Enterprise Artificial Intelligence (AI) efficiency programme outcomes

Research Question

Which published AI strategies — corporate, national, or sector-specific — are explicitly designed around business efficiency as the primary objective, what measurable outcomes have they produced, and what design choices distinguish effective efficiency-focused AI programmes from those that underperform?

Findings

Executive Summary

Efficiency-focused AI programmes deliver measurable outcomes when three conditions align: high-quality operational data, end-to-end process integration (not isolated pilots), and active change management that changes how people work rather than just providing tools. ANZ Bank's $1.9B in productivity savings since 2019 and 40–55% faster software delivery exemplify what is achievable at scale with sustained investment. The same body of evidence shows that 74–95% of AI projects globally fail to deliver visible business value — primarily because organisations treat AI as a technology project rather than an operating model transformation. For NZ-scale organisations, the most actionable findings are that buy-first strategies enable fast entry at sub-$5k setup cost, and that the efficiency ceiling appears when routine task automation is exhausted and deeper gains require process redesign, not more AI tools.

Key Findings

1. Only 26% of organisations globally have scaled AI to generate visible business value. BCG's October 2024 survey found 74% of companies struggle to achieve and scale value from AI, despite most running pilots. McKinsey's 2024 survey corroborates: 80%+ of companies using AI see no significant earnings gains yet.

2. AI leaders outperform peers by a measurable margin. BCG data shows AI leaders achieve 1.5x higher revenue growth, 1.6x greater shareholder returns, and 1.4x higher returns on capital versus peers. Accenture found companies with AI-led processes achieve 2.5x higher revenue growth and 2.4x greater productivity.

3. The ANZ Bank case study is the most richly documented efficiency programme in the ANZ/NZ region. ANZ achieved $1.9B in productivity savings since 2019; 40–55% faster code development via GitHub Copilot across 3,000+ engineers; home loan origination deployment time reduced from more than one year to six weeks; risk model Gini coefficient improved from 0.78 to 0.82 processing 200,000 accounts in 30 minutes. ANZ's ROE in institutional banking doubled between 2016 and 2024.

4. Data readiness is the primary technical blocker. 85% of failed AI projects cite data as a core issue (multiple enterprise surveys). 61% of companies report their data is not yet ready for generative AI. Organisations with higher data maturity see significantly more financial impact from AI.

5. People and process change matters more than algorithms. BCG's 10-20-70 rule states that 10% of AI transformation is algorithms, 20% is data and technology, and 70% is people and process change. Projects fail most often because the operating model is not changed, not because the AI model is wrong.

6. Failure to scale beyond pilots is the dominant failure mode. Only 12–16% of AI initiatives reach enterprise scale. Most stall when moving from proof-of-concept to integration with core workflows. Common causes: brittle models, absent governance, no executive sponsorship, and insufficient change management.

7. Generative AI at the frontier shows diminishing returns on further model scaling. Doubling training compute now yields roughly 1% quality improvement for the largest models. This "efficiency ceiling" applies to AI developers, not deployers — for businesses using AI-as-a-tool, the ceiling is reached when routine task automation is exhausted and further gains require process redesign.

8. NZ businesses report meaningful efficiency gains at low entry cost. AI Forum NZ's third AI Productivity Report found 91% of NZ businesses report efficiency improvements from AI, 77% report operational cost reductions, and 25%+ see annual benefits exceeding $50,000. Setup costs are now under $5,000 for 75% of organisations, driven by off-the-shelf tools.

9. 68% of NZ SMEs report no plans to assess or invest in AI, significantly higher than comparable economies. The NZ AI Strategy projects $76B GDP uplift by 2038 with efficiency as the dominant near-term mechanism, but this assumption relies on a cohort that currently has limited AI engagement.

10. Healthcare AI shows the highest density of quantified outcomes. 71% of US hospitals used predictive AI by 2024. Documented outcomes include up to 35% reduction in adverse clinical events, 40% reduction in appointment wait times, and significant reductions in administrative staff hours for scheduling and billing.

11. Microsoft/IDC report 3.7x average ROI on generative AI, with leading companies achieving up to $10.3 returned per $1 invested. This is vendor-sponsored research; treat as directional, not independently audited.

12. Build vs. buy decisions follow a consistent pattern among high performers. Start with buy (off-the-shelf) for speed and proof of value; shift to build for customised solutions at scale where differentiation matters. Organisations that start with build often waste 12–18 months before generating business value.

Assumptions

• Assumption: Global case study findings (ANZ, BCG survey, McKinsey) are directionally applicable to NZ-scale organisations. Justification: NZ organisations operate in similar market conditions and use the same technology platforms. Scale differences matter at the margins (less proprietary data, smaller engineering teams) but the success and failure patterns are structural, not size-dependent.
• Assumption: Self-reported outcomes from corporate case studies are directionally accurate even if not independently audited. Justification: Multiple independent sources (McKinsey, BCG, Accenture, AI Forum NZ) converge on similar efficiency ranges. The pattern is consistent enough to treat as real even if specific figures are inflated.
• Assumption: The ANZ Bank case study is a near-peer for NZ private sector organisations at scale. Justification: ANZ operates extensively in NZ and its institutional platform transformation is frequently cited in NZ business and technology reporting.

Analysis

The central finding is a bimodal distribution in AI efficiency outcomes: a minority of organisations (roughly 20–26%) scale AI to generate visible business value, while the majority run pilots that do not progress. The gap is not technical — it is structural. Successful programmes share four characteristics: a clearly scoped business problem with quantifiable outcomes; operational data that is clean, governed, and accessible; a deployment model that embeds AI in existing workflows rather than running alongside them; and change management that modifies how people work.

The ANZ case is illustrative because it spans multiple deployment types — software delivery (GitHub Copilot), risk modelling (deep learning with Nvidia), process automation (Temporal workflows), and administrative AI (Copilot for M365). Each deployment addressed a specific efficiency gap with a measurable KPI. The $1.9B productivity savings figure is a cumulative multi-year outcome, not a single tool's result.

The BCG 10-20-70 rule resolves an important confusion: organisations frequently over-invest in the 10% (model selection, vendor evaluation) and under-invest in the 70% (retraining staff, redesigning processes, changing incentives). This is why technically capable organisations still fail to realise efficiency gains.

For NZ-scale organisations, the efficiency ceiling comes earlier. NZ's SME-dominated economy means most organisations have limited historical operational data, smaller engineering teams, and less capacity for bespoke model development. The evidence suggests these organisations should default to off-the-shelf solutions targeting well-documented efficiency use cases: customer service automation, administrative document processing, scheduling and resource allocation, and code assistance. The AI Forum NZ data showing $50k+ annual benefits at sub-$5k setup cost is plausible for these use cases — they are the most commoditised segment of AI deployment.

The $76B GDP uplift projection in NZ's AI Strategy is optimistic given that 68% of NZ SMEs currently have no plans to engage with AI. Even if all other conditions held, the SME engagement gap represents a structural risk to the projection.

Risks, Gaps, and Uncertainties

• Most quantified efficiency outcomes are self-reported by either the organisation or a vendor with commercial interest. Independently audited case studies are rare.
• Air NZ and Fonterra were listed as sources in the research item but public quantified outcome data for both is thin. Efficiency gains are cited in general terms only; company-specific hard numbers are not publicly available.
• The OECD case study database and NZ Productivity Commission sources were not consulted for this iteration. They may hold independently verified data not captured here.
• NZ SME AI adoption data is drawn from AI Forum NZ reporting; the primary dataset methodology is not described in the summaries accessed.
• The efficiency ceiling concept applies differently to AI deployers (businesses using AI tools) vs. AI developers (model trainers). The research conflates these two ceilings in parts of the literature; they require separate analysis.

Open Questions

• What are the independently audited (not self-reported) efficiency outcomes for NZ's largest AI deployments? (Could become a separate research item.)
• What specific efficiency use cases are most tractable for NZ SMEs given their data and capability constraints? (Practical guide scope.)
• Does the Jevons paradox apply to efficiency-mode AI — does automating tasks create enough new demand to neutralise net efficiency gains? (Macroeconomics research item.)
• What is the change management methodology used by ANZ and other high-performing organisations? Is it codified? (Process research item.)

sources

• [ ] McKinsey Global Institute: AI productivity reports (2023, 2024)
• [ ] BCG "AI at Scale" case study library
• [ ] Gartner AI hype cycle efficiency use case data
• [ ] Harvard Business Review case studies on AI-driven operational transformation
• [ ] OECD case study database: AI in public sector efficiency
• [ ] NZ Productivity Commission reports referencing technology-driven productivity gains
• [ ] Specific sectors: ANZ banking AI programme; Air NZ operational AI; Fonterra precision agriculture AI