When to bring in an AI consultant, and when not to.

Almost every article answering this question is published by a firm that sells the service, and they arrive, with impressive consistency, at the conclusion that you should buy the service. This one is published by a firm that sells the service too. So rather than argue the case, here is the test we actually apply before quoting, including the situations where we tell people not to engage us.

The expensive mistake is not choosing the wrong firm. It is engaging anybody at all when the constraint you have is not the one consulting solves. A capable supplier doing competent work against the wrong constraint still produces a bad outcome, and it does it slowly and at a professional rate.

First, diagnose the constraint.

Three quite different problems get described in the same seven words - "we need help with AI" - and they have three different answers.

• Clarity. There is commitment and budget, and no agreed view of what to build or in what order. External help is genuinely valuable here, because the work is comparative: someone who has seen twenty of these can tell you in a few weeks what would take you several months to reason out from first principles. It should be time-boxed and it should end.
• Capability. You know what to build and nobody in the organisation has built one before. External is usually right, on one condition: that capability transfer is a written deliverable with acceptance criteria, not a line in the closing slide.
• Capacity. You know what to build, your own people could build it, and they are committed to other things. This is not a consulting problem. It is a hiring or contracting problem, and buying it at consultancy rates is the most expensive way in the market to solve it.

In our experience most organisations that approach a consultancy have diagnosed a capacity problem as a capability problem. That misdiagnosis is flattering to consultants, which is exactly why you should be wary of a firm that never raises it.

Four situations where bringing someone in earns its cost.

1. There is an executive or board commitment, and no plan underneath it. The risk being managed here is not delivery risk, it is the risk of spending a year and a large budget on the wrong thing. A short, time-boxed assessment that ends in a prioritised plan is cheap insurance against that, and it is one of the few pieces of consulting work whose value is easy to check afterwards: either you can now name what you are building and why, or you cannot. This is what an AI readiness assessment is for.
2. The work requires joining source systems that have never been joined. Integration across systems with no shared identifier is specialist, unglamorous and largely one-off. It is a poor candidate for building internal capability, because the learning is not recovered - you do it once. It is also, reliably, where most of the cost and nearly all of the risk of an AI programme actually sits, well ahead of the modelling. That work is the data foundation, and if a proposal treats it as a preliminary rather than the main event, the proposal has mispriced the job.
3. Regulatory evidence is part of the deliverable. Work touching APRA-regulated operations, public-sector AI obligations or health data carries an evidence burden that is real, billable work rather than overhead. The obligation stays with you, but the artefacts are produced by whoever does the build. Our notes on CPS 230 and AI and on what the NSW AI Assessment Framework asks of your consultant set out what that means in practice. The useful screening question is simple: ask a prospective supplier to describe the evidence pack they will hand you at the end. A firm that has done this before answers immediately and specifically. A firm that has not, generalises.
4. It is a first-of-type build and your team will own it afterwards. External for the first one, internal from the second, is a sound pattern and it is how most durable capability actually gets built. It only works if the handover is scoped as a deliverable - documentation, runbooks, a named person who has operated the thing while the supplier watched - rather than left to the final fortnight. If you want the arrangement to run the other way round, with senior specialists working inside your team to your priorities, that is a different shape again: see forward deployed engineers.

Three situations where it is not.

• The constraint is capacity. If you know precisely what needs doing and simply need more hands doing it, hire or contract. A consultancy delivering business-as-usual work at consultancy rates is the most expensive staffing arrangement available, and it quietly creates a dependency that becomes harder and more costly to unwind each quarter it continues.
• The organisation has not agreed what the priority is. A consultant cannot settle an unresolved disagreement between executives about what matters. What actually happens is that you pay professional rates to hold the argument inside a workshop, and the output is a document everybody signs and nobody uses. That disagreement is resolvable internally, and resolving it first is free.
• You want someone else to be accountable for the outcome. Accountability does not transfer with a contract. This is most explicit in the public sector, where the obligation sits with the agency no matter who performs the work, but it holds commercially too. A supplier can carry delivery risk, and a well-written fixed scope makes it carry a good deal of it. It cannot carry your risk. If the appeal of engaging a firm is that there will be someone else to point at, the engagement has already failed.

Consultant, contractor, or hire?

These get compared on day rate, which is the least informative axis. Compare them on what you are actually buying.

• A permanent hire is the lowest cost per unit of work and by far the slowest to start. Right when the capability is permanent, the work is continuous, and the knowledge has to stay in the building.
• A contractor is fast to start and effort-based. Right when you know exactly what needs doing and need more hands. You direct the work and you own the outcome, which is the trade: speed and flexibility in exchange for carrying the thinking yourself.
• A consultancy is the highest cost per day, and should be bought for the things the other two cannot supply quickly: comparative experience across many similar programmes, a method that has survived contact with reality, a team rather than an individual, and accountability for a named deliverable. If what you are buying is hours, you are overpaying for them.

A blunt but effective sanity check: try to write the job description. If you can write it, you want a hire or a contractor. If you genuinely cannot write it, that difficulty is the consulting problem, and it is worth paying someone to resolve.

The fortnight before you engage anyone.

A surprising share of what makes an engagement expensive is determined before it starts, and almost all of it is within your control rather than the supplier's.

1. Name the decision the work has to inform, and the date it is needed. "We will decide by December whether to build X" is a brief. "We want to explore AI" is not, and it will be priced as though it is not.
2. List the source systems in scope and who owns access to each. This single page changes more proposals than any other document you can produce.
3. Start the access approvals now. Elapsed time waiting for credentials and security review is one of the largest real costs in these engagements and the one most often left out of proposals entirely. It is also the one you can eliminate before day one, for nothing.
4. Name one internally accountable person with authority to decide. Not a steering committee. Engagements slow down at the speed of the slowest decision.
5. Write down what finished looks like, in one sentence, before anyone quotes. If you cannot, that is worth knowing about your own brief, and it is a cheaper thing to discover now.

An organisation that arrives at a first meeting holding those five things gets a smaller, cheaper and more accurate proposal. From any firm, not only from us.

Decide in advance how you will know it worked.

The characteristic failure of AI consulting is not a poor deliverable. It is a good deliverable that changes nothing. Agree the test before the work starts, and make it something observable from outside the project: a decision actually taken, a process running in production and being used by the people it was built for, a named person able to operate and extend the thing without the supplier in the room. A report is an input to a result. It is not one.

It is worth agreeing in advance what happens if the answer comes back "not yet", too. An assessment concluding that your data cannot yet support the use case has done its job and has saved you the far larger spend that would have followed. If a supplier's commercial model makes that conclusion difficult for them to reach, you have learnt something useful about the advice you are going to receive.

How RUBIX approaches this.

We work to fixed-scope engagements rather than open-ended day rates, so the cost and the deliverables are known before we start, and the engagement has a defined end rather than a renewal. These are the four ways clients usually begin, described as durations rather than prices, because the price depends on the state of the data underneath:

• AI readiness assessment - two to four weeks. A prioritised plan before you commit budget. Detail on how we run one.
• Data foundation build - eight to twelve weeks. Core systems joined into one trusted, AI-ready layer.
• AI governance framework - four to eight weeks. Making AI use safe, accountable and defensible. See AI governance consulting.
• Embedded delivery - ongoing, with defined review points rather than an open tap.

If you are not sure which of those describes your situation - or whether you need any of them yet - the free AI readiness snapshot takes about ninety seconds and will show you where the gaps are before anybody quotes you anything. Broader detail on scope and approach sits on our AI consulting in Australia page.

General information only, not commercial, legal or financial advice. Durations are indicative and are not a quote. Current as at August 2026.