A hypothetical board meeting. Four people quote four reports, all of them correctly:

• A director: 95% of AI pilots fail. That is MIT's count of firms with zero P&L impact from custom GenAI tools, based on 52 interviews and 153 conference respondents in early 2025.
• The CFO: 44% of companies run AI at enterprise scale. That is McKinsey respondents judging their own scaling, 1,719 of them, mid-2026.
• The CTO: open models carry a third of all AI traffic. That is developer traffic through one model router, OpenRouter.
• Procurement: open models are 11% of enterprise use, and shrinking. That is API usage at 495 US enterprises, surveyed by Menlo Ventures.

Each report measures something different. Ask who was asked and what was counted, and most contradictions dissolve. What remains is where the real uncertainty sits.

The position: no single report tells you where enterprise AI stands. Read them side by side and treat the disagreements as the finding. Apply the same test to your own AI metrics.

The short version:

• Use is near universal, value is not: around 9 in 10 organisations use AI somewhere; 37% report any EBIT impact (McKinsey), and 20% see revenue growth from it (Deloitte).
• Spend rises, unit cost falls: US enterprise GenAI spend tripled to $37B in 2025 (Menlo), while token prices fell 41% in six months (Ramp).
• Multi-model is already the norm: 81% of large enterprises run three or more model families (a16z).
• "Agent" has no shared definition: 16% of enterprise deployments are true agents (Menlo), 40% of large companies are scaling agents (McKinsey), 69% of Australian leaders say they use autonomous agents (Deloitte).
• APAC is under-measured: two of the major reports are US-only and none has an ASEAN enterprise cut, yet Singapore shows 61% population adoption and 9.21% of global OpenRouter tokens.

In this article:

Part 1: What the reports say

• What the reports measure: buyers, workers, telemetry and indices, and who paid
• One word, several meanings: adoption, production, value and agent
• Where the reports agree: six findings that hold across methods
• Where they contradict each other: six disagreements and what causes them
• What the reports miss: APAC: the region the benchmarks under-sample
• How to read the next AI report: seven questions before you quote a number
• Where this is going: how the evidence base changes next

Part 2: The reports, one by one

• Twelve reports, one card each: sample, method, headline numbers, what to watch for

Part 1: What the reports say

What the reports measure

Twelve sources cover most of the enterprise AI conversation in 2026. They fall into four kinds, and the kind decides what a number can mean.

The four kinds:

• Buyer surveys (Menlo, a16z, McKinsey, Deloitte, MIT) tell you what leaders believe and plan. They are good on intent and spend, weak on what actually runs.
• Worker surveys (BCG, Mozilla) tell you what people do day to day. They are good on usage, weak on business value.
• Telemetry and spend data (OpenRouter, Ramp, Hugging Face) count what actually happens. They are precise about their own platform and blind to everything outside it.
• Indices and forecasts (Stanford, Gartner) aggregate or predict. They inherit the strengths and the biases of their inputs.

Then check who paid. Menlo and a16z are venture investors with portfolio exposure to the vendors they rank, and both disclose it.

McKinsey, Deloitte and BCG sell transformation services, so "value is hard to capture without redesign" is both a finding and a sales argument. Mozilla advocates for open-source AI.

None of this makes the numbers wrong. It tells you which way a borderline judgement is likely to lean.

One word, several meanings

Most apparent contradictions come from four words that every report uses and no two define the same way.

The McKinsey row is the clearest case. The same survey, the same respondents, finds 80% individual productivity gains and 37% enterprise EBIT impact. Both are true.

They measure different things.

When you read "AI adoption", ask which row it sits in before you compare it to anything.

Where the reports agree

Some findings hold across buyer surveys, worker surveys and telemetry alike.

Use is broad, value is concentrated: McKinsey counts nearly 9 in 10 respondents using AI regularly, but only 37% report EBIT impact and about 6% qualify as high performers. Deloitte finds 66% reporting efficiency gains and 20% reporting revenue growth, while 74% still hope for it. BCG shows the leak: 42% of frontline regular users save at least a full day a week, and 66% of those get little or no guidance on what to do with the time.

Spending keeps rising, unit prices keep falling: Menlo puts US enterprise GenAI spend at $37B in 2025, up from $11.5B in 2024. a16z's CIOs expect their average LLM spend to rise about 65%, to $11.6M. McKinsey has 60% expecting AI investment to increase.

At the same time Ramp measures the effective price per million tokens down 41% from its March 2026 peak, and median AI spend per employee at its top 1% of spenders down 9.7%. Total spend and unit cost are moving in opposite directions.

Buying beats building, for now: Menlo finds 76% of AI use cases purchased, up from 53% a year earlier. MIT finds external partnerships reaching deployment about twice as often as internal builds. The first crack: McKinsey reports 32% of respondents avoided buying software because coding agents let them build it internally.

Coding is the first workload at scale: it is the largest departmental AI category in Menlo's data at $4.0B, 55% of departmental spend. On OpenRouter, programming grew from about 11% of tokens in early 2025 to over half. Hugging Face reports AI agents became the largest user of its hub in July 2026.

Multi-model is the norm: a16z finds 81% of large enterprises using three or more model families, up from 68% a year earlier. Even the CIOs who prefer closed models run several of them. Portability is no longer a theoretical concern.

Work changes before headcount does: McKinsey has 39% expecting AI-driven headcount reductions next year but only 14% reporting actual reductions last year. BCG finds 47% of workers spending more time directing AI than doing the work themselves. The Stanford AI Index records employment of software developers aged 22 to 25 down nearly 20% since 2024: the first measurable effect lands on entry-level roles, not across the board.

The Stanford AI Index sometimes draws its corporate adoption figure from McKinsey's survey. Its 2026 economy chapter labels the 88% organisation-adoption figure as McKinsey data, so that number appearing in both is one measurement quoted twice, not two confirming each other. Agreement only counts when the methods are independent.

Where they contradict each other

The same numbers look contradictory because the denominators change.

1. Do AI pilots fail?

Cause: MIT asks the hardest question (measurable P&L impact of custom-integrated tools) of a small, early sample. Menlo counts bought products, which arrive production-ready. McKinsey asks respondents to judge their own scaling.

The "95% fail" headline survives because it is the most quotable, not because it contradicts the others. It measures a different and narrower thing.

2. How big is open source?

Cause: each source sees a different slice. US enterprise procurement buys closed APIs. Developers experiment with open models heavily and route a lot of traffic through them.

Self-hosted open models never appear in API or card-spend data at all, so Menlo and Ramp undercount them by design. Mozilla's production gap (51% vs 63%) is the honest middle: open models get used everywhere, but reach production less often, mostly for operational reasons. Mozilla also puts open models at about a third of real-world usage and 4% of revenue: lots of tokens, little vendor revenue, because the value accrues to whoever runs them.

3. Who leads the vendor race?

Cause: "used in production somewhere" favours the vendor with the broadest footprint. API spend favours the vendor that dominates coding, the largest workload. Card spend favours whichever tool individual teams buy themselves.

4. How common are agents?

Cause: definition, almost entirely. Gartner's "agent washing" point explains the spread: assistants, RPA and chatbots relabelled as agents inflate self-reported numbers. The gap that matters is Deloitte's: 69% using agents against 22% governing them properly.

5. Is there ROI?

Cause: productivity is easy to feel and hard to bank. Time saved only becomes EBIT when someone redesigns the workflow, cuts a cost or sells more. All four reports point to the same mechanism from different angles.

BCG's number is the most actionable: strategy and redesign move the result five times more than better tools.

6. Where is the workforce effect?

Cause: the effect is real and narrow. It hits entry-level and routine roles first, and outsourced functions before internal staff. Company-wide surveys average it away.

What the reports miss: APAC

Line up the reports by geography and a gap appears.

The usage-based sources show APAC ahead.

Singapore's population adoption beats the US by a wide margin. Singapore alone generates more OpenRouter tokens than China, and more than Korea, Japan and India combined, although billing location may flatter it: OpenRouter uses billing geography, and enterprise accounts can aggregate activity across multiple regions under one billing entity. East Asian developers use open models more than anyone.

The buyer surveys that set boardroom benchmarks barely sample the region. So an APAC company comparing itself to "the market" is usually comparing itself to US enterprises. That benchmark understates open-model use, overstates concentration on two US vendors, and misses the procurement questions that dominate here: Deloitte's Australian respondents say 72% consider a vendor's country of origin, a question US surveys rarely ask.

No source in this set publishes an ASEAN enterprise benchmark.

How to read the next AI report

A new report with a new headline number will appear next month. Seven questions sort a useful number from a quotable one.

Apply the same questions to your own internal AI dashboard. "Adoption" in your company probably means licences issued. Value means something else.

Where this is going

The evidence base is moving in two directions at once. Surveys still set the boardroom numbers, but the faster sources are telemetry and spend: Ramp updates monthly, OpenRouter counts tokens, and Hugging Face counts model activity. Those sources will not replace surveys because each is platform-specific, but they make it harder for "adoption" to stand in for usage.

The next useful editions are the recurring ones. Ramp is already monthly. Stanford's AI Index is annual, McKinsey's State of AI is annual, and a16z describes its CIO survey as annual.

Menlo's December 2025 report is the spend baseline to watch because its 2024-to-2025 jump from $11.5B to $37B is one of the few hard enterprise-spend series in the set.

The question will also shift from pilots to production evidence. MIT's 5% production figure for task-specific GenAI, Menlo's 47% production conversion for AI purchases, and McKinsey's 44% enterprise-scale figure cannot be reconciled until reports separate custom builds, bought software, workflow integration and P&L impact.

Model choice will become less useful as a headline. a16z already has 81% of large enterprises on three or more model families, and Ramp has token prices down 41% in six months. Once model switching is normal, the sharper measurement is cost per completed task, not vendor share or tokens alone.

The weakest part of the evidence base is still APAC enterprise data. Singapore shows 61% population adoption and 9.21% of OpenRouter tokens, but no report in this set publishes an ASEAN enterprise cut. That gap will matter more as vendor origin, open models and regional procurement rules become part of the buying decision.

Part 2: The reports, one by one

Twelve reports, one card each

Each card has the same two tables: how the report was built, then its key numbers.

Menlo Ventures: 2025 State of Generative AI in the Enterprise

The most quoted source on enterprise AI spend and vendor share. Menlo combines a buyer survey with a bottom-up market model, which makes it strong on money and weak on anything that does not show up in a purchase.

How to use it: for US enterprise spend, vendor share and the build-vs-buy split. Do not use its 11% open-source figure as a global number: self-hosted models are invisible to an API spend model, and the sample is US only.

a16z: Enterprise CIO survey

A small, senior and deliberately selected panel. It shows what the largest companies are doing, which is often where the rest of the market goes twelve months later.

How to use it: for the direction of travel at Global 2000 scale, especially multi-model adoption and incumbent-vendor preference. Do not read 100 hand-picked leaders as the market average.

McKinsey: The state of AI in 2026

A long-running annual management survey on AI. Its strength is the time series: largely the same questions, asked every year, across regions and company sizes.

How to use it: for trends in scaling, agents and EBIT impact, and for the gap between individual productivity (80%) and enterprise value (37%). Remember that every number is a respondent's own judgement.

Deloitte: State of AI in the Enterprise, 2026

A large survey of senior leaders that is unusually specific about which benefits have materialised and which are still hoped for. The Australian cut is one of the few published APAC breakdowns.

How to use it: for the realised-vs-hoped gap (20% revenue growth against 74% expecting it) and for agent governance. Note that fieldwork ran in August and September 2025.

MIT NANDA: The GenAI Divide, State of AI in Business 2025

The source of the most repeated statistic in enterprise AI. The report itself is careful: it measures measurable P&L impact from integrated, task-specific GenAI, and labels its findings preliminary. The way it is quoted is not.

How to use it: for the gap between generic tools that staff adopt on their own and custom tools that rarely reach production, and for the case for external partners. Do not quote "95% of AI fails" without the qualifiers.

BCG: AI at Work 2026

The largest worker survey in this set, and the best source on what AI does to daily work. It says little about spend or vendors, which is exactly why it complements the buyer surveys.

How to use it: for where productivity leaks: time saved that nobody redirects, and the finding that strategy moves impact five times more than tools.

Mozilla: State of Open Source AI 2026

The first edition of a developer-side view of open models. Its most useful number is the production gap: developers use open models widely but deploy them less often than closed ones.

How to use it: for open-model usage among developers and the reasons it does not always reach production. Weigh it against Mozilla's open-source advocacy.

OpenRouter and a16z: State of AI, a 100-trillion-token study

Measured traffic, not opinions. OpenRouter routes requests to hundreds of models, so its logs show which models developers actually call, at what volume and for what kind of task.

How to use it: for model mix, open-weight share, the rise of reasoning and coding workloads, and Asian usage. Do not treat developer and startup traffic as enterprise production.

Ramp AI Index, September 2026

A monthly index built from what US businesses actually pay for. It is the fastest-moving source in this set and the only one that tracks unit prices and spend per employee month by month.

How to use it: for paid adoption among US businesses, vendor momentum and token price trends. It cannot see self-hosted models or AI bought inside cloud contracts.

Hugging Face: State of Open Models, Summer 2026

The view from the largest model hub: what gets released, downloaded and adapted. It is the best source on which open models developers build on, and on the shift in open-model leadership to Chinese labs.

How to use it: for open-model momentum, licence patterns and the dominance of small models in practice. Downloads are not deployments.

Stanford HAI: AI Index 2026

An annual compendium rather than a survey. It is the best single source for investment, research output, benchmarks and country comparisons, including Singapore's lead in population adoption.

How to use it: for macro figures and international comparison. Check where each number comes from before treating it as independent confirmation of another report.

Gartner: agentic AI predictions

Forecasts from Gartner's analysts. Its agentic AI release is useful less for the headline cancellation rate than for naming "agent washing": products relabelled as agents without agentic capability.

How to use it: as a planning assumption and a reality check on vendor claims. These are predictions, not measurements.

Sources

• Menlo Ventures: 2025 The State of Generative AI in the Enterprise (December 2025)
• a16z: Leaders, gainers and unexpected winners in the Enterprise AI arms race (January 30, 2026)
• McKinsey: The state of AI (August 25, 2026)
• Deloitte: The State of AI in the Enterprise (2026)
• MIT NANDA: The GenAI Divide, State of AI in Business 2025 (July 2025)
• BCG: AI Is Reshaping Jobs Faster Than Companies Are Reshaping Work (June 3, 2026)
• Mozilla: State of Open Source AI 2026 (September 2026)
• OpenRouter and a16z: State of AI (December 2025)
• Ramp AI Index, September 2026 (September 9, 2026)
• Hugging Face: State of Open Models, Summer 2026 (August 14, 2026)
• Stanford HAI: The 2026 AI Index Report (April 2026)
• Gartner: Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (June 25, 2025)
• rAInvent: Open weights don't make enterprise AI cheaper. They make it portable. (October 3, 2026

)