How the agents are governed
This is an AI system. Dynocortex's answers, briefs, alerts and drafts are produced with AI models, and the product says so where you meet them.
Models. On the hosted service, we supply the model. If you run it yourself, you choose your provider.
What agents can see: only what the person they're working for can see in the ERP. Payroll, bank accounts, employee personal data, email addresses, phone numbers and passwords are excluded by default.
| Agents can | Agents can't |
|---|---|
| Read reports and records the person may see | See more than that person |
| Prepare drafts for a person to approve | Submit, cancel or delete anything |
| Send alerts and briefs | Write anything a person hasn't approved, even if text in the data tells them to |
Human approval. Every draft is approved by a person, on a card showing exactly what will be created. If the source changed since, approval is refused. The approved draft still waits for a person to submit it.
Logs. Every draft is recorded in a log that shows changes, deletions or reordering. The oversight report shows, per draft, who proposed it, who approved it and when, and what became of it. It's ready for an auditor.
Stopping it. Separate switches stop drafts, alerts, the brief and the agents' access to the ERP. The widest one works on the next request.
Known limits. Answers can be wrong. Every figure states company, currency and period, and opens the page it came from, so it can be checked.
For EU deployers: the oversight report and the stop switches support the human-oversight and logging duties of the EU AI Act.