Why can AI do so much while business change takes so long?

Scope

Australian SMEs; evidence from dated Australian and international studies

Confidence

High on the access–implementation distinction; lower on which barrier dominates

My starting point

Access to a model does not establish that a business can use it well. My infrastructure submissions ask for practical training, help to build workflows and the ability to judge outputs. Nominal compute allocations are a weak measure if firms cannot turn them into useful work. Tasmania submission, 01/10/2026.

This is also the question connecting the series to my research: how Australian SME leadership teams interpret AI and relate it to resource commitments, routines and strategic responses. Buying a tool and changing how a firm works are different things. The research is inductive; its questions should not be mistaken for established findings.

What the evidence supports

The National AI Centre’s December 2025–February 2026 wave reports that 43% of Australian SMEs had some level of AI adoption. Its monthly survey has at least 400 respondents and is weighted by industry, state and size. This is a dated self-report measure, not the October adoption rate, the share with deep integration or measured return on investment. Reported trust, relevance and skills barriers identify things respondents say matter; they do not prove why a firm does or does not adopt.

OECD’s 2025 SME report finds limited preparation among users, including training and guidelines. Its seven-country sample excludes Australia. The useful lesson is a reason to examine implementation support, rather than a statistic to relabel as Australian practice.

What should make us cautious

A gap between a demonstration and everyday use is not automatically evidence of irrational resistance. A task may require customer trust, access to protected information, integration with an old system or a level of reliability that the demonstration did not test. A smaller firm must also decide whether the benefit warrants its time.

Anthropic’s expectations survey illustrates disagreement about the future: staff expected AI to do an average 75% of their current work hours by 2030, compared with 28% among employed US respondents. The public sample is weighted and staff responses are unweighted. These are beliefs, not proof that insiders are correct or businesses are behind.

What I would look for

Support should help a firm specify a useful job, choose an appropriate tool, protect information, assign responsibility and assess the result. The question is whether people can repeat useful work and handle failure, not whether they attended an introductory session. That position is already present in the submitted pieces.

A practical evaluation would include rework, output quality and staff experience alongside saved time. If a trial fails, changing or stopping it should remain possible. These are evaluation criteria, not a claim that a particular programme has already succeeded.

18/10/2026: adoption alongside dependence

The possibility of a compounding frontier lead strengthens the case for examining dependencies alongside adoption. It does not remove the value of useful work with today's models. For a small business, a practical trial can assess output quality, review effort and cost while also testing whether its information and workflows can move to another provider. These are proposed evaluation criteria, not a measured finding from my research. The companion frontier article explains why cheaper access can support local gains without providing ownership or assured continuity.

What would change this answer

Comparative evidence that reliable low-cost tools routinely produce sustained gains without additional support would reduce the weight placed on implementation assistance. Evidence that current assistance delivers little would change the proposed form of support. Longitudinal Australian studies that distinguish model capability, organisational cost and changes in routines would help identify which constraints actually bind.

Limits of this edition

The series does not claim that all SMEs should adopt the same tools, or that non-adoption is always a problem. There is no validated diagnostic or firm-level causal finding from my ongoing PhD being offered here.