AI capability per dollar is improving rapidly, but enterprise AI bills are rising because usage, context, reasoning depth, agent loops, tool calls and infrastructure scale faster than unit prices fall. The right economic unit is therefore not the token, licence or model benchmark. It is the cost and value of an accepted business outcome. Enterprises that treat AI as a variable-cost operating system—measured by workload, quality and decision value—will make better build/buy, model-routing and scaling decisions than those managing only model price.
Research question and scope
Core question: What does enterprise AI actually cost at scale, how should value be measured, and when do declining model prices translate—or fail to translate—into better business economics?
Scope: Model/API costs, agentic workflows, infrastructure, human review, software substitution, workload routing, FinOps and value measurement. Capital-market infrastructure economics are covered only where they affect enterprise unit costs.
Method: Synthesis of Stanford cost/performance data, enterprise surveys, FinOps benchmarks, current agentic-economics research and hyperscaler/infrastructure evidence. The report explicitly separates price-per-token from fully loaded workload economics.
1. The paradox: intelligence gets cheaper while AI spending rises
AI has one of the fastest price-performance improvement curves in modern computing. Stanford’s AI Index documented a more than 280-fold decline between November 2022 and October 2024 in the inference price required to reach GPT-3.5-level performance on MMLU—from about $20 to $0.07 per million tokens—while hardware price-performance improved roughly 30% annually. S1 The 2026 AI Index shows continued convergence of top-model capability and rapidly expanding compute capacity, shifting competition increasingly toward cost, reliability and domain performance. S2
Yet enterprise spending is rising. McKinsey reports that enterprise LLM spending tripled in the twelve months to the end of 2025, 93% of qualified respondents in an Enterprise AI FinOps survey exceeded AI budgets, and one fifth of respondents to its 2026 State of AI survey say operating costs have constrained AI use. S3 FinOps Foundation’s 2026 survey finds 98% of respondents now manage AI spend, up from 31% two years earlier, and identifies AI cost management as the leading skill gap. S4
| The central economic insight Falling unit prices do not guarantee falling total cost when the quantity and complexity of consumption expand faster. Agentic AI intensifies this effect because one user request can trigger many model calls, retrievals, tool invocations, retries and validations. |
|---|
2. Why token price is not the economic unit
Per-token pricing remains useful for procurement and architecture comparison, but it is a weak management metric. McKinsey’s July 2026 research argues that enterprises increasingly pay for systems whose costs scale with usage rather than seats; the same nominal “task” can consume dramatically different amounts of context, reasoning and tool execution. S5
| Layer | Cost drivers | Why it varies |
|---|---|---|
| Model inference | Input/output tokens, reasoning/test-time compute, modality | Task complexity and model choice |
| Context/retrieval | Embedding, search, document loading, context size | Enterprise knowledge scope and retrieval design |
| Agent loop | Planning steps, retries, self-checks, subagents | Reliability threshold and workflow length |
| Tools/APIs | Search, databases, SaaS/API calls, code execution | External service pricing and call volume |
| Infrastructure | Cloud GPU/CPU, networking, storage, private compute | Latency, sovereignty and utilisation |
| Human control | Review, approval, exception handling, rework | Risk level and error rate |
| Operations | Evaluation, observability, security, governance | Production maturity and compliance |
3. The correct metric: cost per accepted outcome
8i recommends measuring the economics at the smallest business unit that has value and can be quality-controlled: an accepted claim, resolved case, approved document, merged code change, completed order, reconciled account, or decision made. This prevents two common errors: celebrating low token prices while the agent loops expensively, and claiming “hours saved” where output requires extensive review.
- Define the outcome and baseline: what does the process cost and achieve today?
- Measure direct AI cost: models, retrieval, tools and infrastructure.
- Add control cost: human review, escalation, audit and rework.
- Adjust for quality: accepted output / total attempted output.
- Measure realised capacity or financial effect, not theoretical time saved.
- Include failure cost and downside risk for consequential workflows.
- Compare alternatives: human-only, software automation, packaged AI, custom AI and no-action.
4. Value is broader than labour substitution
The 2026 evidence increasingly challenges a labour-savings-only ROI model. McKinsey’s “decision dividend” framing argues that major gains can come from faster decisions, better use of existing assets and opportunities that would otherwise be missed. S6 This matters because many high-value AI applications—pricing, sales prioritisation, maintenance, fraud detection, product development—change revenue, quality or asset utilisation rather than simply remove head count.
| Value mechanism | Example measure | Common measurement trap |
|---|---|---|
| Labour efficiency | Minutes/case, cases/FTE, review hours | Assuming time saved becomes cash saved |
| Throughput | Orders, releases, cases or analyses per period | Ignoring demand constraints downstream |
| Quality | Defect rate, first-pass yield, resolution quality | Measuring output quantity instead of acceptance |
| Decision speed | Time-to-decision / opportunity capture | Not valuing avoided delay |
| Revenue/growth | Conversion, win rate, price realisation, retention | Attributing all change to AI |
| Risk avoidance | Loss prevented, control failures avoided | Ignoring low-frequency high-impact downside |
| Strategic option | Experiment cost, time to build capability | Treating optionality as guaranteed return |
Sources and notes
- Stanford HAI, AI Index 2025 — inference cost trend. [Source](https://hai.stanford.edu/ai-index/2025-ai-index-report/research-and-development)) — Historic price-performance baseline; >280× decline at GPT-3.5-level benchmark through Oct 2024.
- Stanford HAI, 2026 AI Index — Technical Performance. [Source](https://hai.stanford.edu/ai-index/2026-ai-index-report/technical-performance)) — Top-model convergence and performance context.
- McKinsey, Is that AI agent worth it? Agentic economics (13 Jul 2026). [Source](https://www.mckinsey.com/capabilities/quantumblack/our-insights/is-that-ai-agent-worth-it-agentic-economics-and-the-modern-operating-model)) — Enterprise cost escalation and agentic economics.
- FinOps Foundation, State of FinOps 2026. [Source](https://data.finops.org/)) — 1,192 respondents representing $83bn+ annual cloud spend.
- McKinsey, Cost versus value: Managing agentic AI system performance (8 Jul 2026). [Source](https://www.mckinsey.com/capabilities/quantumblack/our-insights/cost-versus-value-managing-agentic-ai-system-performance)) — Why per-token pricing is insufficient for enterprise economics.
- McKinsey, The decision dividend: How AI creates economic value (26 Aug 2026). [Source](https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/the-new-economics-of-ai)) — Value mechanisms beyond labour savings.
8i TECHNOLOGIES
The Economics of Enterprise AI
From cheap tokens to expensive systems: unit economics, value capture, cost governance and build-versus-buy in 2026
Reference date: 15 September 2026
See what matters next.
| Executive answer AI capability per dollar is improving rapidly, but enterprise AI bills are rising because usage, context, reasoning depth, agent loops, tool calls and infrastructure scale faster than unit prices fall. The right economic unit is therefore not the token, licence or model benchmark. It is the cost and value of an accepted business outcome. Enterprises that treat AI as a variable-cost operating system—measured by workload, quality and decision value—will make better build/buy, model-routing and scaling decisions than those managing only model price. |
|---|
Research question and scope
Core question: What does enterprise AI actually cost at scale, how should value be measured, and when do declining model prices translate—or fail to translate—into better business economics?
Scope: Model/API costs, agentic workflows, infrastructure, human review, software substitution, workload routing, FinOps and value measurement. Capital-market infrastructure economics are covered only where they affect enterprise unit costs.
Method: Synthesis of Stanford cost/performance data, enterprise surveys, FinOps benchmarks, current agentic-economics research and hyperscaler/infrastructure evidence. The report explicitly separates price-per-token from fully loaded workload economics.
1. The paradox: intelligence gets cheaper while AI spending rises
AI has one of the fastest price-performance improvement curves in modern computing. Stanford’s AI Index documented a more than 280-fold decline between November 2022 and October 2024 in the inference price required to reach GPT-3.5-level performance on MMLU—from about $20 to $0.07 per million tokens—while hardware price-performance improved roughly 30% annually. [S1] The 2026 AI Index shows continued convergence of top-model capability and rapidly expanding compute capacity, shifting competition increasingly toward cost, reliability and domain performance. [S2]
Yet enterprise spending is rising. McKinsey reports that enterprise LLM spending tripled in the twelve months to the end of 2025, 93% of qualified respondents in an Enterprise AI FinOps survey exceeded AI budgets, and one fifth of respondents to its 2026 State of AI survey say operating costs have constrained AI use. [S3] FinOps Foundation’s 2026 survey finds 98% of respondents now manage AI spend, up from 31% two years earlier, and identifies AI cost management as the leading skill gap. [S4]
| The central economic insight Falling unit prices do not guarantee falling total cost when the quantity and complexity of consumption expand faster. Agentic AI intensifies this effect because one user request can trigger many model calls, retrievals, tool invocations, retries and validations. |
|---|
2. Why token price is not the economic unit
Per-token pricing remains useful for procurement and architecture comparison, but it is a weak management metric. McKinsey’s July 2026 research argues that enterprises increasingly pay for systems whose costs scale with usage rather than seats; the same nominal “task” can consume dramatically different amounts of context, reasoning and tool execution. [S5]
| Layer | Cost drivers | Why it varies |
|---|---|---|
| Model inference | Input/output tokens, reasoning/test-time compute, modality | Task complexity and model choice |
| Context/retrieval | Embedding, search, document loading, context size | Enterprise knowledge scope and retrieval design |
| Agent loop | Planning steps, retries, self-checks, subagents | Reliability threshold and workflow length |
| Tools/APIs | Search, databases, SaaS/API calls, code execution | External service pricing and call volume |
| Infrastructure | Cloud GPU/CPU, networking, storage, private compute | Latency, sovereignty and utilisation |
| Human control | Review, approval, exception handling, rework | Risk level and error rate |
| Operations | Evaluation, observability, security, governance | Production maturity and compliance |
3. The correct metric: cost per accepted outcome
8i recommends measuring the economics at the smallest business unit that has value and can be quality-controlled: an accepted claim, resolved case, approved document, merged code change, completed order, reconciled account, or decision made. This prevents two common errors: celebrating low token prices while the agent loops expensively, and claiming “hours saved” where output requires extensive review.
- Define the outcome and baseline: what does the process cost and achieve today?
- Measure direct AI cost: models, retrieval, tools and infrastructure.
- Add control cost: human review, escalation, audit and rework.
- Adjust for quality: accepted output / total attempted output.
- Measure realised capacity or financial effect, not theoretical time saved.
- Include failure cost and downside risk for consequential workflows.
- Compare alternatives: human-only, software automation, packaged AI, custom AI and no-action.
4. Value is broader than labour substitution
The 2026 evidence increasingly challenges a labour-savings-only ROI model. McKinsey’s “decision dividend” framing argues that major gains can come from faster decisions, better use of existing assets and opportunities that would otherwise be missed. [S6] This matters because many high-value AI applications—pricing, sales prioritisation, maintenance, fraud detection, product development—change revenue, quality or asset utilisation rather than simply remove head count.
| Value mechanism | Example measure | Common measurement trap |
|---|---|---|
| Labour efficiency | Minutes/case, cases/FTE, review hours | Assuming time saved becomes cash saved |
| Throughput | Orders, releases, cases or analyses per period | Ignoring demand constraints downstream |
| Quality | Defect rate, first-pass yield, resolution quality | Measuring output quantity instead of acceptance |
| Decision speed | Time-to-decision / opportunity capture | Not valuing avoided delay |
| Revenue/growth | Conversion, win rate, price realisation, retention | Attributing all change to AI |
| Risk avoidance | Loss prevented, control failures avoided | Ignoring low-frequency high-impact downside |
| Strategic option | Experiment cost, time to build capability | Treating optionality as guaranteed return |
5. Model routing and architecture become financial controls
As leading-model capability converges and smaller models improve, using the most capable model for every step becomes economically irrational. Stanford reports top model performance clustering tightly by early 2026. [S2] McKinsey’s work on inference-cost reduction highlights continuing gains from hardware, model efficiency, quantisation and system architecture. [S7] The economic architecture therefore moves toward model routing: use the cheapest model/system that meets the quality and risk threshold for each step.
- Route low-risk extraction/classification to small or specialised models.
- Use frontier reasoning only where marginal accuracy has economic value.
- Cache stable outputs and retrieved context.
- Reduce unnecessary agent loops and duplicated validation.
- Batch or schedule latency-tolerant workloads.
- Consider private/on-prem capacity for stable, high-volume workloads only after utilisation and total-cost analysis.
6. Cloud/API versus owned/private compute
Deloitte’s 2026 infrastructure work argues that some enterprises are reaching a crossover where predictable, high-volume workloads justify private or on-prem infrastructure, while API/cloud remains attractive for variable demand, rapid model access and experimentation. [S8] This is not a universal threshold: cost depends on utilisation, hardware depreciation, energy, operations, model cadence and sovereignty requirements.
| Deployment | Economic strength | Economic weakness | Best fit |
|---|---|---|---|
| API / managed model | Low fixed cost, fast access to frontier models | Variable unit cost, provider margin, less control | Variable or evolving workloads |
| Cloud reserved/dedicated | Capacity commitments and integration | Commitment risk, still variable service layers | Large predictable workloads needing cloud ecosystem |
| Private/on-prem | High utilisation can lower marginal cost; control | Capital, talent, refresh and idle-capacity risk | Stable high-volume or sovereignty-sensitive workloads |
| Hybrid routing | Optimises per workload | Architecture/operations complexity | Mature portfolios with clear workload classes |
7. Agent economics amplify both value and cost
Agentic systems change the curve because they substitute sequences of software and human coordination, not one inference. McKinsey’s “Is that AI agent worth it?” notes that agent cost can spiral even while tokens become cheaper, because agents consume repeated inference and tool calls. [S3] This is why agent economics should be modelled as a workflow tree: probability of each branch × calls × cost × intervention × outcome value.
| 8i unit-economics formula Expected economic value per task = probability-adjusted value of accepted outcome − (model + context + tool/API + infrastructure + human review + exception/rework + allocated governance cost + expected failure loss). |
|---|
8. Build versus buy is changing
McKinsey’s August 2026 State of AI survey reports 32% of respondents saying their organisation had chosen not to buy at least one software product or feature because it could be built internally with agentic coding tools. [S9] This is an important early signal, not evidence of broad SaaS collapse. It suggests that the cost of creating “good-enough” internal software is falling for narrow workflows, which raises the bar for packaged software differentiation.
- Buy when the workflow is standard, compliance/integration burden is high, and the vendor’s data/network effects matter.
- Build when process logic is proprietary, the feature is narrow, internal integration is the value, or packaged pricing exceeds build/operate cost.
- Avoid confusing code-generation speed with lifetime product economics: maintenance, security, ownership and change management remain real.
9. FinOps for AI becomes value governance
The FinOps Foundation’s 2026 report shows AI moving into normal technology-value governance: 98% of surveyed practitioners manage AI spend, but visibility, allocation and ROI remain major challenges. [S4] The discipline should therefore extend from chargeback to demand shaping: approve high-cost reasoning where value justifies it, route routine tasks cheaply, set budgets/limits per workflow, and surface cost at decision time.
10. 8i economic maturity model
| Level | Management behaviour | Primary metric |
|---|---|---|
| 1 — Access | Licences/API enabled; spend tracked globally | Monthly AI spend |
| 2 — Visibility | Cost allocated to teams/apps | Cost by user/application |
| 3 — Unit economics | Cost linked to workload and quality | Cost per accepted task |
| 4 — Portfolio value | Workloads compared and routed by value/risk | ROI / value per dollar by workflow |
| 5 — Dynamic optimisation | Models, infra and workflows continuously optimised | Marginal value vs marginal cost |
11. 8i assessment
| 8i view The near-term economics of enterprise AI favour disciplined expansion rather than blanket cost cutting. Unit model prices will continue to fall, but total spend can rise sharply as AI becomes embedded in operations. Competitive advantage comes from turning cheaper intelligence into better workflows faster than cost and complexity accumulate. |
|---|
12. What leaders should do now
- Create workload-level AI P&Ls for material production systems.
- Adopt cost per accepted outcome and intervention rate as core metrics.
- Separate labour capacity released from actual financial benefit realised.
- Route models by quality threshold; reserve expensive reasoning for high-value uncertainty.
- Set agent budgets, loop limits and tool-call controls.
- Re-run build-versus-buy decisions for narrow software capabilities, but include lifetime ownership.
- Link AI FinOps with product/process owners so cost optimisation cannot degrade business value.
Source register
Sources are prioritised toward primary institutions, transparent methodology, observable behavioural data, large-sample surveys and current production evidence. Vendor evidence is identified as such and used cautiously where it provides proprietary usage data or live implementation detail.
[S1] Stanford HAI, AI Index 2025 — inference cost trend. Source — Historic price-performance baseline; >280× decline at GPT-3.5-level benchmark through Oct 2024.
[S2] Stanford HAI, 2026 AI Index — Technical Performance. Source — Top-model convergence and performance context.
[S3] McKinsey, Is that AI agent worth it? Agentic economics (13 Jul 2026). Source — Enterprise cost escalation and agentic economics.
[S4] FinOps Foundation, State of FinOps 2026. Source — 1,192 respondents representing $83bn+ annual cloud spend.
[S5] McKinsey, Cost versus value: Managing agentic AI system performance (8 Jul 2026). Source — Why per-token pricing is insufficient for enterprise economics.
[S6] McKinsey, The decision dividend: How AI creates economic value (26 Aug 2026). Source — Value mechanisms beyond labour savings.
[S7] McKinsey, Frontiers of compute: technologies to reduce AI inference costs (25 Jun 2026). Source — Inference architecture and cost drivers.
[S8] Deloitte, Tech Trends 2026 — AI infrastructure compute strategy. Source — Cloud/private compute economics and deployment drivers.
[S9] McKinsey, The state of AI in 2026 (25 Aug 2026). Source — Budget constraints, EBIT contribution and software substitution.
[S10] McKinsey, The cost of intelligence (20 Jul 2026). Source — CIO demand and cost governance.
[S11] Deloitte, AI tokenomics and infrastructure Dbriefs (27 May 2026). Source — 2026 AI infrastructure survey themes and tokenomics.
[S12] Reuters, AI investment boom puts Big Tech free cash flow under pressure (22 Jul 2026). Source — Capital-market evidence on infrastructure spending and return pressure.