An HR agent that acts: preview, approval and undo
Most HR chatbots answer questions. Rafiq AI is designed to do the work. That changes what has to be true before you let it near employee records. Here is how we designed it.
Rafiq AI, the AI agent in RafiqHR, completes HR tasks instead of only answering questions. It uses the same tools and permissions as the user, shows a preview before changing anything, sends sensitive actions such as pay, promotions and terminations through human approval, and writes a receipt for every action that can be undone within 24 hours.
An assistant that answers "how many leave days do I have?" is useful. An assistant that files the leave request, routes it to the right manager and updates the team calendar is far more useful, and far riskier. Rafiq AI, the agent inside RafiqHR, is built to do the second thing. This post describes the rules we set for it.
It uses the same tools as you, with your permissions
Rafiq AI does not have a back door into the database. Every action it takes goes through a registered tool, the same operation the screen would call, and each tool declares the permission it needs and its risk level. When an employee asks Rafiq AI for something, Rafiq AI can do only what that employee could do by clicking. A manager's Rafiq AI sees the manager's team, and nothing else.
Nothing changes before you see it
For anything that changes data, Rafiq AI first shows a preview card: what it is about to do, to which record, with which values. Nothing happens until the user confirms. If the request is ambiguous, for example two employees named Sara in the same department, Rafiq AI asks rather than guesses.
Sensitive actions go through approvals
Each company chooses how much autonomy Rafiq AI has for each role: answer only, suggest, act with approval, or act within limits. Some actions are never automatic, whatever the settings: paying salaries, promotions, disciplinary action, terminations, government submissions and bank files always wait for a human approval through the company's usual approval chain. Rafiq AI can prepare a payroll run and explain its variances; it cannot approve one.
Every action leaves a receipt
When Rafiq AI acts, it writes a receipt to the audit log: who asked, in what words, which tool ran, with which inputs, who approved it, and how to undo it. Where a tool supports it, the action can be undone for 24 hours. That makes Rafiq AI's work reviewable in the same place as everyone else's.
Arabic first
Most HR questions in Saudi companies are asked in Arabic, so every tool has an Arabic and an English description, and we test Rafiq AI with scenario suites in both languages. A request in Arabic gets an answer in Arabic, with dates and numbers formatted the way the reader expects.
The system still works without it
Rafiq AI is a shortcut, not a gate. Every screen in RafiqHR works directly, and if the AI provider is unavailable, Rafiq AI says so and the rest of the product carries on. Each company can also choose its own AI provider and model, or a private model for on-premises installations.
You can read more about Rafiq AI on its feature page.
