What an AI consulting engagement actually looks like.

There is a decent amount written about how to choose an AI consulting firm in Australia, and a little about what it costs. There is almost nothing about what the engagement itself is actually like once the contract is signed - which is odd, because that is the part you are buying.

The gap matters. Most AI programmes that disappoint do not fail at the selection stage. They fail because nobody agreed, in writing, what each phase had to produce before the next one started. Below is the shape we use, and the shape most competent firms use under different names.

Phase one: framing, and the question of whether to proceed.

A good engagement starts by trying to talk you out of the expensive version. The first two to three weeks should establish what decision or process you are actually trying to improve, what it is worth in dollars if it improves, and what data would be needed to do it. That is it. No architecture, no model selection, no platform.

The output is a written recommendation that includes the option of stopping. If your consultant's framing phase has never once concluded "do not build this yet", you are not being advised, you are being onboarded. We set out the four situations where an external consultant earns its cost and the three where hiring or waiting beats it in when to bring in an AI consultant.

Phase two: the readiness check nobody wants to run.

This is where AI programmes actually die, and it is the phase most often skipped because it is unglamorous. Before anything is built, someone has to establish whether the data behind the use case is accurate, connected, permissioned and available at the frequency the use case needs.

The answer is frequently no. That is not a failure of the engagement; it is the engagement doing its job early rather than late. The expensive version of this discovery is finding out in month five, with a model built and no trustworthy data to run it on. A structured AI readiness assessment scores data foundations, use cases, skills, technology and governance, and produces a gap list with costs attached, so the decision to proceed is made on numbers rather than optimism.

Phase three: build the thin slice, not the platform.

The strongest single predictor of an AI engagement that ships is that it was scoped along one decision, end to end, rather than across the whole data estate. One workflow, one dataset, one audience, running in production - not a pilot in a notebook that someone demos and nobody uses.

Two things separate a thin slice from a proof of concept. A thin slice runs on governed data with real access controls, and it has a named human accountable for its output. A proof of concept has neither, which is why so many of them never graduate. If the plan describes a "pilot" without saying what would make it production, you are buying a demonstration.

Phase four: handover, or you have bought a dependency.

The last phase is the one most likely to be quietly dropped when timelines slip, and it is the one that determines whether you own anything at the end. Handover means your team can run, change and debug the thing without the consultancy in the room: documented pipelines, transferred repositories, a runbook, and at least one cycle where your people operate it while the consultants watch.

Ask what handover looks like during procurement, not at the end. A firm whose commercial model depends on you never quite being able to run it yourself will answer that question vaguely.

The five things a statement of work must actually say.

Most disputes we hear about trace to one of five omissions. Whoever you engage, insist the contract is explicit on all five:

One - who is in the room. Named people, with their seniority and the proportion of their time. Pitch teams and delivery teams are frequently different teams. Two - what happens if the data is not ready. The most likely finding in phase two needs a pre-agreed path, price and exit, or it becomes a variation negotiated from a weak position. Three - what software is being resold. Independence is a factual question about revenue, not a values statement. Ask what the firm earns if you pick a particular platform. Four - how the work passes governance. Whichever regime applies to you, the obligation is yours and the evidence has to be produced by the delivery, not retrofitted after it. Five - what you own and how you take it. Code, models, prompts, documentation and data, named, with the handover milestone that transfers them.

Those five are also the five questions we suggest asking every firm you shortlist. Our methodology-led comparison of the top AI consulting firms in Australia applies them across specialist independents, the Big Four and boutique machine-learning shops, including where RUBIX is the wrong choice. If the commercial models are what you are weighing up, the four you are likely to be offered are set out in what AI consulting costs in Australia.

How long the whole thing should take.

For a single well-defined use case, framing is two to three weeks, readiness is three to four, the thin slice is six to twelve depending on the state of the data, and handover is two. Call it three to five months to something running in production that your team owns.

Programmes quoted at eighteen months before a single thing reaches production are not more thorough. They are structured so that the first honest test of whether the approach works arrives after the budget is committed. Whatever the size of the organisation, the correct number of production deliverables in the first quarter is at least one.

RUBIX runs this shape from Melbourne and Sydney for clients across the country - see AI consulting in Australia for the national practice, or AI consulting in Melbourne for the local team.

General information only. Timeframes are indicative and depend on the state of your data. Current as at September 2026.