AI's Year of Truth: Eight Consultancies, 11 Reports, One Uncomfortable Conclusion

I opened the latest AI report, then ten more - McKinsey, BCG, Bain, Deloitte, KPMG, Accenture, Capgemini, PwC, all 2025-26 editions. Eleven reports. I expected them to disagree: each with its own list of trends, its own pretty words. Instead I got rare agreement. The kind where these same reports’ own headlines contradict themselves.

TL;DR

• Agentic AI is the number-one trend at all eight firms, and in every one of their own datasets it stalls on adoption.
• Agents are in production at 11% of companies, at full scale at 2%. Meanwhile 90% of CEOs expect ROI from them already in 2026.
• Deloitte’s CTO puts the split at 93% of AI spend on technology and tooling, 7% on everything else: culture, change, redesigning the work. Exactly the opposite of where the value sits.
• The gap isn’t about models, it’s about processes, data and people. So it’s about work, not magic. Hard, but it works.
• That’s exactly why the potential is huge: the market paid for the hype up front, and execution is lagging behind.

The gap nobody puts on a slide

The number-one trend at all eight is agentic AI. And right there, in their own numbers, it’s failing across the board.

See for yourself. Deloitte: agents are in production at 11% of companies, 42% are still writing the strategy, 35% haven’t even started. Capgemini, on its sample: at full scale, 2%. KPMG: 24% get ROI across several AI use cases, and that’s 7 percentage points lower than a year earlier - meaning the industry went backwards. McKinsey: 78% use AI in at least one function, but only 1% call themselves mature.

Worth slowing down on that 2% and 24% for a second - the fine print matters. Capgemini’s own ladder underneath the headline: 2% at full scale, 12% at partial scale, 23% still piloting, 61% just exploring, so “2%” is one rung on a real staircase, not a wall. KPMG’s own number for “have agents somewhere” is 88%, worlds apart from Deloitte’s 11% in production - eight firms, eight different definitions of the same word, which is exactly why these headline percentages don’t stack into one funnel. Among KPMG’s 24% who do get ROI, the split is stark: leaders report 4.5x, the average is 2x, and the 53% who can’t close that gap say the same thing every time - not enough people who know how to finish the job.

Now the other half of the picture. BCG: AI budgets doubled in a year and will double again in 2026 - from 0.8% to 1.7% of revenue. 90% of CEOs expect measurable ROI from agents this year. 94% will keep investing even if it doesn’t pay off within the year.

And the real surprise: Capgemini measured that trust in fully autonomous agents fell over the year from 43% to 27%. It’s falling, in their own words, “from experience, not from fear” - meaning people tried it and got disappointed. There’s the gap: more and more money pouring in, almost nobody in production, and the ones who tried it now trust agents less. The year of truth.

Why it doesn’t take off - and why that’s good news

The most interesting part isn’t “how much”, it’s “why”. And here the reports unexpectedly describe exactly what I do every day.

Gartner (quoted by Deloitte): by the end of 2027 more than 40% of agent projects will be abandoned - “not because the technology doesn’t work, but because companies automate broken processes instead of redesigning them”. Deloitte calls things by their names: “agent washing” - when ordinary automation gets repainted as agents, and “workslop” - when a badly built agent makes the process slower.

Everyone’s formula for success converged on the same thing. PwC: 80% of the value is in redesigning the work, 20% in the technology. BCG: 10-20-70, where the 70% is people and processes. Bain puts it in numbers: a bare coding assistant gives 10-15% productivity, the same assistant plus an end-to-end process redesign gives 25-30%. Twice the difference, and all of it is in the organization, not the model.

Why the gap between 10-15% and 25-30% is that wide has a plain mechanical answer: writing and testing code is only 25-35% of the whole development cycle, so speeding up just that slice barely moves the total. The higher-band teams didn’t get a better model, they zero-based the whole cycle and rebuilt it around the agent instead of bolting one onto the old process - Intuit did exactly this, using generative AI to standardize code and documentation across every role that touches a product, not just engineers. And Bain’s EBITDA number isn’t a survey guess either: it’s outcomes tracked from real clients Bain walked through 2023 and 2024, after they had already pushed AI past the pilot stage into core workflows.

And the line I keep coming back to is from Deloitte’s Tech Trends 2026, where their CTO says it plainly: 93% of all AI spend goes to the tech and the tooling, and only 7% to everything else - culture, change, learning, redesigning how the work actually runs. Deloitte publishes no sample behind that split, so read it as a practitioner’s estimate rather than a survey result. It still lines up with every number above.

Plus the foundation. Fewer than 1 company in 5 is ready on data; the main blockers are that data can’t be found (48%) and can’t be reused (47%). Capgemini’s line belongs on a wall: “if your data isn’t ready for AI, your business isn’t ready for AI”. Bain adds a down-to-earth point: the first step of any AI transformation is cleaning out old and broken data, “sometimes up to 80%”, and you can’t cut a corner here.

I read this and see my own work described in someone else’s words. My thesis at KIT is exactly about this - how manufacturing SMEs should choose where to put AI instead of sticking it everywhere. The working method: first break the process down in BPMN, then a pilot. Exactly what gets skipped when 93% of the money goes into tooling.

The layer few people see

Under all the talk about “agents” there’s an engineering layer, and the reports speak about it more precisely than you’d expect.

MCP - Anthropic’s protocol, the one an agent uses to reach data and tools - grew from ~100 servers in November 2024 to ~7000 by July 2025. A 70x jump in eight months. But Bain cools it down right away: “MCP is not USB”. It standardizes the syntax, not the meaning: two agents can “talk” over the protocol and still not understand each other. Whoever sets the semantic standard - how an invoice bot agrees with a payment bot - will capture the next big wave of value.

A whole discipline called “agent-ops” has appeared: FinOps for agents (the token bill runs away into the tens of millions because an agent thinks non-stop), a registry of models and agents, “every agent has an owner, a KPI and a rollback plan”. The average Mittelstand factory has none of this. And vendor neutrality is now direct advice from Bain: “choose vendors so as to limit lock-in and keep optionality”.

And on humans in the loop. 90% of organizations consider human oversight useful or at least cost-neutral; SAP advises starting with “read-only AI” - the agent recommends, the human decides; the EU AI Act flat-out requires oversight for high-risk systems. A live lesson: Klarna publicly rolled back replacing support with AI and brought people back. Bain puts it nicely: for now these are “Iron Man suits, not autonomous Iron Men”.

What’s actually on the shop floor, and why the wind blows toward the EU

Manufacturing here isn’t an abstraction. PTC and Microsoft are building a multi-agent setup that reads and acts on its own through PLM, ERP and MES. Siemens has built an industrial orchestrator, a “digital foreman”. Foxconn counted more than 400 million dollars in savings from an AI model across 200+ plants. Toyota replaced 50-100 mainframe screens with a single supply-chain agent. BMW already drives cars under their own power along kilometer-long routes inside the plant and is testing humanoid robots at its Spartanburg site. All of this is end-to-end redesign, not a bow on top.

For a factory the cloud often doesn’t fit physically: you need latency under 10 ms on the line, plus data sovereignty and IP protection. Deloitte gives a rule I can apply straight to a client’s invoice: on-prem wins as soon as the cloud starts costing more than 60-70% of the price of your own hardware. The new normal is three tiers: cloud for elasticity, on-prem for stability, edge for instant response.

And the political wind is blowing toward my niche. Deutsche Telekom and Nvidia are building an industrial AI cloud for European manufacturers. EU InvestAI is 200 billion euros, including 20 billion for AI gigafactories. The EU AI Act is already coming into force. The regulation everyone treats as a brake can, in the EU niche, be sold as a head start.

Five things that break the mold

The reports also gave up some real surprises - the stuff that doesn’t make the headlines.

One. AI isn’t a job killer, it’s a job creator. The most AI-heavy companies grew headcount faster: 52% versus 36% at the least AI-heavy ones (PwC, analysis of a billion job postings). But the prize concentrates wildly - 20% of companies take 74% of all the gains, and the top 20% by productivity grew 163%, five times the average.

Two. On juniors everyone reads only the headline. Yes, 49% of CEOs expect AI to cut junior hiring. But in the numbers there’s a twist: junior roles that “grew up” (started requiring 10+ traditionally senior skills) rose 35% since 2019, while the ones that didn’t fell 10%. The bottom rung isn’t disappearing - it’s being pulled upward.

Three. The hyped skill is in oversupply, the boring one is scarce. There are 3.6 times more prompt engineers than demand, while Python (0.4x) and machine learning (0.5x) are in hard short supply (McKinsey). Everyone ran for the trendy thing, and the bottleneck is good old engineering.

Four. The wage premium for AI skills is 62%, and in manufacturing 73%. Demand specifically for agentic skills grew 985% in a year - faster than all 13 trends at McKinsey. As Google’s director of AI go-to-market put it: “90% of your organization will need training just to understand how this even works”.

Five. A couple of numbers that make you sit up straighter. The inference token got 280 times cheaper in two years - and the bills grew anyway, because agents burn tokens faster than the token gets cheaper. Code agents already write 30% of the code at Microsoft. And a quiet bomb under the noise: 97% of security people consider quantum risk material within a 10-year horizon, and only 9% have a plan. The “harvest encrypted data now, decrypt it with quantum later” scheme is already running, and voice phishing with voice cloning grew 442% in half a year.

The conclusion that seems unpopular

Putting it together. The technology works - where it’s been carried through to the end: 10-25% added to EBITDA at the leaders per Bain, tens of percent in productivity, real factory cases with names attached. But “carried through” means 2-11% of companies. Between “everyone’s investing” and “almost nobody’s in production” there’s a gaping hole, and that hole isn’t about models, it’s about processes, data and people. So it’s about work, not magic.

Hard. But it works. And that’s exactly why the potential is huge: the market paid for the hype up front, and execution is lagging behind. The winner isn’t whoever buys agents fastest, it’s whoever first fixes the process and the data underneath them. It’s boring, it’s exactly my job - and eight of the biggest consultancies just said the same thing, each in its own numbers.

This is the first run of my radar. I’ll go through these reports twice a year and work out what actually moved and what stayed a slide. In the next part I’ll dig one branch deeper - most likely the EU-sovereign stack for the Mittelstand, because that’s where the money, the regulation and my profile all meet. To be continued.

Sources

All the numbers come from the firms’ own fresh reports (2025-26 editions). The full list of what I went through:

• McKinsey - Technology Trends Outlook 2025 (13 trends) + MGI “The race takes off”
• BCG - AI Radar 2026 (survey of 2360 executives, 640 CEOs) + Most Innovative Companies 2025
• Bain - Technology Report 2025 (the meatiest of the lot)
• Deloitte - Tech Trends 2026
• KPMG - Global Tech Report 2026 (2500 executives, 27 countries)
• Accenture - Technology Vision 2025
• Capgemini - TechnoVision 2026 + Rise of Agentic AI 2025
• PwC - Global AI Jobs Barometer 2026 (analysis of ~1 billion job postings) + AI Business Predictions 2026