AI guardrails for agents: filters, forbidden topics, fact checks and human handoff
We put a layer of checks between the model and your customer. It catches forbidden topics, invented prices and personal data, and hands hard cases to a person.
What hurts
The bot promised a customer a discount that does not exist, or quoted a price that is wrong. The customer took a screenshot, and now you either honour it or argue.
A clinic fears the agent will start giving medical advice. A bank fears the bot will comment on a credit decision. A prompt saying "do not talk about this" does not always work.
When the bot does not know the answer, it makes one up, confidently and smoothly. The customer cannot tell fact from fiction, and your manager learns about it from a complaint.
What we build
- A list of forbidden topics, promises and phrasing agreed with you
- Input filters against attacks and abuse
- Output checks before sending: prices, discounts and facts from your data
- Masking of personal data in answers and logs
- Handoff rules: topic, customer emotion and model uncertainty
- A provocation test set for regular runs
- A dashboard of guardrail triggers and false blocks
What you get
- A guardrail layer built into your chatbot or voice agent
- Documented rules you can change without a developer
- A provocation test set with results
- A trigger dashboard with alerts configured
- A guide for managers who take over handed-off conversations
How we measure the result
- Number of answers stopped or corrected by guardrails
- Share of normal answers blocked by mistake
- Number of complaints about false information from the bot
- Share of conversations handed to a person, and the reasons
Systems and stack
Industries
Questions
How much do guardrails cost?
From $5,000 over 3–5 weeks. The price depends on the number of rules and channels, and whether facts must be checked against your product or price data. Voice agents add a speed requirement, because checks cannot delay the reply.
Will the bot start refusing everything?
That is a real risk of overly strict filters. So we measure both blocked bad answers and wrongly blocked good ones. Thresholds are tuned on your real conversations, not guessed.
Do guardrails guarantee the bot never makes mistakes?
No. They reduce risk significantly, but language models sometimes make mistakes. So critical items such as prices and discounts are checked in code against your data, and disputed cases go to a person.
Which bots does this work with?
Website chat, WhatsApp, Telegram and Instagram bots, and voice agents over SIP. Any model: OpenAI, Gemini, Anthropic or self-hosted. We run test suites on the Happ platform agents ourselves before every change.
What do you need from us?
A list of things the bot must not say, access to price and product data, sample real conversations and one person who approves the rules. We sign an NDA before access.
What happens after launch?
New products and promotions change the rules. The rules interface lets you update them without a developer. We can also review triggers and tune the filters under an SLA.
Related services
More in this section
Describe the task. We reply within one business day, the first call is with the founder.
Describe the task. The first call is with the founder.
We reply within one business day. 30 minutes to tell you whether it is worth doing and what it costs.