Sixteen years in signals intelligence, then a decade delivering production AI inside Australian government operations — systems that make decisions about real people, under real accountability, and are still running on Monday.
Back to the frontSome years ago RAND ran an AI readiness assessment on an organisation I was working in. I was one of many people they interviewed. I gave them a great deal, and I was quietly annoyed about it at the time — not about the findings, but about the access. They could ask questions I could not ask. People made time for them who would not have made time for me. My analysis was not weaker than theirs. My position on the org chart was the problem.
The report they produced agreed with my own read at about ninety per cent. The other ten per cent was worth the engagement on its own, and I could not have reached it from where I sat.
That is the case for bringing in someone from outside, and it is narrower than the case usually made for it. It is not that the outsider knows more. It is that inside an organisation, anyone with a view on AI is also understood to have a stake in the answer — a team to protect, a budget to defend, a position taken in a meeting eighteen months ago that they are now expected to hold. Questions are heard as moves. Findings are read as bids. People calibrate what they say to who is asking, and the internal diagnosis comes out partial in a way that cannot be seen from inside it.
Someone with no history in the building, no team in the restructure and nothing to win from the conclusion gets different answers to the same questions. Much of what comes back will be what somebody internally already suspected. That is not a wasted engagement. It is the first time it has been sayable.
National securityIntelligenceLaw enforcementBorder securityMigrationSupply chainInternational trade and mail
I built the applied AI and data science branch behind that work and ran it for a decade. One system it delivered received a national data science award. Another delivered a 228% productivity improvement for a small team of specialists doing work that could not be scaled any other way. The branch also built the risk models supporting Australia's migration stream, which is where the position in judgement 02 comes from: where an automated decision affects a person's circumstances, someone has to be able to intervene and answer for it. That is not a principle read somewhere. It is one that had to be designed for.
Emeris is not an AI implementation shop, a policy-writing exercise or a vendor sales channel. The work starts with finding out what is actually true, then deciding what needs to change.
The failure modes in that work — a control routed around by Thursday, a capability that left with the team that built it, an escalation path nobody had walked — are not unique to government. They appear wherever decisions are automated at scale under scrutiny. Regulated financial services and critical infrastructure are arriving at them now from a different direction — through prudential expectations and security obligations rather than through operational exposure — with the same questions unanswered. The obligations are new. The failure modes are not.
The less flattering half of the record is the more useful half. I have killed projects that should not have shipped. I have been in the room when a production model degraded and the escalation path turned out to be theoretical. Frameworks are straightforward to write and I have written some. Having owned a system while it was going wrong is the part that is difficult to fake, and it is the part that shapes how Emeris works.
No deck, no discovery process, no proposal arriving three days later unless you ask for one. It is a diagnostic conversation, not a sales pitch.
Thirty minutes is usually the difference between a description and a diagnosis.