AI consulting
Australian enterprises do not have an AI strategy problem. They have a systems problem.
The bottleneck is not another model vendor. It is Salesforce, SAP, the core, and the APIs nobody wanted to own. A Sydney view of how to actually ship.
Most Australian boards now have an AI agenda. Very few have a system that an agent is allowed to touch.
That is not a model problem. It is a systems problem. The customer record still lives in Salesforce. The ledger still lives in SAP or a core that predates the cloud. The workflow still lives in a shared inbox and a person named Helen. Until those things are addressable — with contracts, permissions, and audit — a chatbot is a costume.
What “strategy” usually means
In the rooms we sit in, AI strategy has come to mean three artefacts:
- A slide with a landscape of vendors.
- A proof of concept on a CSV that nobody will refresh.
- A risk paper that correctly says “we are not ready” and then stops.
None of those ship software. The firms that are moving treat AI as another consumer of the same APIs they should have had anyway.
Start with a workflow that already has a system of record
Pick one process with a clear owner, a measurable failure, and a system that is already the source of truth. Claims. Onboarding. Quote-to-cash. A partner portal that currently is a spreadsheet.
Then ask four questions:
- Can we read the record without a human export?
- Can we write back, idempotently, with an audit trail?
- Who is allowed to approve the write?
- How will we know the agent got worse this week?
If you cannot answer those, you do not need a model bake-off. You need integration, data contracts, and a product owner.
Australian constraints are not optional colour
Privacy Act reforms, APP, APRA CPS 234 and 230, sector regulators, and the political reality of on-shore data are not “later.” They are design inputs. The advantage of a Sydney practice is not patriotism. It is that we have already had the argument with legal, and we still have to ship.
What a 90-day programme actually looks like
Weeks 1–3: diagnose the workflow, map systems, write the eval set from real historical cases.
Weeks 4–8: thin slice in production — one tool, one write-back, one human approval.
Weeks 9–12: harden, hand over, and decide whether the second workflow is even the same shape.
If your vendor cannot describe that calendar without saying “platform,” keep looking.
The model will change again. Your systems of record will not. Build for the latter.
