Hirra, the AI Payroll Assistant

What Hirra is, and what it is not

Hirra is a chat panel inside SmartCompaire, labeled Beta. The panel says plainly that Hirra can make mistakes, so check figures that matter.

Payroll reconciliation in SmartCompaire runs on fixed matching and grading rules, not AI. Hirra does not decide whether a row passes; it reads the results and helps you interpret them and record your team’s decisions.

Hirra is told to base every figure on data it read in the current turn, never to estimate, and to say when a list is cut off.

What you can ask the AI payroll assistant

Hirra works through 40 typed tools that call the SmartCompaire API. Most of them are read-only. If you have a project open, Hirra is told to treat “this project” as that one.

Projects and setup

Find a project by partial name, check setup progress, and list mappings and source or target pay rows.

Runs and results

List comparison runs, open individual result rows, and ask where the over-threshold differences are concentrated.

Breakdowns and trends

Get totals and breakdowns by bucket, code type, employee, pay code or status, and the over-threshold trend from run to run.

Defects and KD/AD

Review defects by status, priority, bucket, type or assignee, and look up known or acceptable differences (KD/AD).

Thresholds and rules

Check the thresholds and compare-status rules that apply to a project.

Jobs and reports

Ask whether an import, run or report has finished, and list the reports already generated.

Charts

Ask for a bar, line, donut, scatter or heatmap chart of rows Hirra has already fetched, each with a caption.

Project summaries in one request

Ask Hirra to summarize a project, or what is wrong with it, and one read-only overview covers:

Setup and readiness. Settings, completed setup steps, dataset counts, the largest unmatched pay codes with the reason, and employees without a mapping.
The latest run. The most recent completed run and its headline figures, plus recent imports.
Next steps. Up to 10 suggested next steps, each naming the app page to open.
Honest gaps. A part that fails to load is shown as unavailable, not as zero.

Actions Hirra can take for you

For a member or above, these run straight away from chat under the user’s normal permissions. The “Let the assistant take actions” permission is not enforced yet, so there is no separate switch to stop them. Viewers can only read.

Create a project. Start a new payroll or time comparison project.
Raise and update defects. Raise a defect, update it, or resolve or close it. Resolving needs a real resolution comment.
Record KD/AD. Create or update a known or acceptable difference record.
Review one result row. Set the status, override, defect or KD/AD reference, or comment on a single comparison row. The measured pay amounts cannot be changed.
Queue a report. Generate a report as a background job, check on it, or cancel a job.

Actions that need your confirmation

Seven actions are gated behind the user’s explicit confirmation: launching a comparison run, importing pay data (Replace mode deletes the existing rows), bulk-updating comparison statuses, updating project settings, setting a threshold, creating or editing a compare-status rule, and archiving a project.

Hirra does not run these. The panel summarizes what it was about to do and stops. Confirming from chat is not built yet, so today you complete them on the screen that owns them.

Hirra has no tool to delete data, edit mappings or change pay amounts. It fixes nothing automatically and writes nothing back to a payroll or time system.

AI insights reports for a comparison run

An insights report reviews one comparison run, the latest completed run by default.

A fixed set of reads. Totals, breakdowns by bucket, code type, pay code, employee and status, the trend, defects and thresholds, through the same permission-checked path as chat.
One model call. The result is a headline, a summary and up to 20 ranked findings. A run with no rows gets no model call.
Sanity checks on findings. A finding must cite data that was actually read. Row, employee and amount figures larger than the run are removed, and the share-of-variance figure is computed in code. The written text is the model’s and can still be wrong.
Numbers from the database. Tiles and charts on the insights page come from stored database figures, not the model’s text. Finding buttons only navigate.
Stored per run. Each report is saved with the model, tokens and cost, and reused until someone presses Regenerate.

Hirra only sees what you can see

Hirra has no database access of its own and no tool for free-form SQL or arbitrary web addresses. Every tool is a call to the SmartCompaire API.

Your login, your permissions. Hirra acts with the signed-in user’s own login token, so the API’s normal permission checks and organization isolation decide what it can read or change.
Tools filtered by role. Tools above your role are removed from Hirra’s list entirely, and every call is checked against your role again before it runs.
One organization only. Your organization and role come from the verified login, never the browser. A call naming another organization is refused.
Private conversations. Each conversation belongs to one user in one organization. Colleagues in the same organization cannot open it.
An audit trail of completed turns. Each completed chat turn writes one audit record per tool call, including refusals and actions awaiting confirmation, to an append-only AI audit log. A turn that fails or times out may leave its tool calls unrecorded. Admins and owners can read the log through the API only.

Where your data goes

The browser never calls the AI directly. Messages go to the SmartCompaire backend, then to Corra, the internal service that runs Hirra, and then to OpenRouter, a third-party model router. The model keys stay on the server.

Models. For web chat, OpenRouter’s Auto router picks the model for each turn, with configured fallbacks, and the model that answered is recorded. Hirra is not tied to one named model.
Pay data sent to the model. Tool results, which can include pay rows with employee numbers and amounts, are sent to the model without redaction. Generating an AI insights report also sends that run’s summary figures, including per-employee and per-pay-code breakdowns. The requests set no data-retention or training opt-out with model providers.
What is stored. Secrets and personal fields are redacted from tool inputs in logs and audit records. Your typed questions and Hirra’s answers are stored as written, with no automatic expiry.
Voice and avatar. The chat takes text only and never receives file contents. Voice input uses your browser’s speech recognition in US English, and some browsers send the audio to their vendor’s servers. The optional avatar is off by default; when on, Hirra’s reply text goes from your browser to HeyGen and its streaming provider LiveKit to be read aloud.
Privacy policy. SmartCompaire’s own privacy notice is still under legal review. Until it is published, the privacy policy linked from this site is that of SLRIS Solutions, LLC.SLRIS privacy policy

Frequently asked questions

Can Hirra see projects or data I cannot?

No. Hirra uses your own login token, so the same permission checks and organization isolation apply as when you use the app by hand.

Is my payroll data sent to the AI model?

Yes, when Hirra reads data to answer you or writes an insights report. It goes through OpenRouter to the model provider, with no zero-retention or no-training option set.

Does Hirra remember earlier conversations?

No. Each conversation stands alone, and within one the last 20 messages are kept as context. The transcript is saved on the server so the conversation can be reopened, with no automatic expiry.

Who can use Hirra?

Every role, viewer through owner, can use Hirra by default. Viewers can only read, and actions start at the member role.

See SmartCompaire on your own payroll data

Bring a source and a target extract from your parallel run; we will walk through the comparison with you.

Request a demo