Skip to content
All features

AI platform

Ask. Do. Watch.

A multi-agent assistant that answers from your data, proposes actions you confirm, and runs watchers that catch problems before people notice them. It obeys the same roles and scopes as every screen.

app.dinohrms · Dashboard

Good morning, Priya

Saturday, 6 September

⌘J Copilot

Headcount

248

+6 this month

Present today

91%

12 on leave

Open approvals

7

3 waiting on you

Payroll run

Sep

Review · 2 findings

Attendance, last 12 weeks

avg 82%

AI insights · this morning

  • Payroll pre-lock: 2 employees have no bank account on file.
  • Design team is 40% on leave next Thursday.

Three jobs

One assistant, three ways to help.

Each pillar is a different mechanism with a different safety envelope. Watchers never call a model; writes never run without a person.

Ask

Natural-language questions routed to the right specialist. Retrieval over policies, help articles, jobs and past tickets, reranked and cited.

  • Streaming answers with charts and tables attached
  • Memory per user and per company
  • Cached knowledge answers to stay within provider limits

Do

Fifty-nine tools cover leave, attendance, payslips, tickets, approvals, candidates, goals, courses, reports and more. Writes become proposals you confirm, executed as you, under your permissions.

  • Autopilot per tool: off, suggest, confirm or auto
  • Decisions and money always require a confirmation
  • Every run traced step by step

Watch

Twenty-three deterministic watchers run every half hour and after key events: missing punches, leave coverage, payroll audit, SLA breaches, stale goals, hiring stalls.

  • Findings linked to the record they concern
  • Morning digest per person
  • No model call needed, so they never hit a quota

Specialists

Six agents that hand off to each other.

A question is routed to one specialist. If it needs another, the agent hands off, up to three hops, and the trace shows every step.

HR helpdesk

Answers policy and process questions, citing the help centre and published policies.

Self-service

Leave balances, payslips, attendance and requests for the person asking.

Manager

Team coverage, pending approvals and one-on-one briefs for people leaders.

Analyst

Charts and tables from safe read-only views, with numbers verified afterwards.

Recruiter

Screens resumes, spots duplicates, drafts job posts and moves candidates.

Writer

Announcements, review feedback, ticket replies and policy drafts in your tone.

Every run follows the same graph

  1. Guard
  2. Recall memory
  3. Route
  4. Specialist
  5. Tools
  6. Hand-off
  7. Verify
  8. Answer or proposal

Watchers

Twenty-three things it checks so nobody has to.

Watchers are deterministic queries, not prompts. They run every thirty minutes and after key events: a leave request, a payroll calculation, a new hire, an approved resignation. Findings link to the record, proposals wait for confirmation, and a morning digest sums it up per person.

  • Autopilot per watcher, capped by company policy
  • Notices deduplicated per person, per day
  • Zero model calls, so they run on any plan
  • Missing punches
  • Leave coverage risk
  • Payroll pre-lock audit
  • Onboarding kick-off
  • Onboarding check-ins
  • Exit clearance
  • Helpdesk SLA watch
  • Helpdesk auto-draft replies
  • Timesheet reminders
  • Probation reviews due
  • Stale goals
  • Hiring pipeline stalls
  • Policy acknowledgements
  • Attrition risk
  • Overtime and burnout
  • Compensation drift audit
  • Leave about to lapse
  • Roster gaps
  • Requisitions and time-to-fill
  • People data quality
  • Access review
  • Review calibration
  • Document intake

Safety model

It can only do what you can do.

The assistant is a user, not a superuser. Every tool call is evaluated against the caller’s permissions and scopes, and every write waits for a person.

Same roles, same scopes

Tools read through the identical scope filters the pages use. The assistant cannot see what you cannot.

Compensation redacted

Pay figures are stripped from tool output unless the caller holds a compensation permission.

Prompt-injection guard

Retrieved documents are fenced and instruction-like passages flagged before the model reads them.

Confirm before write

Writes are proposals until a person confirms. Confirmation executes under that person, not a service account.

Caps and usage

Daily caps per role, a usage ledger per company and a provider page that shows index health.

Bring your own model

OpenAI-compatible providers: OpenRouter, OpenAI, Groq, Together, Mistral, Gemini or a local Ollama.

Where it shows up

Inside every suite, not in a separate tab.

Small and large

The same assistant at ten people and ten thousand.

What changes with size is policy, not capability: which roles may use which agents, daily caps, autopilot levels and how many companies it can see.

10 – 100 people

Small companies

One admin, no HR department, a system that does the HR desk for you.

100 – 1,000 people

Growing companies

Managers approve from one inbox; HR stops being the bottleneck.

1,000+ people

Large companies

Several legal entities, branches and countries under one login, all of it audited.

Questions

About the AI, specifically.

Is the AI optional?

Yes. Every module works without a model. Watchers are deterministic queries and run on any plan. The assistant needs a provider key; pick one from the AI provider page.

Can the assistant see salaries?

Only for callers who hold a compensation permission. Pay figures are stripped from tool output before they reach the model for anyone else.

Can it take actions on its own?

Reads, yes. Writes become proposals a person confirms and are executed as that person. Decisions and money never run on full autopilot, whatever the autopilot setting.

Which model providers work?

Any OpenAI-compatible endpoint: OpenRouter, OpenAI, Groq, Together, Mistral, Gemini, or Ollama running on your own hardware.

Is there an audit trail?

Every edit, approval, login, session and AI proposal is written to a searchable log per actor and entity. Assistant runs are traced step by step.

Ask it something about your own data.

Sign in as a manager, press ⌘J, and ask who is out next week. Then confirm the proposal it makes.