One working automation illuminates more about AI in your business than any roadmap document.
We built a production system in a domain where being confidently wrong is a safety problem.
Much AI advice comes from people who have never had to decide what a model must refuse to answer.
Proof
Before advising anyone on AI transformation, we built and shipped one: a production system that drafts regulatory compliance documents for Australian road worksites, in a domain where a confident wrong answer is a safety problem rather than an embarrassment.
The system described here is VTRAFFIC, and it is live. Every claim on this page can be checked against it.
Traffic Management Plans and Traffic Guidance Schemes are the documents that govern how a road worksite is signed, staged and made safe. They are prepared by qualified people, audited by road authorities, and wrong details have physical consequences.
The rules are layered: a national baseline, then state and territory overlays that diverge from it. What is compliant in one jurisdiction is not automatically compliant in the next.
The work was slow, repetitive, and highly susceptible to transcription error — the profile that looks like an obvious AI candidate right up until you consider what happens when the model is confidently wrong.
When the system cannot find a value in the source document, it records it as absent. It does not estimate, interpolate, or fall back on what is usually true. A plausible-but-wrong number is worse than a blank field, because a blank field gets checked and a plausible number does not.
How much the system found, and how compliant the document is, are scored separately and never combined into one flattering number. The compliance scorer never sees whether a field was found, so it cannot reward the system for its own coverage. One blended metric would have hidden exactly the failure we most needed to see.
When another data source agrees or disagrees with a value a qualified engineer confirmed, it adjusts our confidence in that value. It never silently replaces it. The human's judgement is the record; the machine annotates it.
The question is not how much the system extracts. It is whether the things it says confidently are actually right. A tool that is right 70% of the time and honest about which 70% is far more useful than one that is right 90% of the time and cannot tell you which 10% to check.
The system is a drafting aid with a professional backstop, and it was designed that way from the start rather than retrofitted after a scare. That framing changes what you build, what you measure, and what you are willing to ship.
Getting a model to produce a plausible draft took a fraction of the effort. Everything after that — grounding it in the actual regulations, handling jurisdictional divergence, deciding what the system must refuse to answer — was the real work, and it is the part most AI programmes never reach.
Every number we report is derived from a run we can point at, and every guard is driven over a known-bad input to prove it can actually fail. A check that has never failed is indistinguishable from a check that cannot.
The features we chose not to automate did more for adoption than the ones we did. Professionals trust a tool that visibly knows its own limits, and refuse one that has been caught guessing even once.
This is the same sequence the AI Transformation practice runs with clients: listen first, ground every claim in something measurable, design for the human who has to sign their name to the output, and be honest about where AI does not belong.
The practice