Keldmoor, a fictional industrial distributor from Ohio, has a problem with its September invoices from Stavlund, a fastener supplier in Milwaukee. One is billed at the old price after a contract amendment. One charges for 12,500 lock nuts when the dock received 12,100. One lists lines that add up to 8,240.00 while the invoice says 8,420.00. And a paper copy of one of them turned up again in the mailroom scan.
The accounts payable team works with an AI assistant, so it asks the obvious: find the disputed invoices, tag them “Do not pay – disputed” and draft an email to the supplier with references to the contract. The question is not whether the assistant can do it. The question is what you let it do on its own.
Our new video, Agents under control, shows how DocReveal answers that question. It runs three and a half minutes, you can watch it at the top of this page, and it is made for finance and operations leaders and for the IT and security teams who sign off on AI. Here is what you will see, and why it matters.
Allow: one key per assistant
An external assistant connects to DocReveal over the open Model Context Protocol (MCP), with an access key of its own. In MCP Manager, the team creates a key named after the assistant and narrows it to what the job needs: documents, search, business rules, labels and the platform tools — nothing else. The key expires in 90 days, its secret is shown only once, and it can be rotated or revoked at any time.
No shared passwords and no all-or-nothing access. Every assistant works under its own name.
Decide: which actions wait for a person
Not every action carries the same risk. For each capability, the team decides how its sensitive actions are guarded: they run automatically, the assistant has to confirm them first, or a person has to approve them. For labels, Keldmoor chooses human approval.
Every sensitive action is marked as such, including assigning, removing and creating labels. Some safeguards cannot be lowered at all: changes to team members always need a person.
Work: the assistant reads, finds and proposes
Reading runs straight away. The assistant lists the September invoices, reads the results of the business rules and pulls the relevant passages from the supply agreement, its amendment and the receiving report. Its answer names three disputed invoices, worth 720.00, 436.00 and 180.00 USD in dispute, flags the paper copy as a duplicate and confirms that a fourth invoice is in order.
Then it asks to apply the label, and that is where it stops. All four label requests wait for approval, and the assistant says so. The email to the supplier is a draft, and it stays one.
Approve: a person says yes
A notification appears in DocReveal, and the approver opens the approvals inbox in Agent Manager. Requests from external assistants and from DocReveal’s own agents land in one place and can be filtered by origin. Each request shows the action, the key that asked for it and when.
The AP lead approves the first request with a comment, “Disputed per MSA-2026-014 – hold payment”, and then the other three. Back in the conversation, the assistant checks the approval status and finishes the job: the label is now on the three invoices and on the paper copy. The fourth invoice stays unlabelled, and the email still waits for a person to send it.
These are the same September invoices you may know from Checked, not trusted and Ask your invoices. There, DocReveal found the problems. Here, an AI assistant acts on them, within limits that people set.
Audit: every call on the record
In the activity log of MCP Manager, every call the assistant made is on the record: what it read, what waited for approval and what ran after the yes, each with its time, capability and key. The team can filter by key, capability or status and export the selected calls as CSV, JSON, XML or YAML for its auditors.
When someone asks who changed what, and when, the answer is a filter away.
What to take away
- Give every assistant its own key. Limit it to the capabilities it needs, let it expire and keep revoking in your hands.
- Decide where a person must say yes. Let reading run on its own and make changes wait for approval.
- Approve in one place. One inbox for every request, with a comment that explains the decision.
- Keep the record. Every call, with its key and time, ready to filter and export.
All companies and figures in the video are fictional demo data.
Agents propose. People approve.