Trust & control
Powerful agents. Your rules.
An agent that can change production deserves the same care as a senior engineer with the keys: limited access, approvals, a complete record, and a stop button you can reach in one tap.
How does DataDack keep an AI agent inside your limits?
Every DataDack agent runs at an autonomy level you set, from L0 observe-only to L5 fully autonomous, chosen per kind of decision and per environment. Anything irreversible - deleting data, changing access, signing contracts - is only ever proposed, and a named person decides. Permissions come from your settings, never from text an agent reads. Every action and approval is recorded and exportable.
Trust & control
Agents act.You decide how far.
Set an autonomy level for every kind of decision, per environment and budget. Prompts never grant permission - only a human can raise a level.
- Scoped by environmentDev, staging and production each get their own level.
- Scoped by money and timePer-action and monthly caps; business hours or always.
- Irreversible waits for youDeleting data, IAM changes and contracts always need a named human.
The six autonomy levels, in full
What each level permits, and where the line sits between proposing and doing.
| Level | Name | What the agent may do | Example |
|---|---|---|---|
| L0 | Observe | Look and report. Nothing changes. | Review your cloud setup and flag risks. |
| L1 | Propose | Prepare the plan or the change - a person carries it out. | Draft a migration plan for your team. |
| L2 | Act with approval | Go ahead once the right person says yes. | Release a new version to production. |
| L3 | Act and notify | Act straight away, tell you, and let you undo it. | Roll back a release that is misbehaving. |
| L4 | Act within budget | Act freely inside limits you set, and report weekly. | Run test environments under a monthly cap. |
| L5 | Fully autonomous | Handle it and keep a record. | Renew an expiring certificate. |
Careful by default
Sensible defaults.Every one yours to change.
- L5
Write code and open a pull request
Always fine - you review before it merges.
- L3
Roll back a bad release
Done at once, you are told, you can undo.
- L2
Release to production
Waits for a named approver.
- L1
Delete data, change access, sign contracts
Only ever proposed. A person decides.
When you can't be reached
Emergencies have rules too.Strict ones.
Only in a real emergency
An outage, a breach or data at risk - and only after the people who could approve have been tried.
Only the smallest safe step
Contain the damage with a step that can be undone. Never delete, grant access or commit money.
Checked before it acts
Every emergency action is independently checked against your rules before anything runs.
Reported straight away
You get a clear account of what happened and how to undo it, and a person reviews every step.
Security model
Least privilege.By construction.
Least privilege
Agents get only the access a task needs, only for as long as it needs it.
See it before it happens
Changes to your infrastructure are shown to you before they run.
Can't be talked into more
Permissions come from your settings - nothing an agent reads can raise them.
Isolated by default
Code an agent writes runs in a locked-down space, away from everything else.
A complete record
Every action and approval is kept and exportable for your auditors.
Your data stays yours
Your code, data and logs are never used to train AI models.
Hire your AI IT company.Before your next hire.
Early access opens to a small group first. Tell us what you would hand over, and we will be in touch.