Industry perspectives

No consensus yet: what industry leaders are actually saying about AI economics

Major consulting firms, investors, academics, analysts, and technology leaders hold genuinely different views on whether AI's economic promise will materialise — and how quickly. Here is what each camp is saying, and why the debate matters for how you govern AI investment now.

This is not unusual. The same pattern played out with cloud economics, ERP, and previous automation waves. Disagreement is a feature of early cycles — but organisations that navigate it well build governance before the consensus forms.

The range of views

From trillion-dollar opportunity to trillion-dollar misallocation

The spread of authoritative opinion on AI economics is unusually wide — and deliberately so in some cases. Below is an honest mapping of where the main camps sit and what evidence they are drawing on.

Broadly Optimist

McKinsey Global Institute

Accenture

Microsoft

a16z (Andreessen Horowitz)

Nuanced / Conditional

Gartner

BCG Henderson Institute

IBM Institute for Business Value

FinOps Foundation

Broadly Skeptical

Goldman Sachs Research

Prof. Daron Acemoglu, MIT

Gary Marcus

Sequoia Capital

Stance classifications reflect each organisation's dominant published position. Most hold more nuanced internal views. Sources listed under each entry below.

Historical parallel

This debate has happened before — with cloud and FinOps

The pattern of rapid adoption, unexpected cost, executive scepticism, and eventual governance discipline is not new. Cloud computing followed an almost identical arc. Understanding that arc is useful for predicting what happens next with AI economics.

Phase 1: Adoption without governance

Cloud / FinOps

2010–2014: Cloud adopted rapidly by engineering teams. Costs invisible to finance. Bills arrive; no one owns them.

AI Economics (now)

2022–2024: AI tools adopted rapidly by teams. API costs buried in expense reports. Shadow AI accumulates outside IT budgets.

Phase 2: The cost shock

Cloud / FinOps

2014–2017: Cloud bills arrive unexpectedly. CFOs see line items they cannot explain. Engineering and finance at odds.

AI Economics (now)

2024–2025: AI API bills arrive unexpectedly. Enterprise contracts hit hard limits. Teams burning through annual budgets in weeks.

Phase 3: The governance response

Cloud / FinOps

2017–2019: Cloud governance frameworks emerge. Tagging, allocation, commitment optimisation, chargeback models.

AI Economics (now)

2025–?: AI economics frameworks beginning to emerge. Cost attribution, unit economics, governance investment, value proof.

Phase 4: Discipline as competitive advantage

Cloud / FinOps

2019–present: FinOps Foundation formalises the discipline. Cloud cost management becomes a core IT competency.

AI Economics (now)

??: The organisations that build AI economics discipline now will operate at lower cost and with clearer ROI as the market matures.

The FinOps Foundation parallel

The FinOps Foundation — which formalised cloud cost management as a discipline starting in 2019 — has explicitly extended its framework to AI and ML workloads. Foundation leadership has drawn a direct parallel between uncontrolled cloud spend in 2014–2018 and uncontrolled AI spend in 2023–present. The same patterns appear: shadow spend in expense reports, no cost attribution, no unit economics, no governance accountability.

The core insight: organisations did not need to wait for cloud consensus to build cloud governance. The ones that did not wait ended up significantly ahead on cost and agility when the market matured. The same logic applies to AI.

The cost reality

AI spend is getting out of control at enterprise scale

Beneath the debate about long-term economic potential, there is an immediate and practical problem: organisations are discovering that AI costs are far less predictable than anticipated, and that existing financial management tools are not equipped to govern them.

The budget burn

Multiple enterprises have reported AI API budgets — sized for a full year of anticipated use — being consumed within weeks. Agentic workflows in particular can generate token volumes orders of magnitude higher than human-initiated requests, compressing spend timelines dramatically.

Shadow AI accumulation

Industry surveys consistently show that a significant share of enterprise AI spend flows through individual expense reports rather than IT procurement — ChatGPT, Claude, Midjourney subscriptions, API keys on personal cards. This spend is invisible to FinOps, ITFM, and TBM processes.

The governance gap

When AI bills arrive, most organisations discover they lack the attribution, tagging, and unit-cost frameworks needed to understand what drove the spend, which teams are responsible, and whether the output justified the cost. The same absence that creates cloud bill surprises is now creating AI bill surprises.

Documented pattern — enterprise AI spend controls

Across 2024 and into 2025, a pattern has emerged at enterprise scale: AI deployment outpaces financial governance. API rate limits get hit unexpectedly. Agentic pipelines generate costs that exceed estimates by an order of magnitude. Finance teams — alerted by variance reports — implement emergency spend caps and mandatory approval gates. Engineering teams perceive this as a brake on progress; finance teams see it as basic cost control.

This dynamic is not unique to any single organisation or vendor relationship. It reflects a structural gap between how AI services are priced (usage-based, variable, often opaque) and how enterprises budget (annual, fixed, allocated by department). Closing that gap is the operational definition of AI economics governance.

The labour debate

AI removes workers — or AI is more expensive than the workers it replaced

One of the most contested dimensions of AI economics is whether the labour savings from AI deployment exceed the cost of the AI itself. Early evidence is genuinely mixed — and some organisations are discovering that the arithmetic does not work in the way they anticipated.

The displacement case

AI systematically reduces headcount and labour cost

IBM (2023): CEO Arvind Krishna announced a pause on hiring approximately 7,800 roles expected to be replaced by AI — primarily in HR and back-office functions. This was presented as a measured, deliberate response to AI capability improving faster than the underlying work.

Goldman Sachs research (2023): Estimated that 300 million full-time jobs globally are exposed to automation from generative AI, with 18% of work globally capable of being automated. The legal, administrative, and knowledge-work categories show the highest exposure.

Klarna (2024): CEO Sebastian Siemiatkowski stated that AI had effectively replaced the work of 700 full-time customer service agents. He later acknowledged customer satisfaction metrics required monitoring and that the company had paused further cuts pending quality assessment.

The cost paradox

AI is sometimes more expensive than the labour it replaced

The verification cost: When AI handles tasks previously done by humans, the output still requires verification for high-stakes applications. The labour saving is real, but the verification cost partially offsets it — and is rarely captured in initial ROI models.

The infrastructure cost: Full-cost accounting of AI deployment — including infrastructure, governance, integration, and support — frequently exceeds the all-in cost of the labour it displaced. This is particularly true for smaller-scale deployments where platform cost is not spread across sufficient volume.

IBM's actual result: IBM later reported that the roles affected by its hiring pause were accompanied by new roles in AI governance, prompt engineering, and system integration. The gross headcount reduction was smaller than the headline suggested, and implementation costs were higher.

The emerging view

The organisations most likely to realise genuine labour economics from AI are those that redesign the work, not just add AI to existing roles.

BCG, McKinsey, and others researching AI productivity patterns consistently find a bifurcation: organisations that redesign processes around AI capability capture 40%+ productivity gains; those that layer AI onto existing processes capture 10–20% at best, and sometimes less when verification and governance costs are included. The determinant is not the AI model — it is the management investment in workflow redesign.

Documented positions

What each organisation is actually saying

These summaries reflect each organisation's published positions as of mid-2025. Sources are listed beneath each entry. Stances are editorially characterised based on the weight of published evidence — most organisations hold more nuanced internal positions.

Broadly optimist

OptimistConsulting

McKinsey Global Institute

“$2.6–4.4 trillion in annual value from generative AI”

MGI's 2023 analysis of generative AI's economic potential placed the aggregate annual value across 63 use cases at $2.6–4.4 trillion. It identified knowledge-worker productivity as the primary driver and positioned AI as a step-change comparable to major industrial technologies. Subsequent MGI research noted that capture of that value requires substantial workflow redesign — not just model deployment.

Source: McKinsey Global Institute, 'Economic Potential of Generative AI', 2023

OptimistConsulting

Accenture

“AI is a full operating model transformation, not a technology layer”

Accenture's research consistently frames AI as an enterprise reinvention opportunity, estimating that organisations applying AI at scale across operating models can achieve 20–30% productivity improvement in targeted functions. Their 'AI: Built to Scale' research series documents that scale requires robust governance architecture — and that most organisations are significantly under-invested in it.

Source: Accenture Technology Vision, 2023–2024

OptimistTechnology

Microsoft

“Copilot cuts time on specific tasks by 10–40%”

Microsoft publishes quarterly Copilot productivity data and has cited time savings in specific tasks — drafting, summarisation, meeting notes — of 10–40%. CEO Satya Nadella frames AI as the most important technology platform shift since the web. Microsoft has committed over $50 billion in AI-related capital expenditure, reflecting high conviction in its own ROI calculations.

Source: Microsoft Work Trend Index, 2024; Microsoft earnings calls

OptimistInvestor / VC

a16z (Andreessen Horowitz)

“The cost-per-task curve changes the unit economics of knowledge work”

a16z argues that the rapidly declining cost-per-task of AI models fundamentally changes the economics of knowledge work — not by replacing workers but by enabling dramatically higher output per person. Partners have consistently argued that the productivity multiplier justifies current infrastructure spend, and that the relevant comparison is not current cost versus current output but current cost versus a 5-year cost curve.

Source: a16z: 'Who's Building the AI Stack', 2023–2024

Broadly skeptical

SkepticFinance / Research

Goldman Sachs Research

“Is AI 'too expensive' to deliver what it promises?”

Jim Covello, Goldman Sachs head of global equity research, published a widely cited 2024 analysis questioning whether AI ROI could justify the capital investment. His argument: AI is best at doing cheap tasks better, but the economic return from cheap tasks is inherently limited. The expensive tasks that would generate significant economic value require human judgement AI cannot yet reliably provide. He estimated the industry needs to generate $600B+ in revenue to justify current compute investment — a gap that remains.

Source: Goldman Sachs, 'Gen AI: Too Much Spend, Too Little Benefit?', 2024

SkepticAcademic Research

Prof. Daron Acemoglu, MIT

“AI will automate less than 5% of tasks at human competency in 10 years”

In 'The Simple Macroeconomics of AI' (2024), Acemoglu argued that AI will affect fewer economically significant tasks than optimists assume — he estimated less than 5% of tasks at human competency level within a decade. He challenged productivity projections as 'wishful thinking', noting that most high-value tasks require contextual judgment, organisational trust, and accountability structures AI systems currently lack. His macro models showed far more modest GDP effects than tech-optimist projections.

Source: Acemoglu, 'The Simple Macroeconomics of AI', NBER Working Paper, 2024

SkepticAcademic / Researcher

Gary Marcus

“Hallucination and reliability costs are systematically underweighted”

AI researcher and author Gary Marcus has consistently documented the gap between AI marketing claims and deployment realities. His central argument: the true cost of AI includes the labour required to verify, correct, and govern outputs — costs that rarely appear in vendor ROI models. For high-stakes applications, these verification costs can exceed the productivity gains, making the economics negative in aggregate even where individual use cases appear positive.

Source: Gary Marcus, 'The Road to AI We Can Trust', ongoing; multiple published papers

SkepticInvestor / VC

Sequoia Capital

“AI's $600 billion revenue problem — does the arithmetic close?”

David Cahn at Sequoia published 'AI's $600B Question' in 2024, updating an earlier analysis of the gap between Nvidia GPU revenue and the AI application revenue needed to justify the compute ecosystem. The piece documented a structural mismatch: the infrastructure layer is valued for a future revenue profile that current applications cannot yet demonstrate. Sequoia remains invested in AI but raised this as a key risk for the cycle.

Source: Sequoia Capital, 'AI's $600B Question', David Cahn, 2024

Nuanced / conditional

NuancedAnalyst

Gartner

“Generative AI in the trough of disillusionment — governance is the missing layer”

Gartner's 2024 Hype Cycle for Artificial Intelligence placed generative AI at 'Peak of Inflated Expectations' moving toward the 'Trough of Disillusionment'. Their research highlights that most organisations adopting AI are underinvesting in governance, cost management, and measurement infrastructure. Gartner analysts consistently note that AI productivity claims are real but highly deployment-specific — and that enterprise-level proof remains thin.

Source: Gartner Hype Cycle for Artificial Intelligence, 2024

NuancedConsulting / Research

BCG Henderson Institute

“AI productivity requires significant workflow redesign — deployment alone does not deliver”

BCG's research into AI adoption patterns shows a consistent finding: the organisations capturing measurable productivity gains are those that redesign the work, not just add AI to existing processes. Their benchmarking distinguishes between 'AI-augmented productivity' (modest, 10–20%) and 'AI-redesigned productivity' (substantial, 40%+) — and finds that reaching the second level requires significant management investment, change capability, and governance discipline.

Source: BCG, 'Measuring AI Productivity at Scale', 2024

NuancedTechnology / Research

IBM Institute for Business Value

“Paused hiring for 7,800 roles — then found AI created new ones”

In 2023, IBM CEO Arvind Krishna announced a pause on hiring for roughly 7,800 roles expected to be replaced by AI, particularly in HR and back-office functions. IBM subsequently found the picture more complex: AI implementation required new roles for prompt engineering, AI governance, output verification, and system integration. The net employment effect was more modest than the initial headline suggested — and the cost of the transition was higher.

Source: Arvind Krishna, IBM CEO commentary, 2023; IBM Institute for Business Value research

NuancedIndustry Foundation

FinOps Foundation

“AI spend is following the same arc as cloud — and organisations are just as unprepared”

The FinOps Foundation, which developed cloud cost management as a discipline, has explicitly extended its framework to AI/ML workloads. Foundation leadership has drawn a direct parallel between uncontrolled cloud spend (2014–2018) and uncontrolled AI spend (2023–present). The same patterns appear: shadow spend in expense reports, no cost attribution, no unit economics, no governance. Their guidance is identical in structure to early cloud FinOps — start with visibility, then accountability, then optimisation.

Source: FinOps Foundation, 'AI / ML Cost Management' framework documentation, 2024

What this means for practitioners

The debate will not resolve itself — build the governance now

The absence of consensus is not a reason to wait. The organisations that built cloud governance before the consensus formed are now significantly ahead. The same logic applies to AI economics.

Do not wait for the debate to resolve

Academic and analyst consensus on AI ROI may take a decade. The cost governance problem exists today. Address the governance architecture regardless of which camp turns out to be right.

Apply the FinOps lesson

Cloud economics became manageable when organisations built visibility, attribution, and accountability — in that order. The same sequence applies to AI. Start with cost visibility before optimisation.

Separate platform from use-case economics

Foundational AI platform investment has different economics from individual use-case deployment. Both need to be evaluated, but on different timescales and with different proof standards.

Include verification and governance costs

ROI models that exclude verification labour, governance overhead, and integration cost will systematically overestimate return. Include the full cost stack before committing to a business case.

Use the lack of consensus as a negotiating point

When vendors and consultants present AI ROI projections as settled fact, the diversity of authoritative opinion is useful context. Insist on scenario analysis rather than point estimates.

Build the internal language

The most durable competitive advantage from AI governance is shared language across finance, technology, strategy, and operating teams. Invest in that before the next budget cycle.

Editorial note on sources

The positions summarised on this page are drawn from published reports, earnings calls, academic papers, and public commentary as of mid-2025. Research papers are cited by author and publication. Commercial research (McKinsey, Goldman Sachs, BCG, Gartner, Accenture) is available through each organisation's website and may require registration or subscription. Academic papers (Acemoglu et al.) are available through NBER and SSRN. This page is updated as significant new positions emerge — it is intended as a reading map, not a comprehensive literature review.