Command Center

#What it is

Command Center is how you ask sSystm to do something without going and finding the screen that does it. You describe what you want — set up a workspace for a new client, give someone access, draft the invoice — and the system carries it out across whichever modules are involved.

It is available at every level of sSystm. It is not an add-on, and it is not a separate assistant sitting beside your workspace. It is how the workspace is driven.

#Request, then decision

The mechanism is worth understanding, because it is the difference between a convenience feature and something you can let near a real client’s data.

The AI model does not perform actions. It asks for one. A request names an action and its arguments — nothing more. The core then decides, on its own authority, whether that request may run: who is asking, what the action requires, whether the person holds it, and whether the target is theirs to touch. Only then does the work happen, and the work is done by the module that owns it, never by the model.

The model’s output is a proposal. Authorisation is not something it participates in.

#Clearance levels

Every action Command Center can take declares the level it requires:

The last row is deliberate. A viewer is a role for reading, and Command Center is a way of changing things; giving it the lowest level would have made “lowest” mean “may write”. So the request is refused before a session is opened, rather than four steps into a loop. An unrecognised role falls the same way, which means a role added without a decision cannot run commands by accident.

Clearance is checked before an action is dispatched, and it is checked against the person who asked — not against anything the model produced. A request for a level you do not hold is refused. It is not negotiated, and rephrasing does not help.

Clearance is a security boundary, so it is included everywhere. It is not something you buy more of.

#Why we wrote this ourselves

There are libraries that do agent loops for you. We do not use one, and the reason is specific.

Those libraries run the action on your behalf. That single step — between the model asked for the permission change and the permission is changed — is the entire security model. A library that owns that step owns your authorisation.

The cost of not using one is real: we write the loop, handle a model that answers with prose instead of a proper request, and count the tokens ourselves. We took that cost knowingly, because the alternative is a security boundary maintained by someone else’s release notes.

#What to actually say

Plain sentences work. You do not need a syntax, and there is no list of commands to learn — describe the outcome you want and name the things involved.

• “Add Anna Berg at Nordkust as a lead, anna@nordkust.se.” Creates the contact. Naming the company attaches it; leaving the status out defaults to lead.
• “What’s in the pipeline this month?” Reads your deals and summarises by stage.
• “Set up a separate workspace for Nordkust, EU region.” Creating infrastructure needs admin clearance. Saying the region means it is not guessed.
• “Give erik@ourfirm.se admin on the Nordkust workspace.” Access changes need owner clearance, and you can never grant above your own level.

Two habits that help: name things the way they are named in your workspace, and say the detail you care about rather than hoping the default matches. If you leave something out it falls back to your organisation’s setting — never to a guess.

#When it says no

A refusal is information, not a bug. The usual reasons:

• “You don’t have clearance for that.” The action needs a level your role does not carry. Rephrasing will not help — ask an owner, or have them grant you the permission.
• It asks a question instead of acting. Something required was missing and it will not invent it. Answer the question and it continues.
• It says the action is waiting for approval. Anything touching your cloud account is held for a person. Nothing has happened yet; find it in Cloudflare → Approvals.
• It cannot find what you named. The record does not exist under that name in this workspace. Check spelling, and check you are in the right workspace.

What will never happen is a refusal it works around, or a claim that something succeeded when it did not.

#Being specific helps

Command Center does not guess at things you did not say. If you ask for a workspace without naming a region, it uses your organisation’s default rather than picking one — because a guess about where data lives is not a guess worth making.

The same applies throughout: say what you mean, and anything you leave out falls back to your organisation’s settings instead of being invented.

#Everything lands in the open

Work done through Command Center appears in your workspace like any other work, marked as done by AI. It shows up on the record it touched, in the place your team would have looked anyway. Nothing happens in a side channel, and nothing arrives unattributed.

If an action needs a person to sign off, it waits visibly for that person rather than proceeding and reporting success. See Permissions & control.