WALKTHROUGH · AI AGENTS
An agent asks. A person approves.
AI agents can do the legwork through RevexOne’s MCP server. This walkthrough follows real requests from a test organisation, from the agent’s request to a person’s decision.
AI agents should take on the routine work without being able to post results or pay anyone.
The agent asked for a run and gave its reason; the draft waited in the approval inbox tagged as the agent’s, and a person decided: one run approved, one sent back with a reason.
Step 1 of 6
An agent gets its own login, its own tools and its own limits.
The org admin adds the agent on Agents & Logins: its name, which MCP tools it may use, and how many changes it may make in a day. The agent never borrows a person’s login, so everything it does is recorded as the agent.
Answers: “Which agents can work in our data, and what may they do?”
- Works for
- Org admin of an India pharma test org
- Connects through
- MCP server
- Tools
- Read plans and runs, request a run, read results
- Daily limit
- 200 changes
- Posts or pays
- No
Step 2 of 6
Through MCP, the agent asks for a calculation run.
At month end the close workflow asks the agent to run Q2. The agent requests the run and gives its reason. The run is calculated as a draft: 3,303 sales lines, 20 people paid, ₹970,339 in total. Nothing is posted.
Answers: “What did the agent do, and why?”
- Asked by
- Agent · close-assistant
- Reason given
- “Month-end close workflow asked for the Q2 run”
- Plan
- PLN-IN-CD-FY27 version 1 · FY2027-Q2
- Result
- ₹970,339 · 20 people · 3,303 lines
- Posted
- Nothing until a person approves
Step 3 of 6
The draft lands in the approval inbox, tagged with its origin.
The approver sees what was asked, by which agent, on whose behalf and why, with the run’s totals beside it. The origin tag stays on the item; it cannot be changed later.
Answers: “Who or what started this?”
| Item | Origin | Reason |
|---|---|---|
| Calculation run · FY2027-Q2 | Agent · close-assistant | Month-end close |
| Calculation run · FY2027-Q2 | Agent · close-assistant | Late stockist load arrived |
Step 4 of 6
One request is approved, the other is turned down, both with a note.
The approver checks the totals and approves the first: “Checked totals against Q1.” The second was asked because a late stock load arrived; the approver rejects it: “Wait for the corrected load.” The agent cannot approve its own request.
Answers: “Who signed this off?”
| Request | Decision | Note |
|---|---|---|
| Month-end close | Approved | Checked totals against Q1 |
| Late stockist load | Rejected | Wait for the corrected load |
Step 5 of 6
An automated job’s first run is a test that a person approves.
When a scheduled job such as the month-end close runs for the first time, it runs as a test with nothing posted. A person approves that first run and chooses how its later runs are approved.
Answers: “Can we let this job run on its own?”
- Origin
- Scheduled job
- Run type
- Test, nothing posted
- Then
- The approver chooses how later runs are approved
Step 6 of 6
Every record keeps who asked, who approved and why.
Months later, an auditor can see that the Q2 run was asked for by an agent, for what reason, and approved by a named person with a note. Paying people always needs a person’s approval. Letting an admin allow a named agent to post specific records is Planned.
Answers: “Was a person in control?”
- Agent close-assistant asked, with a reason
- Draft run ₹970,339, nothing posted
- Inbox tagged Agent
- Person approved with a note
- Posting only after the approval
Values come from a RevexOne test organisation: invented people and companies, real calculations by the product. Panels are simplified; people and the organisation are left out.