Answers from your knowledge
Replies from your website, documents and articles, and says so when it doesn’t know.
Technology · AI Agents
Agents that act, not just answer.
Mevrik AI Agent understands what a customer wants, uses your tools to get it done, and knows when to bring in a person. You set it up in minutes in AI Agent Studio: no code, no flowcharts.
One turn, start to finishagent harness
- Customer
- “Can I move my appointment to Friday?”
- Reason
- Known customer. Changing a booking needs identity checked first.
- Tools
- Google Calendar · find free times on Friday
- Decide
- Offer three times, confirm before changing anything.
- Reply
- “Friday has 10:30, 14:00 or 16:30. Which suits you?”
6named building blocks behind every answer
3ready-made roles: Support, Sales, Lead-gen
4steps in every turn: reason, tools, observe, decide
100+languages, in your brand’s tone of voice
What it does
One agent, many jobs.
Gets things done
Looks up orders, books and moves appointments, sends checkout links and captures leads.
Remembers each customer
Keeps facts like preferences and past purchases, so returning customers aren’t strangers.
Asks before it acts
Confirms before cancelling, refunding or booking, and checks identity before sharing account details.
Hands over gracefully
Passes the conversation to your team with a summary when a person is the better answer.
Speaks their language
Replies in the customer’s language, in your brand’s tone of voice.
Ready-made roles
Start with a role. Shape it to your business.
Each role arrives with sensible skills and a tone of voice. Change the name, the greeting and the skills it uses.
Resolves questions end to end.
Uses every skill you enable: answers, order lookups, bookings, tasks and hand-over.
- Grounded answers from your Knowledge Hub
- Order tracking and appointment changes
- Creates tasks for anything your team must finish
Support agentrole
- Skills
- Knowledge answers · Shopify · Google Calendar · Tasks · Hand-over
- Tone
- Clear, calm and helpful
- Goal
- Resolve the issue in the conversation
Under the hood
Every turn runs through the Agent Harness.
The same engine runs in AX Growth and AX Enterprise. It is built on six named building blocks, so you can see how each answer was made.
- 01
Reason
Reads the message, the customer’s history and what the agent knows about your business.
- 02
Use tools
Calls only the tools you have connected and switched on for this agent.
- 03
Observe
Checks what the tools returned before it says anything.
- 04
Decide
Replies, asks for approval, or hands over to your team.
Six building blocks
Named parts, not a black box.
Tools
Typed actions over MCP, REST and built-in connectors. Each agent only gets the tools you allow.
Skills
Versioned playbooks for jobs like booking, lead capture or hand-over.
Hooks
Safety checks on every turn, such as personal-data redaction and audit logging. The defaults can’t be removed.
Memory
What the agent knows about your business and each customer, with erasure on request.
Context
Picks the right facts for each turn and keeps conversations within budget.
Model Hub
Routes each step to a suitable model. Open to any model.
Use cases
What teams use it for.
Order questions, solved
WHEN“Where’s my order?”
AI AGENTChecks identity, looks up the order and shares tracking
RESULTAnswered in seconds, no ticket
Bookings at midnight
WHEN“Any appointments tomorrow?”
AI AGENTReads the live calendar, books and confirms
RESULTConfirmation by email and SMS
Leads while you sleep
WHENVisitor asks about pricing
AI AGENTQualifies the lead and books a call
RESULTMeeting in the calendar, lead in Contacts
Why it’s different
A chatbot answers. An AI agent gets it done.
| Typical chatbot | Mevrik AI Agent | |
|---|---|---|
| Setup | Decision trees and scripted flows | A role, your website and the tools you connect |
| Questions it can’t match | “Sorry, I didn’t understand” | Reasons about the question, or says it doesn’t know and offers a person |
| Taking action | Sends a link to a form | Books, looks up orders, builds carts and captures leads in the chat |
| Returning customers | Starts from zero every time | Remembers preferences, purchases and open issues |
| Risky actions | Not possible, or no safeguards | Confirms first; refunds over your limit wait for approval |
| Seeing why | A black box | Every step recorded in Trace |
AI Agent Studio
Five tabs. One agent you’re proud of.
Everything about an agent lives in one place. Draft it, test it in preview, publish it, and roll back if you need to.
- 01
Profile
Name, role, greeting, avatar and tone of voice.
- 02
Skills
Switch on what it can do: answers, bookings, orders, leads, tasks, hand-over.
- 03
Knowledge
Choose which folders and pages this agent can use.
- 04
Memory
See what it has learned about your business and your customers.
- 05
Deploy
Pick channels and publish. Every version is kept.
Safe by default
Helpful, but never reckless.
Some rules are built in, whatever you configure.
- It confirms with the customer before cancelling, refunding or booking
- It checks who it’s talking to before sharing order or appointment details
- It only uses tools you have connected and switched on
- When a helper agent works on a sub-task, it can look things up but can’t change anything
Before a cancellationbuilt-in
- Customer
- “Cancel my order #4821.”
- Identity
- Checked against the order’s email
- Confirm
- “Just to be sure: cancel order #4821 for £86?”
- Customer
- “Yes, please.”
- Done
- Order cancelled · confirmation sent
It learns your business
An agent that knows your business like a team member.
Alongside your Knowledge Hub, the agent keeps its own notes about how your business works.
- Learns facts about your business from real conversations
- A fact must come up in different conversations before it is kept
- It refuses to store personal details in its business notes
- Your team’s notes always outrank what the agent concluded
Agent notes · Brightside Dentalmemory
- Learned
- Saturday appointments are mornings only
- Learned
- Parking is free after 6pm
- Rejected
- A customer’s phone number: personal data stays out
- Your note
- Whitening needs a check-up first
Guardrails
Guardrails on every single turn.
Thirteen checks run around every reply. The defaults can’t be switched off by anyone, and Enterprise can add its own.
Personal data redaction
Card numbers and personal details are removed before the model sees them, and checked again on the way out.
Prompt-injection filter
Attempts to trick the agent with hidden instructions are caught before they reach the model.
Tool permissions
The agent can only use tools that are connected, enabled and allowed for it.
Hallucination check
Replies are checked against your knowledge and data before they’re sent.
Brand voice check
Every reply is checked against the tone you set.
Rate, budget and cost limits
Limits per workspace stop runaway costs or abuse.
How guardrails work
Checks before, during and after every reply.
- 01
Before
Kill switch, personal-data redaction, injection filter, permissions and budget checks.
- 02
During
Tool inputs and outputs are validated as the agent works.
- 03
After
Brand voice, hallucination check, outgoing data check and cost meter.
- 04
Always
Every step is written to Trace, and learning signals go to Reflex.
Understands more than text
It listens, looks and acts on what customers send.
Customers send voice notes, photos, screenshots and documents. The agent understands them and takes the next step.
Hears what customers say.
A WhatsApp or Messenger voice note is turned into text, and the agent replies to what was said.
- Transcribes voice notes in the customer’s language
- Shows your team the transcript in the thread
- Acts on the request, like booking or tracking an order
WhatsApp voice note0:12
- Heard
- “Hi, can I move tomorrow’s appointment to Friday afternoon?”
- Tools
- Google Calendar · free times on Friday
- Reply
- “Friday has 14:00 or 16:30. Which suits you?”
- Done
- Moved to Friday 14:00 · confirmation sent
Learns like a new team member
It asks when it should. It remembers what you decided.
Like a new hire, the agent checks with your team when something needs a decision. Your answers and replies become lessons for next time.
- 01
Asks for approval
When an action needs a person, it creates a task and tells the customer what happens next.
- 02
Your team decides
Approve, decline or reply yourself. The agent watches how it was handled.
- 03
It learns
Approvals, rejections and your team’s replies become signals. Facts about your business are added to its notes.
- 04
Next time
It handles similar cases the way your team did, within the rules you’ve set.
From approval to habit
The first time it asks. Next time it knows.
Learning never bypasses your rules. Anything involving money, tools or personal data still follows your approval settings, and bigger changes are proposed for you to approve.
- Facts must come up in different conversations before they’re kept
- Your team’s notes always outrank the agent’s own conclusions
- Reflex proposes improvements; it never rewrites safety settings
- Every lesson can be seen, edited or removed
Learning from a decisionreflex
- First time
- “Can I swap my size after 30 days?” → asks your team
- Your team
- “Yes, for unworn items, one swap.”
- Learned
- Size swaps after 30 days: unworn items, once
- Next time
- Answers straight away, in line with your team
Growth or Enterprise
Same agent. Enterprise adds depth.
AX Growth
For growing businesses. Live in minutes.
- Mevrik AI Agent with Support, Sales and Lead-gen roles
- AI Agent Studio: profile, skills, knowledge, memory and deploy
- Ready-made skills for Shopify, calendars, leads and tasks
- Draft and publish, with version history
AX Enterprise adds
Same platform, more depth and control.
- Custom skills and the full Studio
- Workflows and routing between agents
- Agent Sandbox to test before you publish
- Bring your own model
Questions about AI Agents.
Do I need to write prompts or code?
No. Pick a role, connect your website and tools, and switch skills on or off. Aura can set it up with you in a chat.
Can it take actions without asking?
Only the ones you allow. Cancellations, refunds and bookings are confirmed first, and refunds above your limit wait for a person.
What if it doesn’t know the answer?
It says so and offers a person, rather than guessing.
Your AI agent, live in 5 minutes.
Start free, connect your channels, and watch it resolve real conversations today.