Skip to content

Agent platform

Keeping AI agents in check: approval, permissions and limits

This page shows who approves what, how far an agent may go and how you stop it. The rules apply across four modules, each in the form that suits the module.

In short

An agent proposes and prepares, and a person decides at the approval points. Permissions, limits, emergency stop and log keep its room for action narrow and make every step traceable.

Updated: November 2026

Approval points decide who has the say

Every step of a flow can carry an approval point. The agent stops there and shows a role what it intends to do. Whoever rejects must give a reason, and the reason goes into the log.

In Documents an approval can also run through several levels. Each level records person, time and comment, and a deputy steps in for absent colleagues for a set period. The same person can never decide two levels. Anyone who wants four eyes switches the principle on. If the content of a document changes, the earlier chain lapses and starts again.

A flow or an interface key cannot approve a level. That decision is reserved for people.

By module

Where approval sits

Approval by module
ModuleWhat approval looks like
ENGIVERITY One approval point per step with a role. Open approvals and messages to roles or people are collected in one inbox.
Documents Approval chains with deputies, in pairs if you wish. An action from the chat counts only after confirmation by the same person.
ERP and CRM An approval run for incoming invoices. Cancelling an outgoing invoice also needs an approval.
Ad Pilot Every proposal waits for a person as long as a tenant sits on the strictest level.

Which steps wait for approval

In ENGIVERITY agents propose. A step that writes, such as creating an open point or a report, is only queued until a person approves it. Nothing in your data changes. You enter per agent which writing steps it may exceptionally carry out without asking.

In Documents the templates that ship with it prepare emails to third parties as drafts first, and they go out only when a person sends them. The daily summary goes out by email without approval, and one template hands a checked document over by webhook. A flow you set up yourself can also send emails, call webhooks and file documents once you switch it on.

An action from the Documents chat waits for your yes. Text from a document does not trigger an action there. The AI proposal in the CRM is available on request. There every proposed value needs a verbatim quote from the source, and a proposal does not write on its own.

Ad Pilot

Levels of autonomy

Ad Pilot has three levels for agents and rules, which we set up on request. Since Ad Pilot does not write back into advertising accounts today, every proposal stays a proposal, even on the higher levels. Every new tenant starts on the strictest.

Levels at a glance
LevelWhat happens without approval
Strictest level Nothing. Every proposal waits for a person.
Middle level Changes to budget and target bids, if they can be reversed and stay within a cap. A waiting period and a quota apply. Anything that creates or activates something new still needs approval.
Highest level Everything Ad Pilot provides, without asking. Today Ad Pilot does not write back into advertising accounts, so this level is not in use.

If a value is unknown, approval applies. A change of level requires a reason. New campaigns are created paused, and switching them on is a second step taken by a person.

Permissions limit what an agent sees and does

An agent works within its limits. You decide which projects and object types it covers and whether it only reads or also writes. In ENGIVERITY permissions follow roles such as warehouse, purchasing, accounting, time recording and administrator, and each role sees its own area.

Individual employees' hours are visible only to the time recording role. This function stays blocked until data protection and the works council have recorded their confirmation.

The chat in Documents never has more permissions than the interface. An action appears as a proposal first. It is carried out only once the same person confirms and the documents concerned have not changed in the meantime.

Limits

Limiting steps, time and cost

Every module has its own brake. You set the figures per tenant or agent.

Limits by module
ModuleWhat is limited
ENGIVERITY A step limit and budget per agent, a time limit per step and a cap on repeats. A run ends at its goal, at the limit or on cancellation.
Documents A page quota per tenant and month. Each document is counted once, and an overrun runs in blocks.
Ad Pilot A quota per tenant across all changes, plus upper limits per change and a waiting period.
ERP and CRM With automatic invoice intake in the customer-specific setup a certainty threshold and a value limit apply, and a switch stops the automation. Every document then waits for a person.

The emergency stop halts everything

The emergency stop in Ad Pilot blocks changes immediately, for a single tenant or for all of them. It needs no restart. It can only block and never release anything.

A model outage counts as a normal case. In Ad Pilot a path without the model takes over, and in the ERP tasks are deferred until the model answers again. Less functionality is available then. How models and operation fit together is described on the Sovereign infrastructure page.

Log and review make every step traceable

In ENGIVERITY the step log only grows, and nothing changes afterwards. Every tool call and every start of a trigger has its own entry. A review sorts an agent's statements as supported, contradicting or without a source reference, and counts how many are supported.

Documents keeps a log for every execution of a flow and notes in the audit log when an action came from the chat. In the ERP the log records every model call, without question and answer. The overview of the agent platform shows how the logs read. How to test before launch is covered under dry run and shadow mode.

Questions and answers

Questions about controlling agents

Can an agent send emails on its own?

The templates that ship with Documents prepare emails to third parties as drafts first, and they go out only when a person sends them. Whoever switches on a self-built flow can have it send mail, call webhooks and file documents. Users without administrator rights can change recipients only to addresses of the partner or its domains. Reports to recipients you have set go out without approval. Every email that goes out appears in the audit log.

Who may start and approve an agent?

Roles govern this. In Ad Pilot, starting agents, managing approvals and switching automation on are separate permissions. In ENGIVERITY every approval point names the role that decides.

What happens if the model goes down?

The work stays in the queue and tasks are deferred. In Ad Pilot a fallback path runs without the model. Less functionality is available then.

What if an agent gets it wrong?

Reject the step at the approval point with a reason, or cancel the case. In ENGIVERITY the source reference shows where a statement comes from, and statements without a source can be recognised there.

Does an agent replace employees?

No. It takes routine off people's desks, such as reading, matching and preparing. In ENGIVERITY, steps that write into your data wait for a person's approval by default.

Who should have the last word in your process?

We set approval points, roles and limits with you before the first agent starts.