How is a synthetic agent governed? Brake, intervention and audit log
A synthetic agent is governed through four mechanisms: written rules that define what it may do, a brake any authorized person can pull, human intervention in any conversation, and an audit log that records every action. No agent passes itself off as a person, and a person is always in charge. It is InnovaBlack®'s “accountable by design” principle.
Why does an agent need governing?
Because an agent doesn't just answer; it acts. It books, quotes, contacts, records, and sometimes touches money or customer data. When a program that only informs gets something wrong, an apology fixes it; when an agent that acts gets something wrong, the consequences can be real. Governing it means deciding in advance what it may do, who answers for it and how it is stopped, rather than finding out on the day something breaks.
What rules does an agent have, and who sets them?
Rules say which tasks belong to the agent, what information it may use, up to what amount or condition it decides alone, and when a case must go to a person. The process owner in the business sets them, not the vendor, and they are written somewhere they can be read and changed under version control. What isn't in the rules, the agent doesn't do. This is part of Synthetic OS, the operating system of every synthetic agent.
How do you stop an agent?
With a brake within reach of authorized people: pause one agent, or all of them, without technical permission or a support ticket. When you buy, check how fast it takes effect, who holds the key and what the customer sees while the agent is stopped (a clear notice and a person to help, not silence). A brake that exists only in the contract is not a brake.
How does a person step into a conversation?
The team can see the agent's conversations, take one over at any moment, reply as themselves and hand it back when done. The agent also escalates on its own when a case falls outside its rules, when the customer asks for a person or when it detects a sensitive situation, and it passes the full context so the customer never repeats anything. This two-way path, where a person takes over the agent and the agent calls the person, keeps control with the team.
What should the audit log record?
Every action the agent takes, with the date, what it did, why and with what result; every human intervention and who made it; and every change to the rules. The test of usefulness is simple: if a customer, an auditor or a regulator asks “what happened with this case?”, someone can answer by reading the log, without reconstructing it from memory. It is the basis for auditing and the reason there are no black boxes.
Who watches the agent when nobody is looking?
Continuous monitoring that checks the agent is working in the real channel it serves, not just that the server is on, and alerts a person when something fails. Some failures raise no error and show up only because the agent stopped answering, which is why health is measured in the channel and not on the server. When a failure is a known one, the system can fix it and record that it did; when it isn't, it escalates to a person.
Does the agent pass itself off as a person?
Never. No agent pretends to be human: it presents itself as what it is and, if asked, does not deny it. A person is always available, and the customer can ask for one. It isn't only ethics: a customer who learns afterward that they were talking to a machine stops trusting the whole company.
How do I check all this before I buy?
- Ask to see where the rules are written and who in the business can change them.
- Ask for a demo of the brake: who pulls it, from where and what the customer sees.
- Ask for a sample log and look up one specific case in it.
- Ask what happens at 3 a.m. if the agent fails: who finds out and how.
- Ask how the agent introduces itself to the customer and how to reach a person.
- Check where your data lives and whether it is used to train models.
To judge whether the agent does its job, see how to measure a synthetic agent. If your industry is regulated, confirm with your legal advisers which obligations apply: this page describes a control model, not legal advice.
Who is responsible if the agent makes a mistake?
Responsibilities are agreed in writing between the business and the vendor. What the design guarantees is that every action is logged and that a person is in charge who can step in and stop it. For the legal detail of your case, consult your advisers.
Can the agent decide on its own?
Within the limits the business owner writes for it. Anything beyond those limits, such as an amount, an exception or a sensitive case, goes to a person.
Can I audit what an agent did two months ago?
That is the purpose of the log: every action is recorded with its date and result so you can reconstruct what happened and why.
Does the brake also undo what the agent already sent?
No. A brake stops what comes next. That is why rules set limits before the agent acts, and why the log shows what already went out so a person can correct it.