Figures on this page reflect early customer feedback during pilot use. They are not the result of a controlled benchmark and may change as our customer base grows.

Why generic AI gets payroll wrong

The SCHADS Award is hundreds of pages of clauses, schedules, and exceptions. That's the floor. On top of it, every provider layers their own internal policies, enterprise-agreement variations, and quirks in how their roster system exports data — all of which change how a shift should be classified.

Two providers can run the exact same timesheet through the exact same award and reasonably reach different conclusions, because their internal policies differ. A generic AI model has no way of knowing which interpretation belongs to which provider.

That's the gap CrossVault's AI Rules close.

What an "AI Rule" actually is

An AI Rule is a short, plain-English instruction that travels with every compliance check your account runs. Think of it as standing orders for the model — the same way you'd brief a new payroll clerk on quirks they need to know about your business.

Most rules sit in one of three categories — data labels that vary across rosters, flags you want raised before a run is signed off, or terms in your enterprise agreement that differ from the award default. An illustrative example:

That single instruction, added once, applies to every timesheet processed afterwards — whether the run happens today, next week, or in six months. Your payroll team doesn't have to remember it. Your contractor doing the run doesn't have to be briefed. The rule is in the system.

The roughly 5% satisfaction lift

Customers in early pilot use have given us informal feedback comparing how each run feels with and without their AI Rules layered in. The pattern is consistent: on the slice of shifts a rule actually applies to, reviewers report runs "match how we'd pay it" noticeably more often once their rules are in place.

The lift is not uniform across all shifts — most shifts have no provider-specific nuance and reviewers rate them the same either way. The improvement concentrates on the rule-affected slice, which is exactly where runs used to feel "almost right but not how we'd actually pay it" before the rule was in place.

vs. ChatGPT one-shot

We've also compared CrossVault runs against what operators get from a single ChatGPT prompt on the same file — the way most providers first try to "use AI" for compliance. No agent coordination, no custom rules. The spread is large enough to be visible by eye.

The larger jump — from ChatGPT to baseline CrossVault — comes from the multi-agent swarm. The smaller jump on top is what AI Rules add. Both contribute. Neither is sufficient on its own for the long tail of provider-specific cases.

What this means for your payroll

A five-percent improvement sounds modest. On a fortnight of 800 shifts it works out to roughly 40 shifts reviewers would have flagged or queried — most of them low dollar value, but exactly the kind of consistent edge case that's tedious to catch manually every cycle.

More importantly, AI Rules turn institutional knowledge — the policy document nobody opens, the EBA clause negotiated years ago, the operations manager's "we always do it this way" — into something the system actually applies. Every run. Every shift. Without anyone having to remember.

Add your first rule from the timesheet review screen, or send us your existing internal policy notes and we'll help you convert them into AI Rules.

---

Notes on figures. Percentages on this page are indicative, drawn from informal feedback during early customer use. They are not the result of a controlled benchmark, do not represent any specific customer, and should not be relied on for procurement or legal decisions. Nothing on this page is legal advice — always consult your own employment lawyer or industrial relations advisor for award and enterprise-agreement interpretation.