Your project, stage by stage

In accounts you own, such as your respond.io workspace and WhatsApp Business account, never on a Zelix server. Each platform is independently SOC 2 or ISO 27001 certified. How your data is handled.

Usually two or three named team members, with access scoped to your project. You can remove any of them at any time in your platform's own user settings.

No. Your conversations are never used to train AI models, never shared with other clients and never sold.

We work to Singapore's PDPA. Test conversations use made-up or anonymised details, and customer personal data stays inside the platforms you own.

The AI does not give medical, account or payment advice it has not been given. Those chats go straight to your team, with the conversation attached.

Two running costs, billed to you directly by the platforms: your respond.io plan, and Meta's WhatsApp message fees, which depend on the type and number of messages. How WhatsApp message fees work.

A deposit before kickoff, and the rest as set in your agreement. Milestone instalments are available, by card or bank transfer.

Only if you approve a change. Every new request gets a written scope with its time and cost before any work starts.

No. The 30 days of hypercare after go-live are part of your project. Some plans include a longer optimisation window, and your agreement states yours.

You do: the accounts, the data, the AI setup, the workflows and the integrations built for you. We keep only our own generic tools and methods.

No. Ongoing support after handover is optional, month to month, with 30 days' notice to stop.

Within 5 business days you get a handover pack: every account and who has access, how to remove ours, and how each automation works.

Your AI runs in accounts held in your name, so it keeps working, and the handover pack shows your team or another provider how it is built.

No. It never uses a human name or claims to be human. Asked directly, it says it is an AI in one line and carries on helping.

It never makes up a price, date, policy or availability. Every fact comes from your own information, and when it is not there the AI says it will check and hands the chat to your team.

At kickoff you tell us what it must never say or promise, such as discounts, guarantees or medical advice. We write those rules into the plan you approve and test them before launch.

It replies in the customer's language, including English, Mandarin, Malay and Tamil, and follows a switch mid-chat. The languages for your business are set and tested in your plan.

It replies when customers message you. Anything it sends first, such as a reminder, uses a message template you approved and Meta reviewed.

Send us the chat. In hypercare we fix confirmed misses within 48 hours, weekends included, check whether other chats got the same answer and add a test so it stays fixed.

We can switch it off in under a minute, and every new chat goes back to your team exactly as before.

Meta restricts numbers that send unwanted messages. Your AI answers people who message you and sends only approved templates, which keeps you inside Meta's rules.

That is why it starts small. On launch day it takes a small share of chats while we read every conversation, and it widens only as the results hold.

People try. Before launch we attack your AI on purpose, following the OWASP Top 10 for AI applications, including attempts to override its instructions, pull out other customers' details or get a promise it should not make. It has to hold before it goes live.

It answers from your business information and the current customer's own conversation. Attempts to get other customers' data are part of our adversarial tests, and a single failure blocks launch.

Platforms update the AI models they run on. When that happens, we re-run your safety tests and fix anything that now behaves differently, during hypercare and under a support arrangement.

Every instruction change is versioned and tested before it goes live, and every miss becomes a permanent test. That stops it drifting change by change.

Chats still arrive in your inbox, where your team can answer them, and our alerts tell us when the AI stops replying. The off switch also sends every chat to your team in under a minute.

Yes. Usage caps and alerts are set and checked before launch, so a fault cannot run away unnoticed.

Your industry's rules go into the plan as hard limits on what the AI may say, and we test them before launch. Regulated businesses get our fuller process, with more review at each gate.

It never poses as a human and says it is an AI when asked, in line with where AI disclosure rules are heading, including the EU AI Act. Our AI principles.

Yes. Where Zelix processes your customer data on your behalf, we sign a data processing agreement covering scope, security, breach notification and deletion. Security and data handling.

You approve what it may and may not say in the plan, and we build and test it to that plan. Any change that touches escalation, payments or personal data goes back to you for approval and is re-tested.

That is your decision, and it should not be made on launch day. Chats per person drop and the harder ones remain, so plan any change on 2 to 4 weeks of real numbers.

Always. Anyone can open a chat and take over, and the AI steps back.

We brief them before launch, show them exactly which chats will reach them, and they can take over any chat at any time.

Four or five meetings of 30 to 60 minutes across the project: kickoff, the plan walkthrough, test acceptance, the team briefing and operator training. Plus quick answers to our questions.