How I work with AI

I don't put AI into an exporter's documentation process unless a human who knows export signs off the output, the system's limits are written down before it is built, and you can run it, audit it and switch it off without me.

Every output that leaves the system has been checked by someone who knows what right looks like. A commercial invoice that reads beautifully and says oranges when the shipment is mandarins is worse than no invoice at all, because it gets through.

The check is not "does this look right". The check is against the source: the job file, the contract, the booking, the actual shipment.
Nothing goes to a client, a forwarder, a bank or a border without a named human sign-off. No exceptions for volume, deadlines or cut-off pressure.
The person signing off knows the document type. I did this work by hand for years before I automated any of it, invoices, packing lists, certificates of origin, or market specific certificates, pre-alerts. That is what makes the sign-off worth anything.
The system is tested against real historical shipments before it runs on a live one, including the ones that went wrong.
Every run is logged inputs, outputs, who signed, when, what they changed. If a certificate is queried six months later, the record exists.
Known failure modes are written down and given to you,not discovered by you. Weights that don't reconcile, descriptions that drift from the contract, references transposed between documents, stale container numbers after a roll.

Nothing gets automated unless automating it is the right answer

Most of what goes wrong in export documentation is a design problem, not a technology problem. The same shipment data gets retyped into five documents. Fixing that needs structure, and sometimes structure is a spreadsheet and a checklist, not a model.
I would rather tell you not to buy something than sell you a system that fails quietly.
Every engagement starts with a paid look at how your documentation actually moves today, before any build, any tool, any commitment.
Some of what you do by hand should stay that way, and I'll say so. Judgement calls, exceptions, anything where the cost of a wrong answer exceeds the cost of doing it by hand.
AI is used where it is genuinely better than the alternative, checking, cross-referencing, reconciling, extracting from documents. Not for its own sake, and not because it is what is being sold this year.
One person in your business owns each system; one person on my side is accountable for it. In writing, before anything is built.
Your data is scoped to what the job needs and no more. Where it goes, who can see it, how long it is kept, and what it is never used for including that it is not used to train anything.
The tools underneath are named. Which models, which services, which vendors, and what they do with your data. No black boxes presented as my own cleverness.
Access control, backup and recovery, and a tested position on prompt injection** wherever a system reads documents from outside your business.

You end up owning it, operating it, and able to stop it

A system only one person understands is the same problem you had when it all lived in one person's head — just with a better interface. If my leaving breaks it, I have sold you a dependency, not a solution.
Handover is part of the build, not a phase that gets cut. Documentation, a written operating procedure, and someone in the business trained to run it.
Every system has an off switch, and you know where it is.Turning it off returns you to the manual process, which stays documented.
Anyone whose work the system touches is told it is there, what it does, and what it doesn't do. No AI running silently behind someone's job.
There is a way to say "this output is wrong" and have it looked at by a person with the authority to overrule it — for your staff, for your clients, for importers on the other end.
Fewer certificate re-issues, fewer amendments, fewer problems found at cut-off instead of at the start. If a build doesn't move one of those, it shouldn't be built.
If I can't explain what a system does to the person who has to run it on a Tuesday morning, it isn't finished.

What I won't do
Build a system that makes a decision affecting a person's money, employment or shipment without a human in the line.
Take on work my professional indemnity cover doesn't reach.
Use one client's data to build something for another.
Claim a system is more capable than it is to close an engagement

This is how I work, mapped to the ten guardrails in the Australian Government's Voluntary AI Safety Standard (Department of Industry, Science and Resources).