A request goes in. A change you can defend comes out.
Four steps, and a person is one of them — no hidden path. The only step that can run without a person is one your own admin wrote a rule for, and it never covers a change to something already running.
You are approving a change, not a description of one.
The common failure here is subtle: you approve a summary, and something adjacent to it runs. Here the thing you reviewed is the thing that executes — there is no interpretation step between your yes and the work.
- What it creates, changes or removes — stated plainly, before you decide
- What it will cost, checked against the ceiling you set
- Whether anything existing is replaced or destroyed, called out rather than buried
- Change the request and the previous approval does not carry over
A well-worded request is not an argument.
Your rules are evaluated against what the change would actually do, not how persuasively it was requested. No phrasing talks its way past a limit.
- Spending limits, permitted regions and your admin’s own rules — policy code or ready-made packs — apply before a person is asked to look, and can only tighten the platform’s
- Nobody approves their own request: the person who asked and the person who approves are two people
- No agent or automation approves its own work. The only automatic approval is a rule your admin wrote, and it never covers changing, deleting or replacing what is already running
- Deleting or replacing anything that holds data requires an elevated approver, regardless of any other outcome
- Approvals and refusals are both recorded, with who and when
Verification does not stop at the first successful change.
Success tells you the request was accepted. It does not tell you your infrastructure matches what you asked for — or that it still will next month.
Confirmed at the time
What is actually running is compared against what was requested. A change is not finished because a command exited cleanly.
Checked from then on
Someone edits a setting in a console, a vendor default shifts. Ongoing checks notice when reality has quietly stopped matching your intent.
Corrected the same way
When something has drifted, Intentric prepares the correction and waits. It goes through the same approval as the original request, and is never applied on its own.
What holds true at every step.
Not features you enable. Properties of the way the thing works.
Access
- No AI agent ever holds standing write access to your cloud
- Access is granted for approved work, and expires with it
- Nothing runs under a shared, permanent service account
- The moment access is granted is itself part of the record
Record
- Who asked, who approved, what changed and what resulted — recorded as it happens
- Nothing is reconstructed from logs after someone starts asking questions
- Exported as CSV or JSON or sent to your webhook, and readable by people who were not there
- Refusals are recorded as carefully as approvals
Reading about it only gets you so far.
Run the whole path in the sandbox against a simulated AWS, or walk through a worked example of what your reviewer would see.