Your AI agents work, you stay in control

Yes, and denying it would be a lie. It can misread a document, misapply a rule, or invent a plausible answer when information is missing (what is called a hallucination). We design it so that errors are rare, visible and can be put right:

• it is checked on your real cases before go-live (15 to 30 cases for a Process Agent);
• it answers from your sources, and hands over when it does not know;
• whatever commits the company is approved by a person;
• every decision leaves a record and, with Monitoring, how it works is followed over time.

The acceptable level of error depends on the cost of an error: each action is therefore classified by what is at stake.

You do, depending on what is at stake in each action. The mode of each action is written down before go-live, and you approve it:

• it acts alone on what is reversible and low-stakes, such as the minutes of a meeting;
• it prepares, you approve what goes outside the company, such as the reply to a complaint;
• it prepares, you decide and carry out what cannot be undone, such as a payment.

By default, what goes to your customers is therefore approved by a person. This setting remains yours: for payment reminders, for example, you can let the first courteous ones go out on their own, or exclude certain customers from any automatic reminder.

No. No irreversible action goes out without a person’s approval: it is a design rule, not an option. An agent never does any of these alone:

• pay or make a transfer: it prepares, a person executes;
• sign, or set a price;
• publish without review;
• handle a sensitive dispute: for a delicate complaint, it reconstructs the history and prepares the reply, a person approves it;
• send a formal notice, or settle a strategic decision.

It can prepare, compare and calculate everything. The decision remains yours.

Your company remains responsible for what it does, with or without an agent. That is why actions that commit it remain approved by a person. Each agent also comes with a log of what it does and a trail of its decisions: you can understand an error, correct it and prevent it from happening again. The documentation handed over explains how to take back control in case of error.

No. It only accesses the data and tools it needs for its work, read-only or with write access, as written down at the start. You approve the people, the roles and the access levels; they can be changed and revoked at any time. The documentation delivered with each agent says who has access to what and where the data flows, and the “Where your data goes” page, attached to the proposal or the contract, names the providers involved.

We check every AI agent against your real cases before it goes live: 15 to 30 cases for a Process Agent, difficult cases included. This testing is on us, not on you:

• in your software’s test environment, when its provider offers one, switched on with your agreement on your account;
• otherwise in observation mode: the agent handles the real cases without writing anything into your tools; it proposes, a person checks.

Final approval follows, tied to the agreed result. Any cost of a paid test environment is set out in the quote: you know it before you sign. Afterwards, Monitoring, if you choose it, re-checks on your cases every month, as soon as a process is in service.

Yes: every agent we put into service comes with its traceability and safeguards, included in the project, neither an option nor an extra. Adapted to its scope, they are:

• a log of what it does: what it has handled, the errors, the alerts and the agreed indicators;
• a trail of its decisions, kept for the agreed periods;
• access limited to what is necessary, which you approve and can revoke at any time;
• reference documentation: who has access to what, where the data flows, who approves what, how to take back control in case of error;
• a person’s approval for every irreversible action.

You keep them even without Monitoring; if you choose it, Monitoring maintains them over time.