Agents prepare. People approve.
AI in your ERP should make you faster without making you nervous. Here is exactly what Dynocortex's agents can and can't do.
- You askor an alert suggests the next step
- Agents readthrough your own ERP permissions
- You approvea card showing every line
- A draft appearsin the ERP, not yet submitted
- A person submitsas with any other document
Every step is logged in a sealed record that shows any later change.
They read first, through your ERP's own reports
Agents answer from the ERP's standard reports (profit and loss, receivables, gross profit, budget variance) and from your records. The arithmetic is the ERP's, not the model's. Every figure states its company, currency and period.
They act as you, never as more than you
Each question runs with the asking person's own permissions in the ERP. If you can't see a company, a cost center or a document, neither can the agent working for you. Someone with no ERP account gets nothing.
Payroll, bank accounts and employee personal data are excluded by default, and so are email addresses, phone numbers and passwords. Only your administrator can open one of these up.
They prepare drafts. You decide.
An agent can propose a document: a restock request, a quotation, a purchase order, a follow-up task. You see a card with exactly what will be created: every line, quantity, price and total. Only when you approve does the draft appear in the ERP. It is still a draft: a person checks it and submits it, as with any other document.
If something it was built from changes before you approve (a price, a quantity already ordered), the card is refused and you're shown the new version. Agents never submit, cancel or delete anything.
Everything is logged and sealed
Every draft has a log entry, chained to the one before it, so a changed, removed or reordered entry is detected. An oversight report shows, for each draft, who asked, who approved, when, and what became of it: submitted unchanged, edited, or deleted.
Text from outside is treated as data
Customer names, item descriptions and supplier notes can contain anything, including text written to trick an AI. The agents receive them marked as untrusted, to report on, not to obey. And if a hidden instruction ever got through, it still couldn't write anything without a person's approval.
You can stop it
Your administrator can switch off drafts, alerts, the brief, or all agent access, each on its own. The widest switch takes effect on the next request and undoes with one click.