Cross-note synthesis

Uncovering barriers to practical collaboration

How the published AI CoLab notes, taken together, speak to the Uncovering barriers to practical collaboration evaluation theme. Public, general synthesis — safe to draw on without checking. It generalises the lessons; the sources below link back to the particular workshops (some members-gated).

56 insight notes so far. Not every workshop becomes one — recurring sessions and smaller gatherings usually don't. For some of the larger workshops a transcript is turned into an insight note so its lessons can be shared — with the room's consent, through time, and with people who weren't there.

The barriers that bind are upstream of the technology. The core every note shares is institutional: limits on which tools teams may use at all — platform mandates in some agencies (AI for Insight Notes), security frameworks and cost ceilings constraining choices in delivery work (AI for the Systems You Inherited). Practical AI for Policy People puts the same point to public servants directly — which tool is sanctioned for which sensitivity (protected versus public), guidance on the use of cabinet material, and access itself gated by whether an agency has signed onto the shared service — with the standing reminder that what is technically possible is not an endorsement to do it. What the model can do matters less than what the organisation permits, can afford, and has data ready for. From Pilot to Production shows the same barrier as the thing you must engineer around to ship: quick point-to-point connectors that prototype well rarely survive a production security review, and anything near primary casework faces a slower, heavier governance bar than an internal sandbox — so the workable path is a deliberate two-speed split (fast internal experimentation kept separate from the governed production track) rather than turning AI on everywhere at once.

This was the canonical finding of the program's earliest public-sector sessions. They located the obstacle in structures, incentives and culture rather than the technology — the risk that without changing them an agency keeps "writing the same report in five years" (AI and the future of the public service); a risk-averse culture with siloed, informal use under the shadow of a past automated-decision failure (Generative AI in government policy work); a "gravity well" of bureaucratic rationality that pulls initiatives back and leaves them fragile to a change of leadership (Scaling and sustaining innovation); access as a whole-of-organisation IT and data-sharing hurdle rather than an individual one (The NSW Trend Atlas); jurisdictional misalignment where a problem crosses boundaries the institutions do not (AI for problems humans can't solve); legacy-system rigidity that constrains what can be changed at all (Rethinking the decision landscape); procurement and compliance load and the fragility of a single champion, alongside a felt need for assurance over large models that pushes some teams toward methods avoiding external APIs (Practical AI for risk professionals, Clustering techniques); and the absence of inventories, clear ownership and cross-agency frameworks (Auditing AI). The 2026 delivery notes extend this map rather than discover it.

Each domain then adds a layer the others would miss. Unlocking career potential shows the human-process layer: language fluency, accent and writing style treated as proxies for capability, and uneven scrutiny of diverse applicants. Growing with AI adds the physical and economic layer software-centric thinking overlooks — connectivity, power, ruggedness, high and unpredictable cost, vendor lock-in, supplier failure — plus structural data friction: no benefit-sharing models, no common standards, on-farm systems that do not interoperate. AI for the Systems You Inherited adds the institutional-dynamics layer: sunk-cost pressure to keep failing projects alive, and the unsolved problem of governing how agents interact. Beyond the Principles adds the evidence-and-governance layer: groups would not recommend either a human or an AI process without comparative studies that do not yet exist, smaller agencies questioned whether an independent review function is even affordable, procured black-box tools resist the explainability good decisions need, and ungoverned "shadow" AI use is already outrunning the rules meant to contain it. When the Algorithm Goes Rogue adds the political-economy layer that precedes any build: a business case built on point-estimate savings and pre-announced staff cuts, limited-tender procurement that rewards over-promising, single approved-cloud bottlenecks, in-house capability hollowed out to contractors, and budget pressure that quietly descopes the guardrails — so in high-stakes public deployments the failure is, in effect, procured before it is coded. Tourism data adds the data-estate layer: fragmented, siloed and sometimes paywalled sources with inconsistent geographies and methods, complex data-sharing frameworks where every custodian has reasons for caution, and the legal-ethical exposure of mobile-movement data — all before a single useful tool can be built. Beyond the Black Box adds the comprehension-and-cost layer: fear narratives and uneven AI literacy as the main brake on genuine adoption ahead of any technical limit, explainability methods that are mathematically sound yet fail with real decision-makers, and below-cost hyperscaler pricing that distorts model choice toward oversized, costlier tools. AI and the changing nature of information work adds the records-and-legitimacy layer specific to government: uncertainty over whether AI chat transcripts are discoverable business records, legal review conducted in the shadow of past automated-decision failures, the absence of platforms approved for sensitive or classified material, and very low digital literacy among senior decision-makers — so the most consequential work stays away from the tools until records, FOI and classification guidance is clear, and adoption proceeds as a standing negotiation between what the technology affords and what accountability allows. People and Culture Leadership adds the workforce layer: project-by-project rollouts with no overarching people strategy, training pitched at tools rather than durable behaviours, shifting usage-based pricing and vendor-dependency risk, and the hollowing-out of entry-level roles that would otherwise supply the experienced judgement needed to supervise AI-augmented work. Authoritative by Source adds the public-sector adoption layer: government uptake is slowed less by the technology than by procurement, accreditation, financial delegations and the cost of pilots, while the opacity of vendored AI — third-party components nested inside one another, so data flows to unexpected places — makes the full software bill of materials hard to see; and much public data, though public in principle, stays practically inaccessible until someone does the unglamorous work of structuring it.

A cluster of late-2025 sessions extends the same map outward and upward. At national scale the constraint becomes physical and geopolitical — energy and grid capacity, supply-chain concentration in a few firms, and sovereign exposure to foreign suppliers (AI and the geopolitics of compute). In the economics session the binding gap was evidentiary: a shortage of Australian-specific data and a mismatch between headline forecasts and operational reality that makes confident planning hard (Exploring the economics of transformative AI). Local-innovation practice named the funding-and-culture layer directly — pilots that stall because short-term project money never covers operating costs, per-seat licence costs set against unclear benefit ownership, and an equity gap between confident early adopters and colleagues unsure how to begin (AI in practice). And the values work adds a subtler operational barrier: when a model's stance drifts with language and version, teams cannot reason about its outputs consistently without shared tools to track it (How AI Speaks Our Values). The pattern holds across all of them — the model is rarely the thing in the way.

Some barriers sit inside the collaboration itself, not the institution around it. A strategic-foresight session named the quiet dynamics that degrade group decisions before any technology is involved — unequal airtime, deference to hierarchy and seniority, a collective skew toward risk over opportunity, and discomfort with ambiguity — and treated surfacing them as a precondition for thinking well together (Beyond the Hype). The high-risk-measurement case adds the institutional-process layer specific to systems that judge people: capability histories that do not transfer between agencies, duplicated vetting regimes, drifting "business metadata", and rules of fair assessment that busy panels often do not fully grasp — so automating the measurement without first fixing that understanding only scales the confusion (Getting Hands-On With High-Risk AI). AI Productivity in the Agentic Era surfaces a cognitive version of the same barrier: even a well-defined, long-established role proved hard for participants to re-imagine once agents entered it — a difficulty that is itself an adoption blocker ahead of any technical one — alongside the tension between vendor incentives to drive adoption and the longer-term interests of communities and the public sector.

A missing evidence base is itself a collaboration barrier. Where there are no studies comparing the AI option with the existing process, groups defaulted to wanting to run the two in parallel as a deliberate pilot rather than choose blind (Beyond the Principles) — which only sharpens Growing with AI's point that, without benefit-sharing and common standards, the data such pilots would generate may never be pooled. The blocker is not that the answer is hard to compute; it is that no one has gathered the evidence, and the incentives to share it are weak.

The notes disagree instructively about data risk. Growing with AI argues data residency is largely a government preoccupation — what producers care about is who uses their data and who benefits. AI for the Systems You Inherited shows the government side of the same coin: security frameworks dictating which models may touch which data at all. These are not contradictory so much as two risk postures any cross-sector collaboration has to satisfy simultaneously — one side asking "who profits from my data?", the other "where is my data allowed to go?".

Capability gaps are a collaboration barrier in their own right. Practitioners who do not know what they do not know have no obvious place to turn (Growing with AI), and AI is an asset for people who can validate its output but a liability for those who cannot (AI for the Systems You Inherited). The no-code case makes that gap vivid: a non-engineer can "vibe-code" a working-looking prototype but cannot judge whether the generated code is sound enough to ship — which is exactly why the prototype-to-production step still needs an engineer, and why metered pricing and lock-in bite hardest on those least able to see them coming (Vibe Coding for Curious Humans). Beyond the Black Box names the affective side of the same gap: fear crowds out the bandwidth to learn what is publicly knowable, so the realistic bar is a "plumbing-level" literacy — enough to know the failure modes and who to call, not mastery. Across every note, AI is framed as a mitigating aid for these barriers — never the fix (Unlocking career potential).

Sources

Last compiled 2026-07-03 from 56 published note(s).

Notes contributing to this theme:

Shared with the AI CoLab Alliance — see join.aicolab.org.
This note was synthesised from the workshop recording with AI assistance and human review. It captures general lessons only, is not attributed to individual participants, and was reviewed for sensitive content and shared with the presenters before publication.