Open AI Chat in the sidebar (or Ask in AI mode) to ask questions about your company's records and indexed documents. Chat requires use_ai_chat, the AI module licence, and enabled company/platform AI settings. Individual features and tools have additional controls.
Questions and tasks
Try a first question
Start with a specific request and add a date range when it matters. For example:
| What you need | Example question |
|---|---|
| Your leave | “What is my current leave balance?” |
| Your pay | “Show my latest payslip and explain the deductions.” |
| Your expenses | “Which of my expense claims are still pending?” |
| A company policy | “What do our uploaded policies say about annual leave? Show the source.” |
| A management summary | “Summarise the latest completed payroll run.” |
| Finding a page | “Where can I review attendance corrections?” |
These are starting points, not a promise of access: your permissions, linked employee profile, company settings and available tools determine what the assistant can retrieve. If the question is ambiguous, specify the employee, pay period or document you mean.
With Tools enabled, try asking for your leave balance, recent payslips, expense claims or employee documents. Self-service tools resolve your linked employee profile. Management tools can look up employees, pay runs, leave requests, recruiting records and reports when your permissions allow them.
The assistant can also retrieve excerpts from indexed company documents for policy and onboarding questions. Document-based answers can display source references; tool lookups show cards with status and expandable input/output. Inspect those sources when an answer matters. An instruction to cite sources or avoid guessing does not guarantee that every generated answer is correct.
Enabled task features include resume screening, interview questions, job-description drafting and performance summaries. The assistant can also run a pay run's anomaly check and explain the findings; it never sees the amounts. Availability depends on company configuration. Shift scheduling exists in Arche; see Shift Scheduling for the roster workflow. A feature existing in the app does not mean every operation has an AI tool.
You can attach a PDF, Word, Excel, CSV or text file to a message when an administrator has turned on file uploads in AI Settings. Pay amounts and the private details your company hides are masked before the assistant reads the file. See AI Privacy & PII Handling.
Ask where to find a page when you prefer to carry out a task directly. The assistant's navigation tools can provide application links. The configured chat pipeline uses company documents and registered tools; this guide does not promise public-web research.
Tools and Agentic
| Control | Behaviour |
|---|---|
| Tools | Enabled by default in the web chat. Allows eligible live-data tools. Turning it off removes those tools from the turn. |
| Agentic | Off by default. Requires use_ai_agentic, which only the Admin role holds by default. To give it to another role, contact Arche support. Allows eligible write/action tools in addition to reads. |
There is a self-service exception: with Agentic off, the assistant can still act on your own records, for example submit_my_leave_request, request_my_salary_advance and create_my_expense. Each of these comes back as a proposal, and nothing is saved until you approve it. They still need the operation's permission and the module on your company's plan. Every other write needs Agentic.
An administrator can enable or disable tool categories in MCP Tool Settings (/ai/mcp-settings, manage_ai). Platform restrictions and individual tool permissions also affect the available set. Tools are selected for the current feature and question, so the whole catalogue is not necessarily offered in each turn.
Reviewing an action
Confirmation-gated tools show a proposal card with Reject and Approve. Review the operation and arguments before approving. Reject declines that call. Every call is a separate decision, so you review each one. An approval is bound to you, your company, the tool and the exact arguments you were shown, and it can be used once. A proposal stays open for a time that depends on its risk, and the card shows the time left. If it expires, ask again.
A decided card collapses to one line, for example "Approved", what it was and the time, and you can open it again. If a change would be refused, for example a leave request that breaks a rule, the assistant tells you why and does not offer an approval. Some low-risk tools do not require confirmation at all; Agentic mode is not a promise of a card for every action.
If an action fails, check the tool result and the corresponding application page before retrying. Approval authorises an attempt; it does not guarantee successful persistence.
Company settings
AI Settings controls enablement, provider/model, feature selection, token budget, PII redaction and custom instructions. Viewing settings uses view_ai; changing them uses manage_ai, except the token budget: your company admins set it when your own API key runs your AI, and Arche sets it when Arche's key does.
The new-settings entity defaults are AI disabled, OpenAI with gpt-4o-mini, a monthly budget of 1,000,000 tokens, and redaction enabled. Existing companies and provisioning flows may use different values. The chat checks recorded monthly usage before calling the provider; reaching the budget rejects new requests. This is a pre-request check, not a guarantee that an in-flight request cannot exceed the remaining allowance. A reply stopped part-way, or one that failed after some of its steps had run, still counts towards the budget, including an estimate for the part the provider had not yet reported.
Company keys can be configured, with platform-key fallback. Provider choices include OpenAI, Anthropic, Google, Azure OpenAI, OpenRouter and a custom OpenAI-compatible endpoint; usable models and credentials depend on the selected provider. The settings page offers only the providers the platform has switched on. If the platform switches off the provider a company uses, AI features stop for that company until an admin chooses another provider.
With OpenRouter, one key reaches models from many vendors. The company adds its own OpenRouter key; the platform key is not shared for it. Arche sends every OpenRouter request only to hosts that keep no copy of the data and do not train on it, and the settings page lists only models that meet this rule and support the assistant's tools. OpenRouter can also log prompts on its own side; that is a setting on the OpenRouter account that owns the key, not something Arche controls.
Privacy and audit
Check the answer before acting
Open cited policy sources and inspect the tool results for the relevant dates and records. For a proposed change, check the employee, amount and other details on the approval card. After approving, read the result; a preview or approval alone does not confirm the change succeeded.
If the assistant cannot retrieve something, open the relevant application page or ask your administrator to check your access. Avoid repeatedly submitting a write that may already have succeeded. Never paste passwords or provider keys into chat.
When enabled, redaction masks configured text patterns and known sensitive fields in tool results before the model receives them. Some analytical features deliberately retain pay figures while removing direct identifiers. See AI Privacy & PII Handling for the limits and controls.
AI audit records expose prompt/response information, usage and outcome to users with view_ai. Treat those logs as sensitive: disabling redaction or using an analytical feature changes what may be retained.