We built a compliance check that audits creative before legal approves it.
At Hiaitus we spend most of our time on the same question: where does AI actually earn its place in a business, and where is it just noise? The honest answer is usually narrower than the hype suggests. The work that AI does well is the repetitive, high-volume first pass that a skilled person could do but shouldn’t have to do fifty times in a row. Find that step, build for it, and leave the judgement to the human. That is the whole methodology.
Checked It is what happened when we pointed that lens at Australian advertising compliance. It is our own product, built by Hiaitus Pty Ltd, and it is the clearest example we have of the way we think about applied AI. So this is partly a product post and partly a builder’s note on why it looks the way it does.
The step everyone waits on
Start with the bottleneck, because that is where we always start.
In any regulated category — alcohol, wagering, health supplements, food and beverage, insurance broking, news publishing — compliance review is the slowest step between a finished creative and a live campaign. It is also the step that happens last, when the shoot is done, the budget is spent, and a change is at its most expensive. A concept that needed a small adjustment at the storyboard stage instead comes back flagged after production, and the team is choosing between a costly redraw and a reshoot.
The frustrating part is that most of what makes this step slow is not senior judgement. It is first-pass triage: confirming a mandatory message is present, checking a disclaimer is legible, spotting an appeal-to-minors risk, verifying the responsible-consumption line is there. That work is real and it matters. It is just not what expensive legal hours are for.
So the design goal wrote itself. Move the check earlier. Screen the brief, the storyboard, the concept — so that what reaches legal is already built to pass. In software the phrase for this is “shift left”: catch the problem at the cheapest possible moment rather than the most expensive one. Checked It shifts compliance left.
What that looks like in practice
A marketer uploads a creative — an image, a PDF, or a video — and selects the sector. In about ninety seconds they get back a structured report. (Indicative, not a guarantee; grey areas take the time they take.) Each finding is scored against a clear band: Pass, Human Review Required, or Fail. Every finding cites the specific clause of the specific code it relates to. And every red finding comes with a fix brief — three concrete recommendations, each pairing the action to take with the provision it addresses.
That last part is the difference between useful and not. “This might be risky” tells a marketer nothing they didn’t already fear. “This reads as an appeal to minors under the relevant ABAC provision because the cartoon element is the primary visual, and here are three ways to change it” is something a team can act on this afternoon.
The result is that the check moves from a days-long queue at the end to a self-service step at the beginning. Legal still reviews. It just receives a clause-cited report instead of a stack of raw files, and it spends its hours on the calls that actually need a lawyer.
Why it’s built the way it is
Two engineering decisions carry most of the weight, and both come straight out of the Hiaitus methodology.
The first is that there is no single universal ruleset. Checked It runs six purpose-built engines, one per sector, each grounded in the frameworks that actually govern that category — ABAC for alcohol, the National Consumer Protection Framework for wagering, the TGA Therapeutic Goods Advertising Code for supplements, and so on. A global compliance suite trained on US or EU regimes does not know what an Australian code requires. Localised is not the same as built for. This is depth by design, not breadth for a marketing line.
The second is restraint, and this is the part we are proudest of. A generic model asked a compliance question will answer confidently every time, including when it invents a clause number or a determination that does not exist. In front of a regulator, a fabricated citation is worse than saying nothing. So the engines are deliberately constrained: they can only cite determinations from a verified set, they have to state whether a ruling was upheld or dismissed and apply it only in that direction, and they treat a mandatory disclaimer as the floor rather than as bonus credit. When a call sits on the line, they escalate to human review rather than wave it through. We would rather flag something that turns out fine than stay quiet on something that doesn’t.
That is “Human Intelligence × AI” made concrete. The system does the tireless first pass. The person keeps the judgement, the sign-off, and the responsibility. Checked It is a pre-screen with human oversight, and it says so plainly — legal counsel retains the final word, by design.