Topic · Enterprise AI cost and ROI
Verifying, tracking, and challenging the ROI claims vendors and analysts make about enterprise agentic AI.
The actual numbers. Procurement-ready cost analysis with workload-pinned math and dated pricing-page citations.
Enterprise AI cost analysis is dominated by two failure modes: vendor-supplied ROI calculators that assume their own success, and analyst headlines that quote a single percentage without naming the underlying survey question. This pillar exists because procurement teams need neither.
The McKinsey 17% EBIT-attribution figure, the CMU 30.3% agent capability gap, the GPT-5 Pro $200/month tier, the Salesforce Agentforce $800M ARR run-rate — every cost claim that gets cited at a CIO budget meeting belongs on this pillar with its underlying methodology audited. When the audit passes (Holding), the piece stays. When it doesn't (Partial / Not holding), the piece stays anyway with the correction appended; nothing is quietly removed.
What "cost" actually covers in 2026 enterprise agentic AI: per-token and per-call API economics across model providers, with workload profiles (RAG-heavy versus agent-loop-heavy) pinned in volume terms. Subscription tier value analysis — is the Pro / Business / Enterprise upcharge supported by the published rate-limit + feature delta or not? TCO comparators — model API + vector DB + observability + governance tooling stack versus cloud-native (Bedrock, Vertex AI, Foundry) integrated stack.
Failure-mode cost — what does an agent loop costing 1000× a normal completion actually look like, and what governance prevents it. Effort accounting — the realistic engineering hours per agent shipped, drawn from named-company case studies where the case study survives source verification.
Pricing data ages fast. Every cost-pillar piece declares a dated source-pull and runs on a 30-day review cadence rather than the standard 30–90.
Pillar last refreshed 2026-05-01
What survives review
- The MIT 95% GenAI-pilot-failure claim: what the State of AI in Business 2025 report actually measured — Holding · AM-128
- Mid-market agentic AI ROI in 90 days: what the cited data actually supports vs the vendor pitch — Holding · AM-129
- Agentic-AI vs human workers: the 2026 cost economics CIOs should actually model — Holding · AM-106
- The McKinsey 17% EBIT claim: what the survey actually measured — Holding · AM-033
- Why 88% of agentic AI deployments fail — Holding · AM-029
- The McKinsey 23%: the agentic AI scaling gap — Holding · AM-030
- The CMU 30.3%: the enterprise agent capability gap — Holding · AM-031
- The CFO's agentic AI business case: TCO and ROI — Holding · AM-027
- Build vs buy vs partner for enterprise agentic AI in 2026 — Holding · AM-028
- Google AI Mode restaurant booking: the template for every partner-aggregation vertical — Holding · AM-023
- GPT-5 Pro at $200 a month: what the pricing tier signals to enterprise IT — Holding · AM-003
- The bimodal ROI distribution in enterprise agentic AI: why the high-performing cohort is structurally distinct — Holding · AM-132
- The hidden costs of agentic AI: a CFO's guide to true TCO and ROI modeling — Holding · AM-020
- Production agentic AI cost: the layered optimisation playbook for enterprise CFOs — Holding · AM-061
What has broken
Nothing has moved to Partial or been retired in this topic yet.
Spoke articles
- The MIT 95% GenAI-pilot-failure claim: what the State of AI in Business 2025 report actually measured
MIT NANDA's GenAI Divide report (August 2025) is the source of the 2026's most-cited bear-case statistic: 95% of generative AI pilots fail. The number is a self-reported survey result with a specific methodology, and the way it gets read in procurement decks materially overstates what the underlying data supports. The structural findings underneath the headline are more useful than the headline itself.
- Mid-market agentic AI ROI in 90 days: what the cited data actually supports vs the vendor pitch
The 240% ROI in 90 days framing is the most common mid-market agentic AI vendor pitch in 2026, and the most-cited stat that no audited mid-market deployment has actually produced. Read against the McKinsey 17%, MIT NANDA 95%, and Stanford 12/88 data, the realistic 90-day mid-market ROI band is much narrower and much more useful for procurement than the pitch suggests.
- Agentic-AI vs human workers: the 2026 cost economics CIOs should actually model
Loaded FTE cost vs total agent operational cost does not favour replacement at parity in 2026 for most roles. The math works for narrow, high-volume task categories and breaks for judgment-laden ones.
- The McKinsey 17% EBIT claim: what the survey actually measured
The McKinsey 17% EBIT-attribution figure is the most-cited single statistic in 2026 enterprise agentic AI procurement. The way it is typically read materially overstates what the underlying survey supports.
- Why 88% of agentic AI deployments fail
Stanford 2026 data: 12% of agentic AI deployments clear 300%+ ROI; 88% miss. The distribution is not a capability problem. It is a governance gap.
- The McKinsey 23%: the agentic AI scaling gap
McKinsey 2025: 23% scaling, 39% experimenting. The pilot-to-production chasm is not about model readiness. It is about operational preconditions.
- The CMU 30.3%: the enterprise agent capability gap
Carnegie Mellon 2026: 30.3% task completion for best frontier models. The deployments that work operate within the 30.3%, not around it.
- The CFO's agentic AI business case: TCO and ROI
Most agentic AI business cases fail audit. Three documents survive: TCO with named components, ROI with pre-deployment baseline, scenario-weighted NPV.
- Build vs buy vs partner for enterprise agentic AI in 2026
Most enterprises frame agentic AI as build vs buy. It's a binary on a three-body problem. Partner — the third path — is systematically under-chosen.
- Google AI Mode restaurant booking: the template for every partner-aggregation vertical
Google shipped agentic restaurant booking to eight countries on 10 April 2026. The restaurant vertical is not the story. The story is that eight named.
- GPT-5 Pro at $200 a month: what the pricing tier signals to enterprise IT
OpenAI's GPT-5 Pro tier launched in August 2025 with no benchmarks and a $200/month subscription. The pricing decision is more interpretable than the capability claim. What the tier signals for enterprise procurement and how the McKinsey 17% EBIT-attribution figure cited around the launch should actually be read.
- The bimodal ROI distribution in enterprise agentic AI: why the high-performing cohort is structurally distinct
Enterprise agentic AI ROI is bimodal, not normally distributed. Stanford DEL, Gartner, McKinsey State of AI, and MIT NANDA data converge on the same shape: a small high-performing tail and a much larger struggling body. What separates the two is operational discipline, not model selection — and the 73%/27% framing in the slug captures that pattern more cleanly than the original AI-slop body did.
- The hidden costs of agentic AI: a CFO's guide to true TCO and ROI modeling
Enterprise TCO models underestimate agentic-AI programmes by 40-60%. The surprise is not that the costs are hidden. It is that they are distributed.
- Production agentic AI cost: the layered optimisation playbook for enterprise CFOs
Production agentic-AI bills routinely run several times the POC forecast. The mechanism is structural: token economics, orchestration overhead, context drift, observability. So is the optimisation.
What we're watching next
- Frontier-model inference pricing dropping another order of magnitude before end of Q3 2026.The LLMflation curve has held since 2022. If it continues, the cost-economics math shifts the boundary between replaceable and augmentable categories. If it breaks, the augmentation-over-replacement frame strengthens further.
- Published Fortune 500 / FTSE 100 case data showing residual-team productivity recovery curves materially shorter than the 6-12 month band.The retraining-gap claim is built on the published 2024-2025 cohort. The next round of named-company audits in the back half of 2026 either confirms or contracts the dip estimate.
- McKinsey and BCG publishing late-2026 enterprise gen-AI cost-multiplier studies.The current pilot-to-production cost multiplier is ~2-5×. If aggregate enterprise data shows compression toward 1.5×, the structural-multiplier framing across the cost cluster needs revising. If it widens, the procurement-playbook intervention becomes more urgent.
- GPT-5 Pro and Claude Opus 4 reasoning-mode adoption metrics becoming public.Reasoning models are 5-20× more expensive per call. The procurement question is whether enterprise deployments are routing-by-step (cheap most of the time) or routing-by-deployment (always expensive). Adoption data settles which pattern actually holds.
Primary sources we trust for this topic
A curated list of primary research, regulator guidance, and vendor documentation for enterprise ai cost and roi. Populated on the quarterly refresh — not a link dump, not competitors.
This pillar page is refreshed quarterly. Last refresh: 19 Apr 2026. Next refresh: 18 Jul 2026.