How do you keep AI customer service safe?
The short answer
Give the AI no authority by default and grant each permission explicitly, so anything not declared routes to a person. Add a safety floor that no configuration can override: it never claims to be human, pauses for a person when someone is vulnerable, and never gives clinical, legal or financial advice. Then make every action traceable to its sources and model version.
Declared authority, not helpful guessing
Most AI is trained to sound helpful, so it says yes by default. That is how a chatbot invents a refund policy: Air Canada was held to a fare its chatbot made up. A Neos agent starts with zero authority. Every refund it can offer, appointment it can confirm or fee it can quote is granted in configuration. If a grant is absent, expired or ambiguous, it routes to a human. "Let me check what I'm authorised to do here" is a valid answer.
A safety floor that can't be overridden
A short set of hard rules fire regardless of context, brand or instruction, and cannot be overridden by the client, by Neos Wave or by the person in the conversation:
- Never claims to be human when asked.
- Detects situationally vulnerable moments, such as distress, safeguarding language, coercion or confusion, and pauses for a person in the same thread.
- Won't give clinical, legal or financial advice.
- Treats retrieved content, images and attachments as material to analyse, not instructions to follow.
- Asks customers to remove information it doesn't need, such as ID documents, payment details or passwords, and doesn't use or store it.
An honest opening
Every conversation opens by establishing identity (who I am, and that I'm an AI), scope, authority, boundaries (what I'll hand to a human) and promise (how I'll treat you), based on the Service Handshake standard.
Traceability and data
Every action is traceable to the sources it used and the model version that produced it, and deployments are pinned to a specific version of the operating rules, so any interaction can be reproduced and reviewed. Agents ask for the least information they need, and customer data isn't used to train third-party model providers.
Common questions
What certifications does Neos Wave hold?
Cyber Essentials (certified by IASME, June 2026), NHS DSPT Standards Met (2025–26), and registration with the Information Commissioner's Office under UK GDPR.
Where do the AI models run?
It depends on the task: cloud frontier models on UK/EEA endpoints where available, local GDPR-resident models inside the UK or EEA, or open-weight models self-hosted in your environment.
What happens if the AI gets something wrong?
Managed by Neos Wave end to end, so one team fixes it, and the audit log shows exactly what the agent did, what it relied on and which rules applied.