Skip to content
AI Agents

An AI that knows
your business.

Grounded in your content. Trained on your tone. Honest about what it doesn't know. When confidence drops it hands off to a human with a clean summary, never a fire drill.

Under 2 s first response68 % avg. deflection rateEU-hosted, GDPR-ready
Connect your knowledge
Website URL, PDFs, text snippets, Q&A pairs, or Shopify, synced and indexed in minutes.
Set tone and guardrails
Choose a voice. Set topics to refuse. Add phrases to always use.
03
Test against real history
Run against the last 500 conversations before going live.
04
Go live with a safety net
Start in suggest-only mode. Graduate to auto as confidence grows.
AI
Confidence score
98 % · Grounded · Sources: 3
<2s
First AI response
68%
Avg. deflection rate
97%
Grounding accuracy
0
Hallucinations shipped
How it works

Your knowledge. Your tone. Your rules.

Connect your existing content and the AI is ready in under ten minutes. Retrieval-augmented generation means it only answers from what you've given it. No hallucinations, no invented policies.

  • Syncs with Shopify, your website, PDFs, text snippets, and Q&A pairs
  • Replies in the customer's language natively
  • Confidence scores logged on every single response
  • Set topics to refuse in the agent's system prompt
Knowledge sync
Website · PDFs · Shopify · Q&A pairs
97 % grounding accuracy
Only answers from verified sources
Multilingual
Replies in 100+ languages, auto-detected per visitor
Smart handoff
Below 70 % confidence → human + summary
Safety by design

Enterprise-grade guardrails, not bolt-on afterthoughts.

Every safeguard was designed in from day one, not added after a customer complaint.

PII redaction

Emails, phone numbers, IBANs and card numbers are stripped before any request leaves the EU. Nothing sensitive touches the model.

Topic guardrails

Set the topics your agent should refuse in its system prompt. Keloa politely declines and escalates to a human.

Human in the loop

Run in suggest-only mode indefinitely. Graduate one intent at a time as confidence grows.

Audit log

Every prompt, every source, every confidence score, logged, searchable, and exportable on Scale.

Multilingual replies

Auto-detects the visitor's language and replies in any of 100+ languages, same agent, same knowledge, no separate setup per language.

Smart handoff

Below 70% confidence, a human takes over with a one-paragraph summary. Never a fire drill.

Web search

When your knowledge base doesn't have the answer, agents can search a trusted set of websites you choose — and answer from them.

FAQ

AI agent questions, answered.

What does 'AI agent' mean in Keloa?

A configured AI persona with a name, a system prompt, a tone of voice, attached knowledge sources, and a set of tools (Shopify order lookup, custom HTTP, human handoff). It only acts inside the conversation, it reads messages, retrieves from your knowledge base, optionally calls configured tools, and replies. It is not an autonomous browsing agent.

How is this different from a chatbot?

A chatbot follows scripted decision trees you write. The Keloa agent reads your knowledge base, applies a system prompt, and uses judgment within the constraints you set. You do not write intents, entities, or dialogue trees. You give it your content, define tone and refusal topics, and the underlying language model handles the rest. Changes are instant: edit a Q&A pair and the next reply reflects it.

Can we control the AI's tone?

Yes. Brand voice is extracted from your website during onboarding as a starter. The most reliable way to lock tone is to add a few Q&A pairs in the exact voice you want. The system prompt is editable so you can add explicit rules ('use 'je' not 'u' in Dutch', 'never apologise', 'always sign off with the agent name'). There is no fine-tuning step.

What happens to PII in customer messages?

Email addresses, phone numbers, IBANs, and card numbers are stripped before any request leaves the platform to the language model provider. Nothing sensitive ever touches model inference. The unredacted customer message stays visible to your operators in the inbox so they can still help if the conversation escalates.

How accurate is the grounding?

Across deployed workspaces, around 97% of replies pass grounding (every claim traceable to a cited source). Replies without citations are flagged automatically. Below the configured confidence threshold (default 70%), the AI hands off to a human instead of being sent. Three layers prevent hallucinations: retrieval-only generation, citation flagging, and confidence-based handoff.

Can the AI agent call our own API?

Yes, via the custom HTTP tool. Configure URL, method, headers, and optional placeholders. The AI calls it like any other function when the description matches the customer's question. 30 second max timeout, 64 KB response cap, one automatic retry on 5xx errors. Same tool the flow builder uses.

How many AI agents can we run?

Caps are per plan: Starter 1, Growth 2, Business 4, Scale unlimited. Common patterns: one agent per language, one per channel (widget vs email), or one per brand if you run multiple stores in a single workspace. Each agent has its own system prompt, knowledge attachment, and tool configuration.

See the AI in action.

We'll train your agent on your content in under ten minutes. Test it before you commit.