Controls & governance
Controls for AI agents that act in your systems
In Konnectify, an AI agent can only use the tools you add, can't change the fields you pin, pauses when it reaches its spend limit, and waits for a person wherever you require approval. Every step it takes, and why, is recorded in the Event Logs.
An instruction asks. A control decides.
You can tell an AI agent “never refund more than $500” in its instructions, and most of the time it will listen. A control is different: the refund tool won't run above $500 without a person's approval, whatever the agent decides. Konnectify uses instructions for judgement and controls for limits.
- Help customers with refund requests.
- Check the order before you refund.
“Never refund more than $500.”
Refund request: $640- The agent decides whether to refund
- Refund request: $640
Refund over $500 needs approval
Approve refund of $640?
Order #4821 · Requested by Refund agent
1Before a run
Decide what each agent may touch
Choose the level for each job
Run it as a fixed Workflow, an AI workflow with one AI step, or an AI agent. Use the least AI the job needs.
Add only the tools it needs
An agent can call only the actions you add. If “delete” isn't in its toolbox, it can't delete.
Pin fields
Set values the agent can't change, such as the Jira project, the Okta department or the Zendesk group.
Set allowed values
Limit the choices on a field, such as the ticket priorities an agent can set or the plans a customer can move to.
Write the rules down
Instructions say what the agent must and must not do, and when to hand over to a person.
Test on sample data
Run the agent on test records before it goes live.
2During a run
Stop and ask when it matters
Approval at any level
Require a person's approval before any step, in a Workflow, an AI workflow or an agent.
Approval by value
Ask only when it matters, such as refunds over an amount or downgrades above a limit.
Where your team works
Approval requests arrive in Slack or Microsoft Teams. The run waits up to 7 days for a decision.
Handover to a person
When an agent isn't sure, it passes the job to your team with a summary instead of guessing.
Run and spend limits
Each agent has limits on how many times it runs and how much it spends. When it reaches one, the agent pauses and your admin is alerted.
3After a run
See what happened and why
Event Logs
Every run is recorded: the trigger, each step, what changed in each system and who approved it.
The agent's reasoning
For each step, see why the agent chose it.
Retries and alerts
Failed steps retry 3 times. If a step still fails, an error workflow can alert your team.
Retention
Logs are kept for 30, 60 or 90 days depending on your plan, and about 6 months on Enterprise.
Which controls apply to which level
| Control | Workflow | AI workflow | AI agent |
|---|---|---|---|
| Only the tools and actions you add | Yes | Yes | Yes |
| Pinned fields and allowed values | Yes | Yes | Yes |
| Human approval, including by value | Yes | Yes | Yes |
| Test runs on sample data | Yes | Yes | Yes |
| Event Logs | Yes | Yes | Yes |
| Instructions and handover to a person | No | No | Yes |
| Reasoning for each step | No | The AI step | Every step |
| Run and spend limits per agent | No | No | Yes |
Who can see and change what
Single sign-on on every plan
Your team signs in with your identity provider.
Roles
Users build and see their own agents and workflows. Admins see and manage everything in the account.
Your data stays yours
Konnectify doesn't use your data to train AI models.
When Claude or ChatGPT uses your apps
An MCP server you build in Konnectify gives Claude or ChatGPT only the tools you add. Connections use OAuth 2.0, and every tool call is recorded in the activity log.
Reviewed by people who check these things for a living
Konnectify is SOC 2, ISO 27001, GDPR and HIPAA compliant. Our trust centre has the details your security team will ask for.
SOC 2
ISO 27001
GDPR
HIPAA
What's not available yet
We'd rather you hear it from us:
- Masking personal data in what agents read or write.Not yet
- Versions and rollback for agents.Not yet
- Approval on MCP tool calls.MCP servers limit tools and log every call, but don't yet pause for approval.Not yet
Need one of these? Tell us in your demo; it shapes what we build next.
Show your security team how it works
We'll walk your reviewers through the controls on a real agent: what it can touch, where it stops and what it records.