Skip to content
Trustworthy AI

An AI that followsyour rules

Before talking to a customer, the agent learns from your company, goes through tests and goes live under control. After that, every reply can be explained.

Tolky

WhatsApp service

I bought the annual plan yesterday. Can I cancel and get my money back?
Yes, you can. Under our policy, purchases made within the last 7 days can be canceled for a full refund. Would you like me to open the request?
Yes. And tell me which stock will go up tomorrow.
I can't help with the stock market, that's outside what I handle. I've already opened your cancellation, and someone from finance will confirm the refund with you.
Handed off to finance with the history

Investigate reply

Instructions used

Support mission · Scope and limits · Don't make things up

Knowledge base excerpt

Cancellation policy, section 2: within 7 days, full refund

Decision

Request human service from finance

Decision confirmed by a second AI
From draft to conversation

Teach, test, publish and monitor

Step 01

Teach

Knowledge base, mission and the agent's limits.

Step 02

Test

Simulations with personas and success criteria.

Step 03

Publish

Draft, comparison and restore of the previous version.

Step 04

Monitor

Live conversations and the trail of every reply.

  1. Step 01

    Teach

    Knowledge base, mission and the agent's limits.

  2. Step 02

    Test

    Simulations with personas and success criteria.

  3. Step 03

    Publish

    Draft, comparison and restore of the previous version.

  4. Step 04

    Monitor

    Live conversations and the trail of every reply.

In detail

What happens at each step

  1. 01

    Teach

    The AI learns from what your company already knows

    You feed the knowledge base with files, links and text and define the agent's identity: mission, context, personality and tone. Governance blocks say what it can do, what it must refuse and when to call someone.

    • Knowledge base with files, links and text

    • Support, sales, engagement, education and productivity missions

    • Scope, refusal policies and quality criteria

    • Uncertainty rule: when in doubt, the agent doesn't make things up

  2. 02

    Test

    Simulate conversations before going live

    In Simulations, you create personas and scenarios, define what counts as success and run conversations in batches. The report shows, turn by turn, where the agent got it right and where it needs adjusting.

    • Custom personas and scenarios

    • Success criteria per scenario

    • Batch conversation runs

    • Turn-by-turn report, with evidence

  3. 03

    Publish

    Changes with a draft, comparison and rollback

    Adjustments go to a draft environment first. You compare it with what's live, publish with automatic backup and, if needed, restore the previous version. Each identity version records what changed.

    • Draft environments separate from production

    • Draft compared with the live version

    • Automatic backup on publish, and restore

    • Version history with a description of each change

  4. 04

    Monitor

    Every conversation shows what to improve

    You follow conversations live and investigate any reply. Each conversation gets a summary, topic and sentiment, NPS is sent at the end and the dashboard shows how much the AI resolved alone and how much needed human support.

    • Live conversations, with AI pause and resume

    • Investigate: instructions, excerpts and calls behind every reply

    • Summary, topic and sentiment for each conversation

    • Automatic NPS at the end of service

Limits

What the agent respects in every conversation

Rules that apply to every message, not just on setup day.

Out of scope, it redirects

Questions unrelated to the agent's mission get a polite redirect.

When in doubt, it doesn't make things up

The agent is instructed to say it will check or call someone, instead of inventing an answer.

A second AI double-checks

On the decisions you choose, a second AI confirms the action before it runs.

Fallback model

You set a reliability threshold and a fallback model for when it isn't met.

People step in by rule

You write when the conversation goes to someone on the team, and it arrives with the full history.

The AI steps aside when someone steps in

The operator can pause the AI at any time. With WhatsApp coexistence, if someone replies from the phone, the AI pauses.

People in control

The AI serves, people decide

Operators see conversations as they happen, pause the AI, take over and hand back. Queues distribute cases by department and show why each one went to whom.

  • Pause and resume the AI in any conversation

  • Queues and automatic distribution by department

  • An explanation of why each case went to each person

  • An introduction message when someone takes over

Frequently asked questions

What people ask most about trustworthy AI

See the AI followingyour operation's rules

In the demo, we build an agent with your rules, simulate conversations and show the trail behind every reply.