupported claims, invented policies or promises, missing answers and claims about actions that have not happened. Where available, it uses a different provider from the one that wrote the reply. |
| Sorting conversations | Suggests a topic, priority and language so the right person sees it. |
| Voice notes, calls, images and videos | Transcribes voice notes and describes images and videos so they can be read and answered. Where a business turns on call transcripts, it summarises each call for the team, and transcribes the call itself if WhatsApp's own transcript is unavailable. |
| Actions | Can propose things like a refund, a delivery update or a class booking. These follow the rules in the next section. |
| Summaries and reports | Summarises long conversations and answers questions about the business's own figures. |
Where people stay in control
- By default, every AI draft is reviewed by a person before it is sent. A business can choose to let some replies send automatically, channel by channel, and can switch that off at any time.
- With automatic sending on, replies about sensitive subjects such as refunds, cancellations and complaints wait longer before they go, so a person can step in. Refunds above the business’s limit, and messages that try to give the AI instructions, always wait for a person.
- Actions that change an order or booking need a person’s approval unless the business chooses otherwise, and are checked against live data before they run.
- Every AI reply and action records why it was made: the sources it used, how confident it was and which model produced it. Staff can see this in the conversation.
- Draft fact checking starts in watch mode, which records results without changing sending. If the platform owner turns on enforcement, a reply intended for automatic sending can be corrected once and checked again. A remaining issue or an unavailable checker leaves the draft for a person. Drafts awaiting human approval are never blocked by this check.
Telling customers they're talking to AI
When a reply is sent automatically, Myla tells the customer once in that conversation that it was written by AI and that a person can step in. Businesses can change the wording but it is on by default. This follows the transparency rules in the EU AI Act and similar laws.
Your data and the AI providers
- We use AI providers listed on our sub-processor page under API terms that do not allow them to train their models on the content we send.
- We do not train our own models on conversations, and one business’s conversations are never used to answer another business’s customers.
- When drafting, the AI can look at how the same business’s team answered similar questions before, with names, contact details and order numbers removed. Each business can switch this off.
- Card numbers, national ID numbers and bank account numbers are removed before text is sent to an AI provider.
- Anonymous benchmarks compare figures such as reply times across at least five businesses. They never include messages, names or contact details.
Limits
AI can be wrong. It can misread a message or state something that is not in the business’s sources. That is why drafts are reviewed by default, why answers are grounded in the business’s own material, and why we test changes against real past conversations before they go live. The business remains responsible for what it sends to its customers.
Myla is not designed for decisions with legal or similarly significant effects on a person, such as credit, employment or insurance decisions, and must not be used for them.
Questions or concerns
If you messaged a business and want a person instead of AI, say so in the conversation. For anything else about how Myla uses AI, write to contact@myla.cx.