A Field Guide to APS AI Adoption

What the data already tells us about how the Australian Public Service is adopting AI, what it isn't telling us, and why the gap between the two matters.

Why this guide exists

The usual question about AI in the Australian Public Service is prescriptive: what should the APS do about AI?

Strategies have been written. Action plans have been published. Capability reviews have been commissioned. The prescriptive conversation is crowded.

A different question is harder to answer and maybe more useful: what is the APS already doing about AI, and how would we know?

This is a measurement problem. The APS is a federation of 155+ agencies with sixteen portfolios, 365,000 employees, and a central workforce database - APSED, maintained by the APSC since 1966, receiving monthly movement files from every agency's HR system - that underwrites the annual State of the Service Report. What APSED supports is an annual statistical publication. Whether it is used, operationally, for a continuous cross-sectional analysis of the cadence of AI adoption? I'm not sure.

Individual agencies know their own hiring. The APSC, through APSED, knows the system - but publishes it annually, in aggregate, with the granularity an annual report affords. Each capability review gives a point-in-time snapshot of a single agency. But the continuous system view - what is happening across the APS this week, this quarter, as it adopts AI - does not exist in any publicly reported form.

What does exist publicly are two datasets, which when read together, provide a more complete picture of APS AI adoption than any currently published source that I'm aware of:

• The Anthropic Economic Index (AEI). Five releases between January 2025 and February 2026, covering 984,000+ real Claude conversations mapped to 19,531 occupational tasks across 165+ countries, with Australia-specific and ACT-specific breakdowns. This tells us what people (including in all likelihood some Australian public servants) are actually doing with AI.
• The APS Employment Gazette. Published weekly by the APSC under Section 40 of the Commissioner's Directions. I've built a structured parser for these resulting in 149,801 vacancy records and 168,861 promotion records spanning January 2010 to March 2026. This can tell us what the APS is actually hiring for, which roles are moving people between agencies, and how the workforce is being structured - at a weekly cadence.

This guide is a practical introduction to what the data reveals, where it falls silent, and what that silence means for anyone trying to understand, plan for, or steer APS AI adoption.

Part I: The Individual Signal

A caveat before the data

The Anthropic Economic Index measures what people are doing with Claude specifically - a general-purpose LLM accessed through chat and API interfaces. It captures the generative text, analysis, and code dimensions of AI adoption, which dominate the personal and professional contexts most public servants encounter day-to-day.

It would be great to see other LLM providers publish similar data. I am not aware of any similar publications from OpenAI or Google or any other major LLM provider though.

It does not capture institutional AI applications that operate on other modalities. Nor is it likely that the Government's official usage of AI from Anthropic would appear in this dataset, as (my guess) is that would explicitly be excluded through commercial agreements and/or other arrangements with LLM providers.

So, with those caveats, the AEI is still the best lens we have for individual behavioural AI usage (but the wrong one for the full picture of institutional AI capability).

Both matter for the analysis that follows; neither is substitutable for the other. Where this guide discusses "AI usage" in the AEI data, it should be read as a proxy for generative-AI adoption, not as a direct measure of government AI capability.

What Australians do with AI

Australia's per-capita AI usage index peaked at 4.22 in August 2025 - the third-highest of any country with 1,000+ observations, behind Israel and Singapore. By February 2026, it had fallen to 3.49 (sixth-highest). Over the same period, the US grew from 3.73 to 3.88, Canada from 3.00 to 3.71, the UK from 2.74 to 3.15.

Two facts to get out of the way about this decline. First, Australia remains well above the GDP-predicted line (172% above what a simple log-GDP regression would predict, ranking 38th of 156 countries) and the absolute level of usage is still high. Second - and this is the part easy to misread - the AEI samples roughly one million conversations per release, so what is falling is Australia's share of a rapidly growing global pie, not necessarily its absolute volume. Over v3→v5 the global sample was held constant (~965k → 989k), but Australia's count dropped 18,753 → 15,906, while the US grew +7%, Canada +27%, and the UK +18% within the same fixed sample.

About a sixth of Australia's drop is a measurement artefact (the "not classified" share of geographies rose from 15.7% to 18.4%); the rest is genuine share loss. Australian Claude usage is almost certainly growing in absolute terms - Anthropic's user base roughly doubled over the same period - but Australia is the only major Anglophone market losing ground in the relative mix.

The more important question than "is Australia still in the top three" is "what are Australians doing with AI, and how does that compare to peers."

Here the data reveals some interesting specific patterns.

Against our peers

Against the global average, Australia appears to dramatically under-index on software development (-3.4 percentage points). While this might look like a "consumer-not-producer" story, it isn't.

The global average is skewed by non-English-speaking countries that use Claude disproportionately for coding, and the confound is tighter than language alone: the top-10 countries by software development share in v5 are Nepal (30%), Tunisia (28%), Egypt (27.5%), India (24%), Sri Lanka (24%), Serbia (23%), Vietnam (21.5%), Turkey (21.5%), Hungary (20%), and Bangladesh (20%) - all non-Anglophone, all low- or middle-income, most with substantial outsourced-developer labour markets. The gap against the global average reflects labour market composition, not a statement about Australian engagement with code.

Against Australia's actual peers - US, UK, Canada - the software development gap disappears entirely:

Australia's distinctive profile, read against actual peers, is not "consumer vs producer." It is: a professionally experienced user base that over-indexes on workplace tasks and under-indexes on some forms of capability building.

The biggest deviations from peers are positive for workplace presentations and IT infrastructure, and negative for STEM homework and job applications.

Self-sufficiency

The AEI measures whether users report being able to complete the task without AI. In the latest release (v5, February 2026), 90.6% of Australian users say yes. That is above the global average of 87.8%, and it is clustered tightly with the rest of the Anglosphere: US 91.3%, UK 90.7%, Canada 90.2%, New Zealand 91.0%. Australia was briefly the highest on this metric in v4 (92.3% vs a peer range of 90.7-91.8%) but has fallen 1.7pp between v4 and v5 while peers held roughly steady.

What this tells us is that Australian users - like their Anglophone peers - are overwhelmingly not using AI to do things they cannot do. They are using it to do things they already know how to do, but faster or better.

AI's potential vs what it's used for

The AEI reports a global average of 12.0 years of education equivalence for what AI demonstrates on task, against a global human-education mean of 11.9 for the users it is working with.

Reconstructed from the published intersection facets, Australia's task-mix-weighted human-education-years is 11.96 and AI-education-years is 12.04, for a gap of 0.08. This is close to the US (0.11) and UK/Canada (both 0.08), but wider than New Zealand (0.04) and Ireland (0.02) - so the gap is small, but not distinctively smaller among Anglophones.

What's more striking is that on a cross-country basis, Australia sits near the bottom of the sampled countries on task-weighted human-education-years (rank 61 of 65 countries with ≥200 observations), along with every other Anglosphere country except Ireland.

The top of that distribution is occupied by markets that use Claude overwhelmingly for coding (Tunisia, Nepal, Sri Lanka, Egypt), which pulls task-weighted complexity up.

Australia's broader task mix mechanically lowers the weighted number - which is a
feature of use diversity, not a capability statement about Australian users.

De-concentration from technology-specific usage

Between August 2025 (v3) and February 2026 (v5), Australia's occupational mix of AI usage shifted. In raw terms, Computer & IT task share fell from 24.2% to 19.8% and Education fell from 13.6% to 7.1%, while Management rose from 1.3% to 2.6%, Sales from 1.6% to 2.5%, and Counselling from 1.6% to 2.5%.

Those raw numbers come with an important caveat. Over the same window, Australia's share of unclassified O*NET tasks rose from 35.7% to 44.5% - an increase of nearly 9pp that has no equivalent in the stable ~7% global unclassified rate. Whatever is driving this (classifier drift between releases, new task types not yet mapped, changes in user query patterns), it inflates the apparent decline in classified categories. Normalising to the classified subset gives a cleaner picture of real movement:

On a like-for-like basis, the Computer & IT decline is about half what the raw numbers suggest - real but modest. The Education decline is genuine and large.

The directional story is that AI use is diffusing outward from purely technical categories toward management, sales, and client-facing work. This matters for the APS because the occupational categories that dominate public service employment - management, business operations, advisory, program delivery - are still where AI usage growth is visible.

The "AI only affects tech workers" window is closing.

The ACT-specific signals

The AEI publishes state-level breakdowns for Australia. The ACT breakdown is small (224 conversations in v5) but revealing:

The ACT is not where Australia's AI usage is concentrated. That is NSW (37.2% of usage, per-capita index 1.19) and Victoria (30.9%, index 1.19). The ACT itself sits at a per-capita index of roughly 0.80 - below the national average. Despite housing the Commonwealth public service and having the highest working-age population share of any Australian jurisdiction, per-capita AI adoption in the ACT is meaningfully lower than in Sydney or Melbourne.

Any framing of the ACT as an AI-adoption leader (at least with publicly available AI) within Australia is wrong.

What the ACT data does show is that when AI is used in the territory, it looks different from the national pattern. ACT usage is more professionally oriented, more successful, and more collaborative than the national average: work-use share is highest, task success rate is highest, directive ("tell AI what to do") usage is lowest, and task iteration ("draft-and-refine together") usage is highest. The ACT is under-represented per capita but qualitatively distinctive.

The plausible reading - though one the AEI data alone cannot confirm - is that the ACT's lower overall usage reflects institutional constraints on casual AI use within the APS (device restrictions, procurement friction, governance caution), while the distinctive collaborative pattern reflects the subset of users who do adopt AI treating it as a work tool rather than a toy.

Whatever AI adoption is happening in the ACT is individual, disciplined, and working - but it is also, measurably, smaller in volume per capita than in either Sydney or Melbourne.

What the individual signal implies

Three implications follow from the AEI data for anyone thinking about APS AI adoption, each stated at the level of confidence the data actually supports:

First, the practitioners who are using AI in the ACT are using it well. Canberra's AI users show the most collaborative, highest-success, most work-oriented pattern of any Australian jurisdiction - not because Canberra uses AI more (it doesn't, on a per-capita basis) but because the subset of ACT users who do engage with AI treat it as a work tool. The behavioural substrate for disciplined institutional adoption exists within the ACT workforce that is already using AI.

Second, Australia's AI user base is becoming more personal and less academic, with a smaller and slower shift toward professional use. The coursework collapse is real, but the share it shed flowed mostly into personal use rather than work use. Any expectation that the AEI trends point toward an imminent surge in professional AI adoption within Australia should be calibrated accordingly.

Third, AI use is diffusing outward from purely technical occupational categories, but more slowly than the raw v3-to-v5 numbers suggest once the unclassified-share confound is controlled for. Management, advisory, and program roles - the bulk of APS employment - are where the growth is visible, but the growth is modest, not dramatic.

A fourth implication is structural rather than behavioural: Australia is the only major Anglophone market where absolute Claude usage has declined across the v3-to-v5 window. Whatever is happening at the individual level is happening against a backdrop of contracting, not expanding, engagement with at least this one major frontier AI provider. The AEI measures Claude usage specifically; other providers may show different patterns. But the Australian-specific contraction is a signal that an "early adopter normalising" frame does not fully capture.

These are behavioural observations about individuals. They say nothing about what the institution is doing. For that, we need the other dataset.

Part II: The Institutional Signal

What the APS is actually hiring for

The APS Employment Gazette is a weekly publication of every Commonwealth vacancy, promotion, engagement, and separation that falls within Section 40 of the Commissioner's Directions. The structured dataset covers January 2010 to March 2026: 149,801 vacancy records and 168,861 promotion records.

It is the longest continuous public record of APS workforce composition publicly available. It is also significantly underused: agencies read it for their own vacancies, the APSC summarises it in annual statistics, and very little cross-sectional analysis happens at the system level, at least that is publicly available.

The AI signal in APS hiring is weaker than the Anthropic data on individual usage, but it is growing and it is differentiating.

AI/ML mentions by year

Using strict search terms ("artificial intelligence," "machine learning," "generative AI," "large language model," word-boundary "AI") across description and duties fields:

A genuine pause in 2023-2024 - roughly flat at ~1.2% - gave way to clear acceleration in 2025 and Q1 2026. The share has roughly doubled in eighteen months. It is not plateauing.

The vocabulary shift is more revealing than the count

Between 2023 and 2025, the terms appearing in AI-adjacent vacancies changed significantly:

The combined governance/safety/ethics vocabulary (responsible AI + AI ethics + AI governance + AI safety) tripled from 9 mentions in 2023 to 30 in 2025. "Generative AI" went from zero mentions to 35. This is the linguistic signature of an institution moving from general AI awareness to operational differentiation - distinguishing between generative AI specifically, AI governance, and AI safety as separate concerns.

AI roles skew permanent

Of the 351 AI-mentioning vacancies in 2024-2025, only 2.8% were purely non-ongoing, against 7.3% for the overall pool. When including roles that offer both ongoing and non-ongoing options, the rates converge (41.9% vs 44.0%). The APS is building standing AI capability, not contracting it on a temporary basis.

Top AI-hiring agencies (2023-2025, strict terms)

The AI hiring signal is concentrated:

Defence leads in absolute terms (as expected given its overall hiring volume - 17.5% of all APS vacancies). But ASD, eSafety, and Geoscience Australia are the agencies punching above their weight relative to total vacancy count.

The spatial, environmental, and statistical agencies on this list - DCCEEW, Geoscience Australia, AIHW, Bureau of Meteorology - are hiring for AI applications that look very different from the LLM-based task augmentation the AEI measures. The SOSR documents the pattern directly: machine learning for locust forecasting, computer vision for weed detection, CNNs for coral reef classification, gradient boosting for border drug detection. The gazette AI-mention signal captures both this institutional, domain-specific ML hiring and the more generic "generative AI" hiring that accelerated through 2025. These are separate AI adoption stories with overlapping vocabulary.

What is not on this list: Services Australia, Treasury, APSC, or most line departments. The AI hiring signal has not yet reached most of the APS.

The structural shifts that frame AI adoption

Three broader workforce patterns in the gazette create the context within which AI adoption is happening. None of them appear in standard APS reporting.

Seniority creep

The share of APS3-level vacancies halved from 5.1% (2010) to 2.0% (2025). APS6 grew from 26.9% to 29.9% and is now the single largest classification by vacancy volume. SES vacancies, which barely appeared on the gazette before 2020 (~0.1-0.3%), now constitute 1.6-1.7% of postings. The workforce is hollowing out at the bottom and thickening in the middle and top. AI adoption is landing on a workforce whose composition is already shifting.

Casualisation surge

The share of vacancies that include non-ongoing arrangements more than doubled from 20.4% (2010-2014) to 41.4% (2020-2025). Part-time eligibility quadrupled from 5.8% to 24.2%. Nearly half the APS vacancy pool now offers flexible tenure. This interacts with the AI-roles-skew-ongoing finding in a specific way: the APS is simultaneously casualising its general workforce and building a small, permanent AI capability core.

Multi-location explosion

In 2010, only 5.3% of vacancies listed multiple states in their location field. In 2025, 43.4% do. Canberra-exclusive vacancies fell from 54.3% (2010) to 33.9% (2025). On paper, the APS is decentralising. In practice, the SES labour market remains Canberra-resident, and APSC census data (not gazette data) suggests many multi-location roles are still filled in Canberra regardless of the advertisement. The gazette records where roles are offered, not where they are filled.

Tech has always led policy

Using consistent keyword matching across the full 2010-2026 dataset, technology roles have led policy roles in every year on record - from +2.6pp in 2010 to +6.2pp in 2025. What has also happened is that policy hiring has collapsed from its 2021 peak of 10.6% to 7.4% in 2025, while tech hiring has held steady at 10-14%. The APS is not transforming from policy to tech. Its policy hiring is contracting while tech hiring is stable.

Part III: What the Gazette Reveals That the Annual Reports Don't

The material above is already more detail than most APS AI adoption reporting uses. But it is the vacancy side of the gazette only. The promotion side contains intelligence that is almost entirely absent from current public APS reporting - and it is intelligence that matters directly to AI adoption.

Cross-agency mobility

Workforce mobility is a stated APS priority. The APS Workforce Strategy and the APS AI Plan 2025 both identify it as essential. The APSC publishes official mobility statistics annually in the State of the Service Report, drawn from APSED. At the system level, the APSC does have visibility into APS workforce movements.

What the SOSR does not do, and what APSED's structure does not support publicly, is continuous cross-sectional analysis at unit-record level. APSED is an annual statistical publication cadence sitting on top of a weekly operational data flow. The gazette - which publishes some of the same underlying movement events weekly, in public - makes the weekly cadence readable by anyone who builds the structured dataset.

The gazette contains every promotion of an ongoing APS employee, with both the person's previous agency and their destination agency. Counting promotions where these differ gives a direct measure of cross-agency mobility:

For comparison, the APSC's official mobility rate - published annually in the State of the Service Report and drawn from APSED - measures a different thing: the percentage of the ongoing employee base who moved between agencies in a financial year, including both promotions and at-level transfers, and excluding Machinery of Government moves. That rate was 13.5% (FY2022), 13.6% (FY2023), 12.5% (FY2024), and 10.5% (FY2025).

The absolute figures differ from the gazette-derived rate because they measure a broader set of movements against a different denominator. What they share is the direction of travel: cross-agency mobility peaked in FY2022-23 and has declined each year since. Both the gazette and the APSED-based SOSR confirm the same substantive pattern, from independent starting points.

In the same years that the APS Workforce Strategy identifies mobility as a priority and the APS AI Plan 2025 emphasises cross-agency capability building, the measurable component of mobility is trending in the wrong direction.

This is a slow divergence between stated policy and observed behaviour, visible in both datasets.

The Services Australia → NDIA pipeline

When you look at which specific agencies are losing talent, and where it is going, patterns emerge that would be invisible to any single agency's HR reporting.

Top 10 agencies by outbound cross-agency promotions (2020-2026):

The single largest inter-agency talent flow in the APS is Services Australia → NDIA: 759 promotions over six years. The second-largest is Services Australia → DSS (173). Services Australia is losing frontline staff into the social services cluster at a sustained rate, and Defence is appearing as a top-three destination for nine of the ten largest losing agencies.

This pipeline is not an anomaly; it is visible in the macro headcount data too. The SOSR reports that NDIA had the largest net increase in total APS headcount in FY2024-25 (+2,756 employees), and Services Australia had the second largest (+1,651). The gazette data shows one of the specific mechanisms driving these flows: a sustained promotion pipeline from Services Australia operational staff into NDIA roles at the next classification level. The SOSR tells us the net headcount movement at the annual macro level. The gazette tells us which specific agencies and classifications are feeding it, at what velocity, and for how long the trend has been running.

These patterns are visible years before they show up in workforce reporting. The question for any agency experiencing unexpected attrition is: is this pattern new, or has it been trending for years and nobody noticed?

Central agencies as training academies

The central agencies - PM&C, Treasury, Finance, Attorney-General's Department, the APSC - are among the APS's most critical policy bodies. They are also, measurably, training academies for the rest of the service.

Outbound promotion share (all years, all classifications):

The APSC - the agency responsible for APS workforce policy - loses nearly a third of its promoted staff to other agencies. PM&C, Attorney-General's, and Treasury all exceed 20%. Finance is the lowest at 18.4%.

The APSC publishes aggregate mobility rates in the State of the Service Report. It does not publicly decompose those rates into the bilateral flows that make visible which specific agencies are training staff for which other specific agencies. The 31.6% outbound rate for APSC itself, or the 24.8% for PM&C, is not something the SOSR surfaces.

Two readings are available. The charitable reading is that these agencies develop high-value talent and the rest of the APS recognises it. The less charitable reading is that these agencies invest in capability that they cannot retain, and the incentives to train staff who will leave are weakening. Both readings are compatible with the data. Neither is visible in any current APS reporting.

For AI adoption specifically, this matters. The central agencies are where AI governance capability needs to sit. If they cannot retain staff in whom they invest that capability, the AI-literate workforce of the future is being distributed across the APS by default.

Part IV: The Three Blind Spots

Reading the AEI and the gazette together reveals patterns that neither dataset shows alone. It also reveals three specific blind spots where important questions cannot currently be answered.

Blind spot 1: Individual vs institutional

The AEI shows that Australian AI users - including ACT users specifically - are sophisticated, collaborative, and successful. The capability reviews show that APS institutional AI readiness is at "Developing" (six agencies) or "Emerging" (three agencies) across all elements related to technology and data. No agency achieved "Embedded" for technology. Only the ATO achieved "Embedded" for data.

The gap between individual and institutional AI capability is therefore measured at both ends but unmeasured in the middle. We know what individuals can do. We know what institutions cannot do. We do not know publicly the shape of the gap between them - how much AI-augmented work is happening informally, without institutional sanction or governance, inside the APS right now.

Blind spot 2: Agency vs system

Every APS agency knows its own hiring, its own promotions, its own attrition. The APSC, through APSED, knows the system - but publishes it annually, in aggregate, and without the specific cross-agency flows that would allow an agency to plan around structural talent pipelines it cannot see from inside its own HR system. The gap between the weekly operational cadence of the gazette and the annual publication cadence of SOSR is where the pattern visibility lives.

The cross-agency mobility data in Part III is the kind of intelligence that individual agencies cannot generate from their own HR systems, because the flows they care about are precisely the ones that leave their boundaries. I hope the APSC has the capability to support agencies with these views (from the APSED).

The Services Australia → NDIA pipeline is the clearest example. From Services Australia's perspective, 759 promotions over six years disappeared into an agency they have no workforce visibility into. From NDIA's perspective, 759 people arrived with Services Australia backgrounds. Neither agency can independently calculate the rate, velocity, or trend of this flow. Neither agency's workforce plan accounts for it. The SOSR reports the macro headcount consequence of the flow (+2,756 at NDIA in FY2024-25) without decomposing its mechanism.

This is the kind of blind-spot that AI adoption makes more expensive. If an agency is planning AI adoption that depends on retaining staff in specific classifications, and those classifications are subject to structural outbound flows, the plan will falter.

Blind spot 3: Hiring vs outcomes

Specific APS AI projects are measured rigorously. The SOSR documents HERMES (Home Affairs) referring around 2,000 consignments since July 2023 and stopping over 400 kilograms of border-controlled drugs, with an estimated $300 million in drugs prevented from reaching the community. The ATO's real-time analytics nudge messaging, using k-nearest-neighbour and neural network models, protected an estimated $92.6 million in revenue - validated by a randomised controlled trial showing nudged taxpayers were 2.5x more likely to review their returns and 3x more likely to adjust them. ReefCloud processes coral imagery 700x faster than manual methods with 80-90% accuracy across 4.5 million survey images. These are outcome-linked, methodologically defensible, and independently auditable results.

What is missing is not project-level AI outcome measurement. It is system-level linkage between workforce composition, AI capability, and program outcomes across the service. Home Affairs knows HERMES works. The ATO knows nudge messaging works.

My guess is that the APSC does not systematically know whether the agencies hiring for "AI governance" in 2025 are delivering better outcomes than the agencies that are not.

When an agency adds "AI governance" to a job description, does anybody measure whether AI governance capability is then operationally present?

The project-level evaluation is there (and reported!); the cross-sectional, continuous, service-wide measurement? I guess we'll wait to find out.

This matters because the entire premise of generalised AI adoption - that AI improves the work across the service, not just in the specific projects chosen for flagship evaluation - is untestable without that system-level measurement.

The AEI can tell us that AI performs at 17+ years of education equivalence on legal analysis, and that 70% of tasks globally are judged successful. It cannot tell us whether a policy brief produced with AI augmentation leads to better policy outcomes for citizens, because the counterfactual is missing.

Only institutional outcome measurement can close this gap, and the gazette indicates that evaluation and measurement language appears in a small and stable share of APS role descriptions.

Part V: What This Means in Practice

This guide has assembled a lot of numbers. The question for any reader with operational responsibility is what to do with them.

If you are a Chief AI Officer

The most directly useful signal in this data is the cross-agency mobility pattern. Your AI capability depends on retaining specific people in specific classifications. Running the numbers for your own agency answers three questions that your current workforce plan probably does not:

1. What is my agency's outbound promotion rate, by classification, relative to the APS-wide median? An outbound rate higher than peer agencies indicates structural talent drain.
2. Which destination agencies are my staff going to, and is the pattern new or sustained?
3. Where is my classification pipeline congested? If APS6 into EL1 is running slower than 2.0 years for your agency, your practitioners will leave.

If you are a workforce planner

The gazette's promotion record is effectively a longitudinal dataset of APS careers. It contains baselines for career velocity, classification bottlenecks, and inter-agency flows that - while latent in APSED - do not appear in any currently published source. Using it to construct counterfactuals - what would this agency's attrition look like if we matched the peer average? - is the kind of analytical capability that could transform APS workforce planning from reactive to predictive.

The AEI data adds a layer that workforce planning has not traditionally incorporated: what those careers are actually doing with AI, and how that is changing over time. Reading the AEI's ACT breakdown against the gazette's ACT-based vacancies and promotions gives a picture of where institutional constraints are binding and where they are not.

If you are thinking about AI adoption strategy

The pattern most relevant to you is the mismatch between individual and institutional capability, and the fact that it is currently unmeasured at system level. Any AI adoption plan that does not explicitly address the measurement gap will be unable to distinguish its own success from failure three years from now. The baseline measurement infrastructure - hiring data linked to outcome measurement linked to usage data - does not exist and will not exist by accident.

The second pattern is the de-concentration of AI usage from technical occupations. The AEI data shows that the categories where AI usage is growing fastest are the ones your strategy probably classifies as non-technical: management, advisory, program delivery, client-facing work. The gazette data shows that these categories are also the largest part of your hiring volume.

If you are a central agency

The most important question is the one that appears nowhere in your current reporting: who has visibility across the system at the cadence that matters? Individual CAIOs can manage their own agencies. Individual capability reviews can assess individual departments. The APSC can compile statistics annually from APSED. But the day-to-day operational view - this week's hiring, this quarter's talent flows, this year's emerging attrition patterns - that would be useful, no doubt.

The cost of not having a system-level reader at operational cadence is that the patterns visible in this guide (cross-agency mobility declining, central agencies losing a quarter to a third of promoted staff, AI hiring accelerating unevenly, classification bottlenecks at APS5 and APS4) compound without any official forum noticing until they become visible in the annual report - or later.

Part VI: The Measurement Case

The argument of this guide is not that APS AI adoption is failing. It is not failing. Where the data reaches, it shows genuine progress: AI hiring is accelerating, vocabulary is differentiating, AI roles are skewing permanent, flagship projects like HERMES and ATO nudge messaging are delivering measured outcomes, and the ACT-based workforce is using AI in exactly the disciplined, augmentation-oriented way the public service should want.

The argument is that APS AI adoption is being measured at the wrong cadence - annually, in aggregate - when the data supports weekly, cross-sectional analysis. The individual behavioural data exists (AEI, publicly released, updated quarterly). The institutional hiring data exists (gazette, publicly released, updated weekly). The authoritative cross-agency workforce data exists (APSED, maintained by the APSC, published annually as the SOSR).

The cross-sectional analysis that would turn these into operational intelligence does not exist in any continuously published form, and the patterns most relevant to AI adoption are precisely the ones that require that continuous cross-sectional analysis to see.

This is a gap with a clear shape:

• It is whole-of-government, not agency-specific (agency-level reporting misses the flows)
• It is continuous, not annual (annual reporting misses the velocity)
• It is structured, not narrative (narrative reports miss the patterns)
• It combines datasets that are currently read separately (neither dataset alone answers the questions)

The APS publishes the data. The APSC maintains APSED and structures the gazette specifically for transparency. The AEI publishes its releases publicly. The analytical tools to combine them are mature and practical.

Will someone eventually build this capability?

Methodology note: all AEI figures are drawn from Anthropic Economic Index releases v1 (January 2025) through v5 (February 2026), with minimum observation thresholds of 1,000 conversations for country comparisons and 100 for state-level comparisons. Peer comparisons use US, UK, and Canada as the reference group. All gazette figures are drawn from a structured dataset of 149,801 vacancy records and 168,861 promotion records spanning January 2010 to March 2026. AI/ML keyword matching uses description and duties fields only (excluding boilerplate "about" sections) with strict terms: "artificial intelligence," "machine learning," "generative AI," "large language model," "natural language processing," "computer vision," "deep learning," "neural network," and word-boundary matches for "AI." Cross-agency promotion classification compares from_agency against destination department fields. Career velocity figures are median intervals between matched promotion pairs for the same individual, excluding records with inconsistent or withheld names; these are directional estimates biased toward faster movers because of left-truncation at 2010. Classification bottleneck ratios compare exact-string vacancy matches to promotion counts by destination classification for 2024-2025. APSED mobility rates and NDIA/Services Australia headcount growth figures are drawn from the APSC's State of the Service Report 2024-25, Table 11 and headcount tables. HERMES, ATO nudge messaging, ReefCloud, locust forecasting, and WeedRemeed case study details are drawn from the SOSR 2024-25 innovation case studies. All figures are reproducible from the publicly available AEI data releases, the APS Employment Gazette, and the APSC's State of the Service Report 2024-25.