The questions Australian leaders are actually asking about AI.

Why hasn't Australia's AI usage lead turned into a productivity bump?

Australia now ranks first of 121 countries on Anthropic's population-adjusted usage index, using Claude at more than six times the rate you'd expect for a country of its size, according to Anthropic's Economic Index update reported by SmartCompany on 7 July 2026. That is not a marginal lead, and it has moved fast: an earlier Anthropic release in March 2026 had Australia at 4.1 times the expected rate and 7th globally. Either figure should show up in productivity statistics within a year or two. It mostly hasn't.

The gap is worth sitting with, because it points to a mistake a lot of Australian leadership teams are making right now: treating usage as the goal, when usage was only ever meant to be the input.

Here's what's actually happening inside most organisations. An employee opens a chatbot to draft an email, summarise a report, or get a second opinion on a spreadsheet formula. It saves them ten minutes. They do it again the next day, and the day after. Multiply that across a workforce and you get exactly the kind of usage figures that put Australia on top of the index. But ten minutes saved by one person doing one task, repeated in isolation, does not add up to a measured operational gain for the business. It shows up as a slightly less frustrating afternoon for that employee, and nowhere else.

That's the core problem: individual, bottom-up AI use is additive at best. It doesn't compound, because nobody has gone back and redesigned the workflow around the fact that a step now takes ten minutes instead of thirty. The approval chain is the same. The handoffs between teams are the same. The reporting cadence is the same. The AI use is bolted onto an unchanged process, so the time saved gets absorbed rather than harvested. High tool usage without process redesign is a leading indicator of wasted opportunity, not realised value.

This is where the vanity metric problem comes in. Seats provisioned, logins per week, prompts per user, these are adoption metrics. They tell you people are curious and the tool is accessible. They tell you almost nothing about whether the business is running leaner, faster, or cheaper as a result. The metrics that actually matter, hours saved per process, cost per unit of output, cycle time from request to delivery, require someone to have gone in and rebuilt the workflow so the saved time falls out the other end as a measurable result. Most Australian firms have the first set of numbers. Very few have started tracking the second.

The distinction that matters for any leader looking at this right now is between AI usage that emerged bottom-up and AI usage that was designed top-down. Bottom-up usage is what you get when you give staff a tool and let them find their own applications for it. It's valuable, and it should be encouraged, but it is inherently limited to the scope of an individual's job. Top-down redesign is different. It means someone looked at an end-to-end process, whether that's a client onboarding flow, a monthly reporting cycle, or a supply chain reconciliation task, and asked which steps could be removed entirely, not just done faster. That's a structurally different exercise, and it's the one that actually shows up in cost per unit of output.

The practical takeaway for Australian leadership is straightforward: run an internal audit that maps where AI usage is happening against where workflows have actually been redesigned. In most organisations, that overlap will be small. The usage will be everywhere, scattered across individual habits and undocumented workarounds. The redesign will be almost nowhere, confined to perhaps one or two flagship projects that got executive attention. That gap, the difference between where the tool is being used and where the process has been rebuilt around it, is the real productivity opportunity sitting unclaimed in every Australian business with a high usage score and a flat productivity line. Treat the usage ranking as evidence of capability and appetite, not as evidence of return. The return still has to be built.

The hidden cost of shadow AI subscriptions

Ask most Australian finance teams what the business spends on AI tools each month and you'll get a confident answer built entirely on the wrong number. It'll cover the enterprise platform contract, maybe a Microsoft Copilot rollout, possibly an OpenAI enterprise agreement. What it won't cover is the ChatGPT Plus subscription the marketing coordinator expensed eight months ago, the coding assistant the dev team lead put on a personal card and forgot to mention, the AI transcription tool a project manager found useful and now can't work without, and a dozen other subscriptions scattered across expense reports, personal cards, and departmental credit cards that never made it into any procurement conversation.

This is shadow IT, rebuilt for the AI era, and it's arguably worse than the version finance teams already know how to manage. The classic shadow IT problem, unauthorised software showing up on the network, at least tends to involve tools substantial enough to trip a security review or an IT ticket. Shadow AI subscriptions don't. They're $20 here, $40 there, sometimes $60 for a more capable tier, and that price point is precisely the trap. A recurring charge that size doesn't clear anyone's approval threshold. It doesn't get flagged in a monthly expense review. It sits below the line of anything anyone bothers to scrutinise, whether one person has signed up or two hundred people across the business have, independently, each made the same quiet decision.

The psychology here deserves naming directly, because it's the whole mechanism. Nobody sits down and decides to commit their employer to thousands of dollars a year in AI tooling. What actually happens is a much smaller decision, repeated many times: this is cheap enough that I don't need to ask. That decision is individually rational and collectively expensive. The scale of the underlying habit is real: card-transaction data from business platform Weel, cited in Scale Suite's analysis of Australian SME AI adoption, puts 30.8 per cent of small and medium businesses as actively paying for AI tools on a card, a figure that by its nature sits below whatever finance believes it has approved.

The result is a business paying full retail, one seat at a time, for tools it could be buying at volume, with none of the leverage that comes from negotiating a single enterprise agreement. No consolidated seat count means no consolidated renewal date, which means no single moment each year where someone reviews whether the spend is justified, whether the tools are still being used, or whether a competitor platform would do the job for less. Contrast that with an enterprise-anchored approach: one or two platforms, licensed at a negotiated rate, with a defined number of seats, a known renewal date, and an owner accountable for whether the spend is delivering a return. That structure doesn't just save money on the subscription line, though it usually does. It creates the single point of accountability that shadow subscriptions structurally prevent. When ten different people are relying on ten different personal subscriptions to do overlapping parts of their jobs, nobody owns the question of whether the money is well spent, because no one person can see the whole picture.

The same Scale Suite analysis notes that nearly half of Australian businesses using AI do not measure its impact at all, which is the cost side of the same coin: because so much of this tool adoption happens bottom-up and informally, the organisation loses visibility not just into spend but into which tools are actually earning their keep versus which ones were tried once and quietly abandoned while the subscription kept renewing.

The fix isn't a blanket ban on individual initiative, staff finding useful tools is a good sign, not a problem. The fix is a light-touch registration process: any AI subscription expensed or reimbursed gets logged centrally, reviewed quarterly, and where there's genuine overlap, consolidated into the enterprise platform with proper procurement terms. Australian firms serious about AI cost control should run that audit now, before the shadow ledger gets any longer, and before renewal season locks in another twelve months of subscriptions nobody remembers approving.

Back office first, or the real business? Where Australian firms are pointing their AI

Walk into most small and medium Australian businesses running AI tools today and you'll find the same pattern: drafting emails, summarising documents, tidying up data entry, generating first pass content. Research into SME AI adoption in Australia from Scale Suite puts it plainly: Australian usage over-indexes on back office work such as invoicing and administration, and under-indexes on the coding and product-development use cases that show up more in global averages.

None of that is a criticism of the firms doing it. Back office, text generation AI use is the sensible entry point, and there are good reasons it's where adoption starts. It's low risk. Getting a drafting assistant to write a first pass client email doesn't threaten to break anything if it's imperfect, because a human still reads and sends it. It's low cost to implement, often requiring nothing more than a subscription and some staff familiarisation. And it delivers a fast, visible win, the kind of quick productivity gain that builds internal confidence and justifies the next step. As a starting point, it makes complete sense.

The problem is when it becomes the whole strategy rather than the opening move. Back office AI use, no matter how thoroughly it's rolled out, caps out at a fairly modest ceiling. You can only save so much time on emails and summaries before you've captured most of the available gain. The larger structural opportunities sit elsewhere, in the parts of the business that back office tools never touch: the core product or service itself, the way work actually moves through delivery and fulfilment, and the logic that governs supply chain and operations decisions.

Consider what a deeper application of the same technology looks like. Programmatic automation applied to genuinely repetitive operational workflows, not just drafting text, but executing multi-step processes end to end. Predictive modelling applied to supply chain and inventory decisions, so a business is anticipating demand shifts and input costs rather than reacting to them after the fact. AI embedded directly into the product or service a customer receives, changing what's actually being sold rather than just how quickly internal paperwork gets done. These applications require more investment, more technical integration, and more organisational change than opening a chatbot tab. That's precisely why they represent the larger opportunity: the barrier to entry filters out competitors unwilling to do the harder work.

This is the real risk for firms that stop at the back office. It isn't that email drafting was a wasted effort, it wasn't. It's that treating administrative AI use as the finish line, rather than the first lap, cedes the larger prize to whichever competitor in the same market is willing to push further. A firm that automates its inventory forecasting or embeds AI driven personalisation into its actual service delivery isn't just saving staff time, it's changing its cost structure and its customer proposition in ways a rival still drafting emails with AI cannot easily match. In a market where Australian businesses are demonstrably comfortable adopting these tools, per Anthropic's own usage data showing Australia now first of 121 countries on population-adjusted use, the differentiator won't be who uses AI. It will be who used it on something that mattered structurally.

The practical takeaway for Australian business leaders: treat current back office AI use as evidence the organisation can adopt the technology, not as evidence it has captured the opportunity. The next planning cycle should include a deliberate look at one operational or product facing process, supply chain, service delivery, core production logic, where AI could be applied beyond text generation. That's where the ceiling is much higher, and where the businesses willing to make the harder investment now will be difficult to catch later.

Could your AI workflows survive a privacy commissioner's audit?

Walk into most mid-sized Australian firms and ask to see the AI policy, and you'll usually get one of two things: a paragraph in the staff handbook telling people not to paste client data into ChatGPT, or a one-page PDF that was drafted in a hurry after someone in finance did exactly that. Ask a follow-up question, though, and the picture changes fast. Which AI tools are actually in use across the business? What data has gone into them, and when? Who signed off on using AI to help draft a credit decision, a hiring shortlist, or a client recommendation? Who reviews that decision if something goes wrong?

For most organisations, there's no answer, and the deadline for that to be good enough is now on the calendar. The Office of the Australian Information Commissioner's automated decision-making transparency obligation under APP 1 commences 10 December 2026, requiring entities to disclose in their privacy policy what personal information feeds any automated decision-making system and what kinds of decisions it makes where a person's rights or interests are significantly affected. AI-assisted lending, hiring and eligibility calls fall squarely inside that description.

MinterEllison's analysis of the OAIC's enforcement posture is blunt about the shape of the problem: a written policy that isn't backed by an enforced process is not governance, it's a liability shield with nothing behind it. A Slack message telling staff to be careful with client data doesn't create a record. It doesn't tell you which tool was used, what was uploaded, or whether a human reviewed the output before it reached a customer. Once the December 2026 disclosure obligation is live, a policy document with no logs behind it is not a strong answer to a regulator's first follow-up question.

This matters more this year than it did two years ago, simply because of volume. AI tools are no longer a novelty a few staff experiment with. They're embedded in customer service, document drafting, recruitment screening, credit and lending workflows, and marketing personalisation across a huge share of Australian businesses. Every one of those workflows now involves a decision about what data goes into a third party model, where that data is processed and stored, and who is accountable if the output is wrong, biased, or leaked. The OAIC's own consultation paper puts this directly to entities: can you name who owns AI risk in this organisation, and can you show how a high stakes AI assisted decision was reviewed before it went out the door? Most boards can't answer either question with anything more solid than trusting their staff to use good judgement.

That's the real exposure. It isn't that Australian businesses are reckless with AI. It's that they've adopted the tools faster than they've built the accountability structure around them, and a regulator conducting a review doesn't care how careful your staff intended to be. It cares whether you can produce a trail: what data went where, who approved it, and what review happened before it affected a customer or a decision.

Serious organisations that have thought this through tend to converge on the same four elements, and none of them require enterprise grade tooling to start.

First, data classification before AI use: knowing, before a tool is even switched on, which categories of data, client identifying information, financial records, health information, anything covered by a confidentiality agreement, are permitted to touch which AI systems, and which are not, full stop.

Second, logging: a record of what was submitted to an AI tool, by whom, and for what purpose, so that "we don't know" is never the answer when something is questioned later.

Third, a nominated accountable owner. Not a committee, not "IT," a named person whose job includes AI risk and who can be asked direct questions about how a specific decision was made.

Fourth, periodic review, particularly for AI assisted decisions that affect customers directly: lending, hiring, pricing, eligibility. If no human ever checks the output pattern over time, nobody will notice drift or error until a complaint or an audit forces the issue.

None of this is legal advice, and this masthead isn't qualified to give it. What it is, is a reasonable summary of the operational maturity gap between a policy on paper and a process that would hold up under real scrutiny. Whether your current setup would survive contact with a privacy commissioner's audit, or a determined plaintiff's lawyer, is a question worth putting to your board and your legal counsel directly, not one to leave to an assumption that having a policy is the same thing as being covered.

Data leaks and hallucinations: the case for sovereign AI infrastructure

Ask most Australian executives what worries them about AI and you'll get one of two answers. Either what happens if sensitive data ends up somewhere we can't control, or what happens if the model just makes something up and someone acts on it. Treated separately, both sound manageable. Treated together, which is how they actually show up in production systems, they describe a genuine structural weakness in how a lot of Australian businesses have deployed AI over the last two years.

The first risk is data residency. When staff or systems send customer records, financial data, or operational detail into a public cloud AI service, that data often crosses borders, gets processed on infrastructure outside Australian jurisdiction, and sits under retention and training use terms that are not always transparent or negotiable at standard commercial tiers. Hard local research quantifying exactly how many Australian businesses have thought this through is still thin on the ground, which is itself worth noting: this is a live governance question ahead of the data most organisations have to answer it with.

The second risk is hallucination: AI systems generating plausible, confidently worded, and wrong information. That's a known limitation of large language models, not a bug that gets patched away. It's manageable in a low stakes context, a drafting assistant a human checks before sending. It's a genuine liability in a customer facing chatbot giving account or eligibility information, or a decision support tool feeding a lending or compliance judgement, where a plausible wrong answer can go straight to a customer or into a file without anyone catching it.

Individually, either risk is a governance problem. Together, they compound: sensitive data leaves the country to reach a system that can also confidently give the wrong answer back, and the organisation often has limited visibility into either failure once it's happened.

This isn't only a risk management story, it's a live infrastructure one. EY's Australian economics team puts the scale of AI infrastructure investment at $31 to 38 billion in additional capital expenditure over the decade to 2036, on top of a projected $95 to 116 billion GDP uplift, with Australia already ranked second globally for data centre investment in 2024 at roughly US$6.7 billion and further facilities worth over $100 billion announced through 2025. Where that infrastructure is built and who controls it is a live question for national economic competitiveness, not a side issue for IT departments, and it sits next to, though it is not the same question as, whether an individual business's own data stays onshore.

That distinction matters. Data centre investment landing in Australia is a national infrastructure story. Whether a specific AI vendor's contract gives your business a clear, enforceable answer about where your data is processed, how long it's retained, and whether it trains someone else's model is a separate, contractual question, and it's the one your business actually needs answered.

None of this means every Australian business needs to abandon public cloud AI tools tomorrow, and this masthead isn't in a position to tell you what your specific regulatory or contractual obligations require, that's a conversation for your legal counsel. What it does mean is that data residency and model reliability are no longer separate line items to review once a year. They're linked risks that show up together in the same workflows, and serious organisations are starting to ask vendors direct questions before signing: where is our data processed, who can access it, what happens when the model is wrong, and is there a local, sovereign alternative on the table. If your organisation hasn't asked a cloud AI vendor those questions directly, that's a reasonable place to start.

"AI-washing": the new director liability

Every investor update this year seems to mention AI. Product pages describe "AI powered" features. Pitch decks lead with machine learning capability as a headline differentiator. Some of that is real. A meaningful share of it is aspirational at best, a rebadged rules engine, a basic automation script, or a genuine AI feature that does far less than the marketing copy implies. The gap between what's claimed and what the software actually does is widening, and it's widening in public, in documents that shareholders, regulators, and journalists can all read.

That should sound familiar, because it's the same pattern that produced greenwashing enforcement over the past several years. Companies made public environmental claims that ran ahead of what their operations actually delivered, statements about carbon neutrality, sustainable sourcing, or emissions reductions that didn't hold up to scrutiny. Regulators and shareholders responded, and greenwashing shifted from a reputational irritation to a genuine source of legal and regulatory exposure for directors and executives who signed off on the claims or let the marketing team run ahead of the substance.

AI-washing is following the identical structural logic, and Australian lawyers are now saying so directly. Corrs Chambers Westgarth, writing via Mondaq, points to the ACCC's own December 2023 greenwashing guidance as a direct template regulators are expected to adapt for AI claims. A company claims AI capability in an investor update, a product listing, or a marketing campaign. If the underlying technology doesn't substantiate the claim, that's a public, provable gap between representation and reality, exactly the kind of gap that gave greenwashing litigation and enforcement its teeth. The mechanism doesn't need new law to bite: existing rules against misleading and deceptive conduct already cover claims that aren't true, regardless of whether the word AI is in them. The Corrs authors put the standard plainly: claims about AI use should be accurate and truthful, genuinely reflect the actual level of AI involved, and not hide or omit important information.

Put that standard directly to your own boardroom: can the board point to evidence, not marketing copy, but actual technical substantiation, for every AI claim the company has made publicly? If a shareholder or a regulator asked your company to prove that its AI powered product genuinely uses the technology it claims, in the way it claims, could you produce that evidence today? For a lot of boards, the honest answer is that nobody has ever been asked to check.

The reason greenwashing enforcement reached directors personally, rather than stopping at the marketing team, is that public claims about a company's capabilities or practices are disclosures the board is ultimately responsible for, particularly when they appear in investor communications or influence how the market values the business. The same logic applies to AI claims. A CEO letter or an ASX announcement that overstates AI capability isn't a copywriting problem if it turns out to be wrong, it's a disclosure problem, and disclosure problems sit with the board and named executives, not with whoever wrote the press release.

The practical response mirrors how serious organisations already handle financial and ESG claims: nothing goes out the door that can't be substantiated on request. If your ESG claims require an evidence file before they're published, your AI claims should meet the same bar. That means someone technical, not just someone in marketing, needs to sign off on what the product actually does before a claim about it becomes public. It means keeping a record of the evidence behind any AI capability claim, the same way a company keeps the evidence behind an emissions reduction claim. And it means boards should be asking, as a standing governance item, whether every public AI claim currently in market could be defended if someone asked for the receipts.

This isn't a legal verdict on where AI-washing enforcement in Australia will land, and this masthead can't tell you whether a specific claim your business has made would survive scrutiny, that's precisely the kind of question to put to legal counsel before the market puts it to you instead. But the direction of travel is clear enough from the greenwashing precedent: the gap between what's claimed and what's real doesn't stay hidden indefinitely, and by the time it surfaces, it's usually not the marketing team explaining it.

Augmentation, not automation: why Australian workers want AI as a co-pilot

The default assumption in a lot of AI strategy decks is that automation means removal: find the task, find the human doing it, replace one with the other. Australian workplace data doesn't back that up. Scale Suite's analysis of AI adoption in Australian SMEs puts more than half of Australian AI usage in the "augmentation" category, meaning a human stays in the loop rather than handing the task over entirely. Anthropic's own July 2026 data, reported by ChannelLife Australia, puts a precise figure on it: 54.5 per cent of Australian Claude usage is augmentation, against a 51.4 per cent global average, alongside Australia's now-first-of-121-countries ranking on population-adjusted usage.

That's not a technology limitation. It's a choice, and a consistent one, made at scale by workers across sectors. It says something about what Australians actually want from these tools: speed and leverage, yes, but not at the cost of losing sight of the judgment calls. When more than half of all interactions keep a human in the loop, that's the market telling you how it wants the product built.

Most local AI strategy still doesn't reflect this. Plenty of firms are still framed around a replace the analyst with an agent logic, chasing headcount reduction as the primary return on investment. The data suggests that's solving for the wrong outcome, at least in the Australian context. Workers here are already telling you, through revealed preference, that the winning model is an analyst working with an AI collaborator, not an agent working alone with the analyst cut out of the loop.

There's a practical reason this matters beyond worker preference. AI tools are still wrong often enough, and wrong in ways that are hard to predict, that full hand-off carries real quality and reputational risk. A model that drafts a client email, flags a risk in a contract, or pulls together a first pass financial summary is doing useful work. A model that sends the email, signs off the contract interpretation, or issues the financial summary without a human checking it is taking on liability that the business, not the AI vendor, will wear. Keeping a person in the loop isn't friction. It's the control point that makes fast AI assisted output safe to actually use.

This is where it becomes a genuine competitive question rather than just a workplace culture note. Firms that design roles around AI doing the first draft, human making the judgment call and owning the outcome, get both things at once: the speed gain from automation and the trust and quality control that comes from keeping accountability with a person. Firms that skip the human accountability step to chase a bigger automation number are trading a real, durable asset, client and market trust, for a short term efficiency metric that doesn't show up the same way on a balance sheet.

The practical shift for Australian firms is in how roles get written, not just what software gets bought. An AI augmented analyst role should specify what the AI drafts and what the human is accountable for signing off on. Performance reviews and KPIs should reward good judgment calls on AI output, not just volume of tasks completed. And when firms are evaluating a new AI tool, the fit question shouldn't just be how much can this automate, but does this let our people do more, faster, while staying answerable for the result.

The takeaway: don't read Australia's preference for human in the loop AI use as a lag in adoption. It's a maturity signal. Firms that build roles around AI as a co-pilot, with clear human accountability for judgment calls, will move fast without breaking the trust that client relationships and regulatory standing depend on. That combination, speed plus accountability, is the actual competitive edge on offer here, and it's currently being left on the table by anyone still measuring AI success purely by how much human involvement they've managed to strip out.

Which Australian sectors are actually ahead on AI? Be careful trusting the answer

Ask most business leaders where their sector sits on AI adoption and you'll get one of two answers: everyone is behind, or everyone is ahead. Both are usually wrong, and both let a business avoid the harder question of where it actually sits relative to its direct competitors.

The clearest real picture available sits in financial services. KPMG's 2026 AI in Finance report, drawn from 1,013 senior finance leaders across 20 countries, finds active AI use in finance functions has more than doubled in two years, with over three-quarters of organisations now using it in financial planning, reporting and commercial analysis. But KPMG's own framing is the more important finding: adoption breadth has stopped being the differentiator. Only 23 per cent of organisations report AI is exceeding expectations, a narrower figure than the headline adoption numbers suggest, and the leaders are the ones building governance, measurement and a workforce equipped to act on what AI produces around that adoption, not the ones with the highest usage count.

That distinction, between having adopted AI and having actually operationalised it well, is the one worth applying honestly to your own sector, whatever it is. It's tempting to reach for a comforting sector-wide story: either everyone in my industry is ahead of me, or nobody is, so there's no urgency. Both stories let a business avoid checking its own position. Sector-wide adoption figures, where they exist and are well sourced, tell you the average. They don't tell you where the specific firms you compete against actually sit, and a firm coasting on its sector's reputation for being fast-moving can be just as exposed as one in a genuinely slow sector, if it hasn't built the governance and measurement discipline KPMG's data says actually separates the leaders.

This unevenness matters most for firms that assume their sector's pace is their own. A sector wide lag, real or assumed, looks from the inside like safety in numbers: if nobody around you is doing it, there's no pressure to move. But that's exactly the condition that produces a first mover window. When AI driven operational integration hasn't become the norm in a sector yet, the first firm in that sector to genuinely operationalise it, not just trial it, gets to set the pace before competitors even recognise the shift is happening.

What this calls for isn't a general AI strategy refresh built on a sector-wide narrative. It's an honest, specific benchmarking exercise: where does this business actually sit against its direct competitors, not a sector average, on the adoption-versus-operationalisation gap KPMG identifies. That means looking past whether the business has adopted a tool and asking whether it has built the governance and measurement discipline to actually convert that adoption into results that beat the 23 per cent who are currently exceeding expectations.

The takeaway: treat any claim about your sector's AI maturity, including this one, as a starting point for checking your own firm's position, not a substitute for it. The businesses that will pull ahead are the ones that closed the adoption-to-results gap first, regardless of what their sector average says.

The AI talent gap: build it in-house or pay a consultancy?

There's a real shortage underneath most Australian businesses' AI ambitions, and it isn't a shortage of enthusiasm or budget. It's a shortage of people. Recent Australian hiring data reported by IT Brief puts it starkly: 99 per cent of employers report hiring hurdles for AI and machine learning roles, 36 per cent specifically cite a scarcity of qualified candidates, and 97 per cent are already relying on workarounds, contractors, remote hires, career-changers, rather than filling roles the conventional way. That shortage is exactly what pushes many businesses toward external consultancies for anything beyond basic tool usage.

The consultancy route makes sense as a first move. When a business has no internal AI capability and a pressing need, whether that's a custom model, a data pipeline, or an integration project, bringing in specialists who've solved similar problems elsewhere is faster than building the skill from zero. For genuinely specialised or one off work, deep technical builds that require narrow expertise the business will only need once, that logic holds up well.

The problem is what happens next. Too many businesses treat that first consultancy engagement as the template for every AI initiative going forward, rather than as a bridge to internal capability. Every new use case becomes another external engagement, another invoice, another six week wait for a specialist's availability. The business ends up permanently renting a capability it uses constantly, and every piece of context about how the AI systems actually work inside that specific business, the data quirks, the workflow edge cases, the reasons a particular approach was chosen, walks out the door with the consultants when the engagement ends.

There's an alternative that doesn't require the business to somehow conjure scarce ML engineers out of a tight labour market, and Australia's own workforce data points at it directly. Hays' FY26/27 Salary Guide found that while 60 per cent of Australian tech employees are already using AI at work, only 22 per cent have had any formal AI training, a gap Hays argues employers should close by building internal capability rather than only competing for external hires. As Hays' Pete Marinis put it, professionals are moving toward organisations where they can build capability, not just earn more. Most of the value in ongoing AI capability doesn't come from deep model building expertise. It comes from people who understand the business deeply and are willing to get genuinely competent with the tools, not just curious about them. Existing business line leaders and technically inclined staff, the ones who already understand what the workflow is supposed to achieve and where it currently breaks, are often a faster and cheaper path to durable AI capability than hiring, or renting, specialists who understand the technology but not the business.

Systematically upskilling that group, giving them real training time, access to the right tools, and permission to experiment on live problems rather than side projects, builds something a consultancy engagement never can: institutional knowledge that compounds. The person who built the AI assisted reporting workflow still works there next year. They know why it was built that way, they can extend it when the business changes, and they can train the next person. None of that survives a consultancy handoff document.

Cost control matters here too, and not just in a straightforward hourly rate sense. A consultancy engagement is priced for a defined scope. Internal capability, once built, can be redirected as priorities shift without renegotiating a contract every time. Over a two or three year horizon, a business running most of its AI initiatives through external consultants is very likely paying more in total than one that invested the same money in structured upskilling of people already on staff.

None of this is an argument against consultancies. They still have a clear and valuable role for the genuinely specialised work, the one off technical builds, the situations that need expertise the business will never need often enough to justify hiring for it. The mistake is defaulting to consultancy dependency as the operating model for all AI work, rather than treating it as the right tool for a specific category of problem.

The takeaway: businesses serious about AI capability should be asking a different question than which consultancy do we hire. They should be asking which internal people are positioned to become the business's own AI capability builders, and what it would take to get them there. Reserve the consultancy budget for the work that genuinely needs it, and build the everyday capability where it will actually stay.

Where AI should never make the call

Most conversations about AI strategy are framed as a capability race: what can the technology do now, and how much of the business can be handed over to it next. That framing has a blind spot. As capability keeps expanding, the harder and more important question for any serious organisation isn't how far AI can go. It's where the organisation decides, deliberately and in advance, that it won't go there, regardless of what the technology can technically do.

This shows up clearly in the questions Australian boards are actually being pushed to answer. The Australian Institute of Company Directors' own guidance on AI use by directors and boards sets out a governance framework built on exactly this instinct: before adopting AI, directors are asked whether the board has documented protocols for its use, whether tools have been vetted for confidentiality and IP protection, and whether directors have had proper training rather than picking up the tools informally. The organisations navigating AI well aren't the ones with the most aggressive automation roadmap. They're the ones that have been explicit, internally and often publicly, about where human judgment stays non-negotiable.

Three categories belong on that list regardless of industry or how capable the technology becomes.

The first is high touch client and stakeholder relationships. AI can draft the email, summarise the account history, and flag a risk before a call. It cannot read the room in a tense negotiation, sense when a client relationship needs a different kind of conversation than the data suggests, or carry the accumulated trust of a decade of dealing with someone honestly. Handing relationship management to automation doesn't just risk a bad outcome in one interaction. It signals to clients that the relationship was never really the point, the transaction was.

The second is ultimate ethical oversight and accountability. AI systems can surface a decision, model an outcome, and flag an ethical tension the humans involved might have missed. What they cannot do is be accountable for the decision in any sense that matters, legally, morally, or reputationally. Accountability requires a person who can be asked why, who can weigh factors the system wasn't built to weigh, and who bears the consequences of getting it wrong. An organisation that lets an AI system's output stand as the final ethical judgment has quietly removed the accountability that stakeholders, regulators, and staff actually rely on.

The third is strategic, board level decision making. AI can model scenarios, stress test assumptions, and surface data a board might not otherwise see. It has no stake in the organisation's long term identity, no accountability to shareholders or staff, and no capacity to weigh the kind of values based tradeoffs that sit underneath genuinely strategic calls, the ones about who the organisation is, not just what the numbers say it should do next. Boards that treat AI generated analysis as an input to their judgment are using it well. Boards that treat it as a substitute for their judgment have outsourced the one thing a board actually exists to do.

The instinct to treat these boundaries as a limitation on AI ambition gets it backwards. In a market where every competitor is chasing the same automation gains, a clearly stated, genuinely held line, here is what we will never hand to AI, and here is why, is a differentiator, not a constraint. Customers trust organisations more when they know a human is accountable for the calls that matter most to them. Regulators are more comfortable with organisations that can point to explicit human only zones rather than vague assurances of appropriate oversight. And staff, especially the ones doing the client facing and judgment heavy work, are more likely to stay and perform well somewhere that has told them clearly their judgment isn't being quietly phased out.

The takeaway: don't wait for a crisis to discover where your organisation's human only lines are. Name them now, specifically, across client relationships, ethical accountability, and board level strategy, and put them in writing somewhere the board, staff, and if appropriate your clients can see them. The organisations that will be trusted most as AI capability keeps expanding won't be the ones that automated the most. They'll be the ones that could say, clearly and credibly, exactly what they never would.