On 21 September the Auditor-General published an audit of an AI system that worked and then stopped being used, because nobody could process what it found. It is the clearest recent example of the failure a buyer of AI consulting services most needs to prevent: advice and a build that produce something the business cannot operate. The other common failure is a scope that grows with no ceiling. A good engagement prevents both inside its first 30 days, and you can tell from the proposal whether it will.

We sell this. Our AI consulting services engagement runs for 30 days, is delivered by the team that would build the result, and ends with a costed plan your board can approve, usually with a first working system behind it. Below is how to judge any AI consulting proposal, ours included.

What happened last week, and why buyers should care

The Australian National Audit Office published its audit of AI and Medicare benefits integrity on 21 September. The health department built an AI-enabled model to detect potential Medicare fraud and non-compliance. Preliminary analysis showed the model was efficacious. It was not used after December 2025, because the staff needed to manually assess the signals it generated could not keep up with the volume. Eight potential matters were referred for preliminary analysis. The audit found no considered assessment before development of data suitability, privacy or cyber security, and no analysis of what the system cost to build.

The model worked. The operating plan around it did not exist.

What a good first 30 days actually delivers

Private buyers are moving the same way as the auditors. In an IT Brief interview published 23 September, Cloudera chief executive Charles Sansbury said: “The experimentation without regard to cost is over because people were surprised.” According to the piece, return on investment is now entering initial customer discussions rather than appearing later in a deployment. Sansbury also said almost every company has kicked the can down the road on data quality.

So month one should answer three questions before anyone talks about models: what AI is already in use, what data can go where, and what the business will do with the output. Thirty days should produce:

• An inventory of AI already in use, sanctioned or not, and where its data goes. The ANAO notes the health department identified unrestricted and unapproved AI use, including through a stocktake completed in June 2026.
• Written data rules that say which categories of data may go to which kind of model, under what terms.
• A ranked register of use cases, scored on value, effort, data readiness and risk, with the reasoning shown.
• A costed plan for the top three, with sequence, architecture, oversight points, measurements and a price for each.
• Where the scope allows, a first working system on the highest-value case, so the plan is tested against real data rather than a slide.

One item most proposals leave out: a capacity check on the humans downstream. If a system will flag hundreds of invoices, claims or enquiries a week for review, somebody has to own that queue, and the plan should say who and how many hours it takes. That is the gap the Medicare audit exposed: cheap to close on paper in week three, expensive to discover in month nine.

How AI consulting fees are usually structured

Nobody publishes a reliable price list for AI consulting in Australia. What is public is how professional services get bought, and government procurement rules are a useful guide because they are written down. Four structures cover most engagements:

• Rates. You pay for time at agreed day or hourly rates. Victoria’s professional advisory services panel lists supplier rates as ceiling rates, and tells buyers to check quotes do not exceed them and to negotiate lower. Rates suit open questions and leave all the scope risk with you.
• Fixed fee against a written scope. The same Victorian guidance tells buyers that if a fixed fee is offered, they should request a breakdown to assess whether it is cost efficient, and to seek a cost breakdown of total fees in any case. A fixed fee moves scope risk to the supplier, which is why a supplier who offers one has to scope properly first.
• Staged or modular work packages. Each stage is priced and released on performance. It works when each stage has its own price, its own deliverable and a point where you can stop.
• Milestone payments. The Commonwealth’s Contract Management Guide lists payment arrangements, including milestone payments, as a standard contract element. Payment follows delivery, not the calendar.

Canberra also caps digital contracts outright. The Digital Transformation Agency’s contract limits and reviews policy, mandatory since 1 February 2020, says digital contracts must not exceed $100 million over their whole life, must not exceed a three-year initial term, and that extension options can only be exercised after a review of the contractor’s performance and deliverables. A private company need not copy the numbers, only the principle: a ceiling, a term, and a review before anything is extended.

Red flags in an AI consulting proposal

Weak proposals share a handful of tells. One is survivable. Two together should stop the signature.

A strategy deck with no working system. A roadmap, a maturity score and a list of opportunities is a document someone else must turn into software. Ask what will be running on your data at the end.

No ceiling on the total. Rates with no cap, stages with no total, or a discovery phase whose only output is a proposal for a bigger phase. The Finance guide warns contract managers that a variation can be implemented without being formalised in writing, so what gets agreed in a steering meeting can quietly become scope.

Lock-in dressed as convenience. The supplier’s own platform, the supplier’s accounts, the supplier’s model keys. Ask who owns the code, the prompts, the evaluation data and the cloud accounts on the day the engagement ends. The Finance guide notes that contracts include provisions on who owns the new material created under them. In AI work, the evaluation sets and prompts are that new material, and they are the hard part to rebuild.

No data or security plan. If the proposal does not say which data will be used, where it will be processed and how the system will be tested against misuse, the Medicare audit is the preview. This is why we treat secure AI system development as part of every build rather than a later workstream.

No named people. The Finance guide notes that in professional services contracts, the skills of particular people may be critical to the value for money assessment, which is why contracts name key personnel. If the senior people in the pitch are not named in the contract, assume you will get someone else.

Questions the board or executive team should ask before signing

1. What is the total ceiling for this engagement, and what exactly triggers a variation?
2. What will be running on our data at the end of the first 30 days, and how will we know it works?
3. What does each stage cost, what does it deliver, and where can we stop without penalty?
4. Who owns the code, prompts, evaluation data, documents and cloud accounts when you leave?
5. Which of our data goes to which model, where is it processed, and who approved that?
6. How many staff hours a week will this system’s output need, and whose team absorbs it?
7. Which named people will do the work, and what happens if they leave the engagement?
8. What is the measurement, the baseline, and the date we will judge it by?

A supplier that cannot answer question six has not thought about operating what it wants to build. One that will not answer question one has told you how the contract ends.

How we run AI consulting, and who it is for

We are a build firm that consults, not a consultancy that subcontracts the build. Our 30-day engagement follows the order above. Week one is the inventory and the data rules. Week two ranks the use cases. Week three produces the costed plan for the top three, with a build, buy or integrate call on each. Week four, where the scope allows, is a working first system on the highest-value case.

It is fixed-price from a written scope, not open-ended hourly consulting. You own everything it produces: documents, code, accounts and data. The plan is written so another firm could deliver it, and you are free to take it elsewhere. Our teams are in Sydney, Melbourne and Brisbane.

A word on fit: our build engagements start at A$100,000. If the plan that comes out of the 30 days does not justify a budget of that order, you still keep the inventory, the data rules and the ranked register, and they are yours to use.

Why not do it in-house? You can, if a data engineer, an integration developer and an AI security specialist are sitting idle. Most organisations have none of them spare. Why not a large consultancy? Because at the end of our first month you should have software running on your data, from the people who will be accountable for the rest of it.

Your next step

Start with the free AI audit. It is a 60-second form: describe what your organisation does and what is painful right now, and a written four-part assessment arrives by email within minutes, naming three specific opportunities, what each would cost and take, what to avoid, and the single next move. If you would rather go straight to a conversation, book a meeting and bring the AI question your board keeps asking. You will leave that first meeting with the shape of the engagement and a fixed price for it.

Frequently asked questions

What should AI consulting services deliver in the first 30 days?

At minimum: an inventory of the AI already in use and where its data goes, written data rules, a ranked register of use cases with the reasoning shown, and a costed plan for the top three. Where the scope allows, a good engagement also delivers a first working system on the highest-value case, so the plan is tested against real data.

How are AI consulting fees usually structured in Australia?

The common structures are time at agreed rates, a fixed fee against a written scope, staged work packages released on performance, and milestone payments. Government buyers are told to treat panel rates as ceilings and to ask for a breakdown of any fixed fee. Whatever the structure, insist on a total ceiling and a point where you can stop.

What is the biggest red flag in an AI consulting proposal?

A deliverable that is only a strategy document, with no working system on your data and no total ceiling on cost. Close behind it is lock-in: code, prompts, evaluation data or cloud accounts that stay with the supplier when the engagement ends.

Why do AI systems that work still get switched off?

Usually because the organisation did not plan for the output. The Auditor-General's September 2026 report on Medicare benefits integrity describes an AI model that preliminary analysis showed was efficacious, which stopped being used because the staff needed to review its signals could not keep up. A good plan sizes the human workload before the system is built.

Should we use a large consultancy or a firm that builds?

If you need a broad organisational strategy, a large consultancy may suit. If you need AI working inside your systems, choose a firm that will build and run what it recommends, so the advice is tested against real data and the people who gave it stay accountable for the result.