LinkedIn
CampusMate uses Agentforce to help students with university life, starting with money. A student explains what has happened in their own words, the fund's rules do the arithmetic, the agents explain, and a person makes every decision.
A student at eleven o'clock at night: the laptop has broken, the rent is due on Thursday, and the shifts have been cut. Universities have hardship funds for exactly this, but the usual route is a long form and a wait while advisers read bank statements by hand. CampusMate makes it a conversation, with a person who decides.
Three Agentforce agents, each with one job:
- Student Agent: chats on the university website, explains the help that could fit straight from the fund's rules, asks for one figure at a time, takes the documents in the chat and sends the application
- Assessment Agent: applies the fund's rules as soon as the application arrives and recommends an amount, with every line showing the rule it came from
- Adviser Agent: helps the student money team in Salesforce with their queue, summaries, the review and the decision letter
Behind them are 28 fund rules held as configuration, 33 flows, 4 prompt templates and 2 small pieces of Apex. Wharfedale University is a fictional university created for the demo.
Status: the demo build is complete and tested on the live site; the demo video is being recorded now.
A person decides
The agents never promise an amount. The Student Agent says "up to", because a person decides. The Adviser Agent can read the figures, but a person checks the documents, and nothing can be approved until they are marked as checked. Before anything is recorded, the agent shows exactly what will be recorded and waits for a yes. Awards over £500 need a second approver, everything is in an audit trail, and a switch turns the assistant off.
When money is not the only problem
If a student says something that matters more than the money, such as skipping meals, the Student Agent pauses the money side, offers a food voucher that day and offers the wellbeing team, but only if the student agrees. It asks what would make things easier, such as email rather than phone, or short messages, and it never asks why.
Rules as configuration, not code
The fund's rules live in custom metadata, so the arithmetic is the same every time and each recommendation line names its rule. In the demo, a student's immediate need is £459, but emergency payments are capped at £250 by rule EP-01, and the recommendation says so. When the adviser asks "Why £250?", the agent answers from the rule lines.
What I learned
- One subagent per job, with deterministic steps for anything that records a decision; otherwise the model can say it has done something it has not
- Confirm in the conversation, store the values, and record them only on the yes
- The site chat is a custom chat on the Agent API, as in VisitMate, so the agent can show cards and take uploads
Next, the same pattern can cover the Wellbeing and Accommodation tabs.