- of transactions checked for fraud and gaming
- 100%
ControlKnow every answer is right.
Tested before release. Scored every day. Hosted where your data must stay.
Before a customer meets a new version, it has passed your test set. After it goes live, every conversation is scored. You see the score, the replays and the reasons.
The right model for every step of your workflow.
Each step runs on the model that performs best for that task, tested on your own conversations. When a better model appears, we test it and switch.
Example conversation
Model
Understand a Darija voice note
« Salam, bghit nkhalless la commande dyali, fin wslat ? »
Local modelBest on Darija in our production data
Find the order in your system
Order API: #48213, out for delivery, 249 MAD to pay
Chosen by testBest score on API calls for this agent
Decide: resolve or hand over
Within the rules: the agent resolves it
Chosen by testBest score on your handover rules
Reply in French with a payment link
Bonjour ! Votre commande #48213 est en route. Voici le lien pour régler les 249 MAD.
Chosen by testBest score on French replies for this agent
The rules your agent is scored on.
You write the rules with us: what the agent must check, say, never say and when it hands over. Every conversation is scored against each one, and you see the pass rate per rule, version after version.
Agent quality · one client agent
v25 → v32
Overall score
92%
Checks
What Control does
A testing lab before every release.
50 synthetic customers, built from your real conversations, test each new version. A regression is caught before it reaches anyone.
Testing lab · v32
Payout delayed, voice note in DarijaCorrect amount and date
PassAngry partner asks for a managerHanded over with context
PassRefund outside the policyRefused politely, offered a credit
PassMenu photo with a price errorFixed in v32 after a regression
Fixed
Every conversation scored in production.
In production, every conversation is scored against each rule. You follow the score release after release, with the replays and the reasons.
Quality score · one client agent
75% to 92%
quality across 8 releases on one client agent
When it isn’t sure, it hands over.
When a request falls outside its rules or the agent isn’t sure of the answer, it passes the conversation to your team with the full context, and tells the customer someone is on it.
Handover
Conversation #20944
- Reason
- Refund above the agent’s limit
- Context
- Full conversation, order #48213
- Assigned to
- Your support team
- Customer told
- Someone from the team is on it
Handed over
An improvement loop you sign off.
Every failed check and every handover becomes a proposed fix: a new rule, a new test or a knowledge update. Nothing goes live without your approval.
Proposed fixes · v33
New rule: confirm the address before bookingFrom 3 failed checks this week
Awaiting youNew test: refund asked in ArabiziFrom one handover
ApprovedKnowledge update: new delivery zonesFrom 12 handovers
Approved
Hosted in the cloud by default, or on your own infrastructure when data must stay in the country.
Trust & security
Replaces QA on a 2% sample, black-box chatbots, and surprises in production.
Our team builds it with yours, runs it, and improves it every month.
How we workControl in production at our clients
Where Control runs and what it connects to
Channels
Systems
- Zoho
- Odoo
- HubSpot
- Salesforce
- Airtable
- Zendesk
- Google Sheets
- All integrations
Questions we hear from IT and compliance teams
Where is our data hosted?
In the cloud by default, or on your own infrastructure. See Trust & security.
Which AI models touch our data?
The choice is documented for each use case. Self-hosted models are available when data cannot leave the country.
Can we audit what the agent said and why?
Yes. Every reply is linked to its version and sources.
See what an agent would handle for your team.
Book a demo with our team: thirty minutes to look at your channels and your volumes, and to show you agents running in production.