Thirty percent. That’s how many CEOs surveyed by PwC — 4,454 of them across 95 countries, interviewed between September 30 and November 10, 2025 — are confident in their company’s revenue growth for the next 12 months. Lowest in five years. Down from 56% in 2022.

And yet 90% of those same CEOs expect their companies to achieve revenue growth and profitability gains in 2026.

Sit with that for a second.

The macro is terrible. Everyone else is screwed. But us? We’ve got a plan. It’s AI plus M&A plus some joint ventures thrown in, and it’s definitely going to work this time. The data disagrees, pretty loudly. But the data doesn’t sit in the board meeting.

30%
CEOs confident in 12-month revenue growth — 5-year low
PwC 29th Global CEO Survey · 4,454 respondents
90%
Expect their own company to grow revenue and profitability in 2026
Same survey, same respondents
95%
AI pilots that fail to deliver marked, sustained P&L impact
MIT NANDA · 300 deployments analyzed · Aug 2025

The Confidence Paradox: Everyone’s Losing, You’re the Exception

There’s a specific cognitive trap here worth naming. In PwC’s data, 74% of CEOs report AI ROI — in KPMG’s parallel survey of 1,350 CEOs conducted August through September 2025, the number is similar. But MIT’s 2025 NANDA report Tier 1: peer-reviewed, 150 interviews + 350 employee surveys + 300 deployment analyses, published August 2025, says 95% of AI pilots fail to deliver “marked and sustained impact on P&L.”

Both numbers are real. They measure different things. CEOs are reporting any measurable value from early pilots. MIT tracked whether pilots scaled into recurring revenue or cost impact. Small early wins are common. Enterprise-wide transformation is rare. The gap between those two things is where capital goes to die — quietly, on a dashboard that still shows green.

“The standard monitoring stack was designed to catch system failures. It wasn’t designed to catch model failures that look like business successes in the short term.”

Editorial synthesis — sources: MIT NANDA (2025), IDC Enterprise AI Survey (2025), Gartner Digital Transformation Report (2025)

Which is, actually, the whole problem. And I’ll come back to that.

The AI Scaling Delusion: What 5% Success Actually Looks Like

MIT’s data is clear enough. About 5% of AI pilot programs achieve rapid revenue acceleration. The other 95% stall. S&P Global data shows 42% of companies scrapped most of their AI initiatives in 2025, up from 17% the year before — the average organization abandoned 46% of AI proof-of-concepts before production. IDC puts the POC-to-production failure rate at 88%.

What’s weird is that companies keep betting bigger. EY-Parthenon’s survey of 1,200 CEOs across 21 countries, November through December 2025, positions 2026 as the year AI transitions from pilot to enterprise scale. The CEOs reporting both cost and revenue gains are two to three times more likely to have embedded AI extensively across products, services, demand generation, and strategic decision-making. So the success stories are real. They’re just not distributed the way you’d hope.

The companies making it share three things that most don’t have going in:

Domain-specific solutions — not generic copilots — succeed at roughly 67% when vendor-led versus 33% for internal builds. MIT NANDA synthesis; treat as directional — vendor categorization is self-reported. Back-office automation — finance, compliance, operations — consistently delivers the highest returns. Sales and marketing pilots? Consistently the lowest. And data infrastructure is the unglamorous prerequisite: 43% of companies cite data quality as their top obstacle.

Second-Order Mechanism — Why You Can’t See the Failure

A failed AI pilot doesn’t announce itself. It produces output — often plausible-looking output — with the same confidence scores, the same dashboard formatting, the same apparent precision as a successful one. The systems built to detect AI performance measure whether the model is running, not whether it’s generating value.

This is the second-order problem. You don’t just fail to generate ROI. You fail in a way that actively resembles early-stage success. Eighteen months in, you have a team, a vendor contract, a quarterly review deck full of “promising signals” — and no measurable P&L impact. By the time the pattern is visible, you’ve committed budget cycles.

The M&A Contradiction: Swapping One Bad Bet for Another

Here’s what’s happening right now. CEOs are bailing on AI pilots — 42% scrapped most initiatives in 2025 — and pivoting to M&A. EY’s survey shows 53% of CEOs plan acquisitions in the next 12 months, up 5% from Q3 2025. Forty-two percent of companies are now competing in sectors they weren’t in five years ago, and 44% of those planners are targeting acquisitions outside their current industry.

They’re fleeing a strategy with a 95% failure rate. And they’re running toward one with a 70–75% failure rate.

The historical evidence here is about as solid as it gets. NYU Professor Emeritus Baruch Lev and University at Buffalo Professor Feng Gu analyzed 40,000 corporate acquisitions over 40 years, published in their November 2024 book The M&A Failure Trap. Tier 1: 40-year longitudinal dataset; data runs through 2018; deal-making dynamics structurally similar — see vintage note below. The 70–75% failure rate has held across decades. What’s gotten worse, according to Lev and Gu, is a “reverse learning curve”: the failure rate is increasing over time. Companies aren’t getting better at acquisitions. They’re getting worse.

Acquisition Type Evidence Relative Risk ⚠ What the Evidence Can’t Tell You
Conglomerate / outside core business Most likely to fail per Lev & Gu 40-year dataset Highest Dataset runs through 2018; post-COVID cross-sector dynamics may differ, though structural reasons for failure (integration complexity, cultural distance) are stable
Bolt-on in core business McKinsey reports 80–85% success rate for bolt-on deals in known markets Lowest McKinsey figure is directional — published in proprietary research with no publicly disclosed population size; treat as indicative, not audited
Stock-based deals in bull markets Higher failure probability when S&P 500 had strong prior-year returns (Lev & Gu) Elevated Causation unclear — bull markets correlate with overconfidence, but overpaying and overconfidence are hard to separate empirically
Operationally weak targets Serial acquirers who “resurrect failing targets” show particularly low success rates (Lev & Gu) High Defined as “operationally weak” post-acquisition, not pre; prospective identification of target weakness is subjective during due diligence
Primary source: Lev & Gu, The M&A Failure Trap (Nov 2024), analyzing 40,000 deals over 40 years. McKinsey bolt-on figure is directional, no independent audit found. Evidence levels: Strong = consistent findings across large longitudinal dataset; Directional = solid analytical basis but no disclosed population or independent audit.

One detail from Lev and Gu that doesn’t get enough attention: CEO compensation structures. Many companies pay acquisition completion bonuses — regardless of outcome. Serial acquirers serve four to five years longer as CEOs than peers who acquire less frequently. The incentives are not aligned with success. They’re aligned with activity.

“Companies aren’t getting better at acquisitions. The reverse learning curve is real. And the incentives explain why.”

Editorial synthesis — sources: Lev & Gu (2024), EY-Parthenon CEO Outlook 2026, PwC 29th Global CEO Survey (2025)

A vintage note, since this dataset is anchored in 2018: Lev and Gu’s 40-year study captures structural mechanisms — cultural integration failure, strategic overreach, incentive misalignment — that are demonstrably stable across economic cycles. The specific dollar figures for deal losses may have shifted. The underlying human dynamics have not. The 2024 publication with data through 2018 remains the most comprehensive longitudinal dataset on M&A outcomes available; no comparable post-2020 study has superseded it.

The Joint Venture Pivot: Same Problems, Different Structure

So if AI is failing and M&A is complicated, surely joint ventures are the answer. Right?

Seventy-nine percent of CEOs plan to pursue joint ventures and strategic alliances in 2026, per EY-Parthenon’s survey. That’s up from 62% in 2025. A 27-percentage-point jump toward a strategy that, depending on which research you cite, fails between 50% and 80% of the time. University of Oregon: ~50%; Forbes aggregate: ~60–80%; Harvard Business Review / McKinsey analysis: comparable to M&A failure rates. Ranges reflect different definitions of “failure” — dissolution vs. missed objectives vs. value destruction

The appeal is understandable. JVs promise speed. You get access to capabilities and technology without the complexity of full acquisition. Lower upfront commitment. More flexibility.

Here’s the catch. The complexity doesn’t disappear. It shifts. Instead of integration challenges, you get coordination challenges. You’re still merging different cultures, processes, and incentive structures — just without the formal ownership alignment that forces resolution. Strategic misalignment can persist indefinitely in a JV in a way it can’t survive in a full acquisition. The Harvard Business Review analysis by McKinsey consultants put it plainly: the success rate for JVs is on par with M&A. Not very good.

Cross-Source Synthesis — Not Present in Any Single Cited Source

Three datasets in combination produce a finding that none contains individually. MIT NANDA establishes that AI failure manifests as short-term signal noise — early pilots produce encouraging local metrics while P&L impact remains unmeasured. Lev & Gu establish that M&A value destruction typically surfaces 18–36 months post-close, after integration costs are sunk. EY-Parthenon’s 2026 data shows CEOs rotating from AI pilots to M&A to JVs within the same 12-month planning cycle.

The synthesis: all three strategies share the same second-order failure architecture. They fail slowly, in ways that produce short-term positive signals, in timeframes longer than most quarterly reporting cycles and most CEO planning horizons. The monitoring gap isn’t accidental — it’s structural. A CEO who rotates from AI to M&A to JV inside 24 months may never hold a failed strategy long enough to observe its failure mode clearly. They just observe the early positive signals, then move on when “momentum stalls.”

The Digital Transformation Floor: It’s Still Broken

While everyone’s chasing the new thing, the foundational problem hasn’t moved. Seventy percent of digital transformation initiatives still fail to meet their objectives in 2025, per Gartner. Gartner 2025; specific survey population not publicly disclosed — directional, treated as Tier 2. A separate Gartner survey finds only 48% of projects fully meet or exceed targets. Failed initiatives cost an estimated $2.3 trillion annually, globally. Aggregate estimate; no independent audit of this figure found; treat as directional order-of-magnitude

The 70% failure rate has been cited for at least five years. It hasn’t moved because the root cause hasn’t been addressed. Organizations keep buying technology expecting it to drive organizational change. That’s backwards. Technology automates existing processes — often including the dysfunctional ones. The transformation has to happen in how people work, and only 25% of organizations invest adequately in change management, per MeltingSpot’s 2025 analysis. MeltingSpot is a software vendor with commercial interest in change management investment; directional, not independently audited

Same mistake. Different decade. Different buzzword. “Digital transformation” is now “AI transformation” and the failure rate is similar.

A Named Failure: The Anatomy of a Stall

Named organizational AI failure cases are structurally scarce — companies don’t publish them, which is itself informative about how these failures circulate. What surfaces instead are practitioner accounts. No named brand published a case study matching the pattern below; the following represents the failure mode as described in MIT NANDA’s aggregate interview data from 150 executives

The pattern MIT documented across their 300-deployment analysis looks like this. A company launches an AI pilot in customer-facing operations — demand generation or sales support. Initial results are encouraging: response times improve, some conversion metrics tick up. The pilot gets expanded. A dedicated team forms. A vendor contract gets signed. At month nine, the P&L is flat. The team is producing dashboards full of activity metrics — leads touched, responses generated, processing speed — but revenue hasn’t moved. Finance asks for the ROI case. The team produces one using the activity metrics. It passes the review.

Month eighteen. Still flat. The vendor proposes a “second phase” that will unlock the real value. Leadership, having invested two years and a meaningful budget, approves it. The alternative — admitting the pilot failed — requires someone to own that decision.

Nobody owns it. The pilot doesn’t die. It zombifies. MIT NANDA describes this pattern across multiple case composites; specific company names are not disclosed in the public report

That’s the failure mode. Not a dramatic collapse. A slow, expensive fog, producing enough signal to survive every quarterly review, never enough to justify the investment.

What Actually Works: The 5% Playbook

The companies that make it work don’t share a secret strategy. They share a set of disciplines that most organizations find politically uncomfortable.

For AI: Back-office first. Finance, compliance, and operations consistently deliver the highest returns — partly because the success metrics are unambiguous. “Invoice processing time reduced 40%” is a measurable outcome. “Improved sales conversations” isn’t. Vendor-led domain solutions outperform internal builds at roughly 2:1. And the data infrastructure investment has to precede the model — 50–70% of budget on data readiness before anything else.

For M&A: Bolt-on deals in core business. Cultural due diligence as rigorous as financial. Start integration planning before the deal closes — companies that wait until post-close lose six to twelve months and the key talent they came to acquire. Lock in critical employees before the announcement, not after. And honestly consider whether internal R&D or organic growth solves the same problem at lower risk.

For JVs: Define exit conditions before you start. Run strategic alignment workshops before term sheets — surface the goal conflicts early, when they’re cheap to resolve. Build governance with clear decision rights; co-management paralysis is a predictable failure mode and it’s preventable. Start with a smaller scope than you think you need. Prove that collaboration works before you depend on it.

The 4-Week Reality Check

Week 1 — Evidence Audit

  • List every active transformation initiative
  • Assign honest probability of success using historical base rates, not exec optimism
  • Calculate total capital at risk in sub-50% probability bets
  • Present unfiltered data to board — no smoothing

Week 2 — Kill Criteria

  • AI pilots: define exact P&L impact required by month 6 or it dies
  • M&A pipeline: eliminate conglomerate deals regardless of banker enthusiasm
  • JVs: create governance with built-in exit clauses before signing
  • Set portfolio failure budget: “We can lose $X on experimentation”

Week 3 — Capability Inventory

  • Data infrastructure: can we actually run enterprise AI? If not, stop pilots
  • M&A track record: what % of last 5 deals hit targets? If under 50%: pause
  • Change management: dedicated transformation teams exist? If not, the digital effort will fail
  • Talent retention: can we keep acquired key employees? Uncertainty = value destruction

Week 4 — Strategic Honesty

  • Present board with base-rate failure statistics for each active strategy
  • Distinguish genuine strategic bets from “we had to do something” activity
  • Reallocate capital from statistically doomed strategies to proven capabilities
  • Set realistic expectations publicly — most pilots fail, most deals disappoint

The Finding That Complicates This Whole Argument

Here’s where I have to be honest with you. The 5% success rate for AI and the 70–75% failure rate for M&A are real — but the companies that beat those odds aren’t random. They’re specifically the ones who’ve already built the infrastructure and discipline that most organizations are still trying to acquire. PwC’s own data shows a 71% combined performance premium (profit margin plus revenue growth) for companies that successfully adapt their business and operating models through transformation.

That’s not a contradiction. It’s the actual problem. The strategies that statistically fail are the same strategies that produce massive premiums for the 5–30% that execute them well. You can’t skip them. You also can’t succeed at them without the foundational capabilities that the failure-rate data says most companies lack.

Which means the choice isn’t “do AI” or “don’t do AI.” It’s “honestly assess whether you have what it takes to be in the 5%, and if you don’t, build that first before you bet on it.” That assessment is harder than it sounds, because the people doing the assessing are the same people whose compensation is tied to moving forward.

“Nearly four in ten CEOs say their company will not be viable in ten years on its current path. Desperation doesn’t improve the odds. It just makes the bet feel more urgent.”

Editorial synthesis — sources: PwC 29th Global CEO Survey (2025), MIT NANDA (2025), Lev & Gu (2024)

For the C-Suite: What This Means in Your Decision Context

For: CEOs and C-Suite Leaders

The Monitoring Problem Is Your Problem, Not Your Team’s

The second-order failure pattern — strategies that fail slowly while producing short-term positive signals — is specifically a leadership governance problem. Your team is incentivized to show progress. The metrics they report up are the ones that survive quarterly review, not necessarily the ones that track actual value creation. That gap between “activity metrics” and “P&L impact” is something only you can close structurally, by changing what gets measured and what triggers a kill decision.

The 12-month CEO planning horizon is shorter than the observable failure window for all three major strategies. AI pilots stall at month 9–18. M&A value destruction surfaces at 18–36 months post-close. JV governance conflicts typically emerge at month 4–8 but become irreversible later. You need kill criteria set before you start — not optimized quarterly alongside the thing you’re trying to kill.

Stop reviewing AI pilots against activity metrics — speed, volume, conversations generated — and treating them as progress. The MIT NANDA data is explicit: companies reporting early AI ROI and companies reporting sustained P&L impact are mostly different companies. If your pilot can’t point to a specific, audited P&L line by month 6, it’s in the 95%.

For: Board Members and Strategy Leads

The Incentive Misalignment You’re Probably Not Asking About

Lev and Gu’s most underreported finding: serial acquirers serve four to five years longer as CEOs than peers who acquire less. Many compensation structures include deal-completion bonuses independent of deal outcomes. This creates a specific governance risk that’s different from general M&A failure risk — it means the CEO’s optimal career strategy may diverge from the company’s optimal value strategy, especially as deal pipelines build during market optimism cycles.

The board’s role here isn’t to approve or reject specific deals. It’s to audit whether the measurement systems — for AI, M&A, and JVs — are tracking value or tracking activity. The difference sounds obvious. In practice, activity metrics are always more available, more frequent, and more encouraging in early stages. Value metrics require patience and sometimes require recording failures explicitly.

Stop accepting “we’re in the learning phase” as a response to flat P&L impact beyond month 9 for AI pilots. The MIT data shows that companies successfully scaling AI were not in a longer learning phase — they had different capabilities going in. A learning-phase narrative for a flat pilot is a sunk-cost justification, not a progress report.

The Question You Actually Have to Answer

Nearly four in ten CEOs in PwC’s survey say their company won’t be viable in ten years on its current path. That creates real urgency. And urgency, without discipline, produces exactly the pattern we’re seeing — simultaneous bets on AI, M&A, and JVs, all with historically brutal failure rates, all launched faster than the monitoring systems can track them.

The companies that succeed aren’t the ones who found a strategy with better odds. They’re the ones who built the capabilities that made the odds better — before they bet. Data infrastructure before AI. Cultural integration rigor before acquisition. Governance clarity before the JV term sheet.

That sequence is boring. It’s slow. It doesn’t show up in an earnings call announcement. And it’s the only thing the data actually supports.

So the question isn’t whether to pursue AI or M&A or alliances. The question is whether you’re genuinely building the 5% capabilities — or counting on your confidence to be the variable that makes a 95% failure rate not apply to you.

It won’t be.