November 12, 2025

65% of Enterprise AI Deployments Are Stalling — The Real Bottleneck

Your Weekly Deep Dive in the AI Next Wave

By Jay Cadmus

What Changed in November 2025

On November 4, Cognizant announced it would deploy Anthropic's Claude to 350,000 employees across engineering, delivery, and corporate functions. Not a pilot. Not a test. A full enterprise rollout combining Claude Code, the Model Context Protocol, and the Agent SDK with Cognizant's existing platforms.

Same week, Forrester published research showing 25% of enterprise AI investments slated for 2026 will be deferred until 2027. Not because the technology doesn't work. Because the gap between vendor promises and measurable business results is widening, not closing.

This contract tells the story. Cognizant is betting its operational future on AI implementation at massive scale. Meanwhile, a quarter of planned enterprise AI spending is getting pushed out because most companies can't figure out how to make it work.

The divide isn't between believers and skeptics. It's between organizations that can implement and those that can't.

The Numbers That Don't Lie

Fortune 500 AI adoption tripled in twelve months. October 2024: 22 companies had deployed enterprise AI platforms. October 2025: 67 companies, representing 13.4% of the Fortune 500.

That sounds like momentum. Until you look closer.

Only 6% of companies with 500 or more employees globally have deployed enterprise AI tools to their workforce. Among those who have, only 35% are achieving their expected ROI. Seventy-eight percent of organizations use AI in at least one function, but a mere 7% have deeply integrated it into operations.

The pattern is clear. Adoption is easy. Pilots are everywhere. Integration is rare. Production deployment with measurable outcomes? Vanishingly scarce.

Computing costs jumped 89% between 2023 and 2025, with 70% of executives citing generative AI as the primary driver. Every single executive surveyed by IBM canceled or postponed at least one AI initiative due to cost concerns.

This isn't a capability problem. The models work. This is an implementation crisis.

Why This Article Isn't Based on the MIT Study You Saw in Headlines

Before we go further, let's address the elephant in the room.

You've probably seen headlines screaming "MIT Study: 95% of AI Projects Fail to Deliver ROI." It's been everywhere. Forbes ran it. Fortune ran it. LinkedIn is full of people citing it.

Here's what those headlines don't tell you.

The MIT Media Lab study ("The GenAI Divide: State of AI in Business 2025") interviewed 52 people and surveyed 153 participants over a six-month window from January to June 2025. That's it. Fifty-two interviews and 153 survey responses to declare that 95% of enterprise AI shows "zero ROI."

Let's be blunt about the problems:

The Sample Size Problem

Declaring industry-wide failure rates based on 52 interviews is not rigorous research. It's a directional signal at best. Any undergraduate statistics course will tell you that 153 survey responses cannot represent the tens of thousands of companies deploying AI globally.

The Timeframe Problem

The study measured ROI within six months post-pilot. Enterprise transformation doesn't work on six-month cycles. ERP implementations take years to show full value. Cloud migrations take years. Declaring AI a failure after two quarters is like measuring the ROI of a college degree six months after freshman orientation.

The Data Quality Problem

Self-reported survey data is inherently noisy. When you ask survey participants if their AI project delivered ROI, you're capturing perception, politics, and incomplete information. Not audited financial results.

The ROI Definition Problem

The study focused primarily on measurable P&L impact within six months, which excludes strategic positioning, operational learning, capability building, and longer-term value creation. Some of the most important outcomes of AI adoption don't show up in quarter-over-quarter revenue.

The Selection Bias Problem (The Big One)

Here's the critical flaw nobody talks about. The MIT study explicitly focused on barriers to GenAI adoption. It asked participants what didn't work and why it didn't work. That's the entire framing.

When you design a study to identify friction points and implementation barriers, you get data about friction points and implementation barriers. What you don't get is a representative picture of the entire landscape. The study didn't balance its inquiry by equally investigating what worked, which organizations succeeded, what practices led to positive outcomes, or how companies overcame those barriers.

It's like surveying hospital emergency rooms to understand national health and concluding everyone is sick. You're sampling the population that's already struggling. You're not talking to the organizations that deployed successfully and moved on to scaling.

This is confirmation bias built into the research design. If your questions are "what went wrong?" and "why did it fail?", your conclusions will reflect failures. If you don't equally ask "what went right?" and "how did successful deployments differ?", you're not studying AI adoption. You're studying AI failure, then presenting it as if it represents the whole.

That's not objective research.

This doesn't mean the MIT study is worthless. It surfaced real barriers: organizational friction, lack of executive buy-in, unclear use cases, integration challenges. Those insights are valid. But extrapolating a 95% failure rate from that methodology, that sample size, that timeframe, and that framing is irresponsible. And news outlets repeating it without reading past the headline are doing their readers a disservice.

The Data That Actually Holds Up

The data in this article comes from different sources entirely:

Forrester Research, analyzing enterprise spending patterns across thousands of companies, looking at both successful and struggling deployments.

Bloomberry's analysis of 76,000 companies, tracking actual Fortune 500 disclosures and global enterprise AI deployments, representing the full spectrum of outcomes.

IBM's research on AI cost structures and executive decision-making across diverse industries and use cases.

Gartner's projections on agentic AI project cancellations, based on enterprise technology adoption patterns over decades.

Primary sources: Cognizant and Anthropic's official partnership announcements, OpenAI's hiring patterns, MCP technical documentation.

These are large-scale studies, longitudinal data, and primary company disclosures. Not 52 interviews framed around what went wrong.

The MIT study got headlines because "95% failure" is a shocking number from a prestigious institution. But shocking doesn't mean accurate. Prestigious doesn't mean methodologically sound. And framing a study around failures doesn't give you insight into the full picture.

When you see a study claiming massive failure rates, ask four questions: How many participants? Over what timeframe? How was success defined? And critically, was the study designed to find problems or to understand the full reality? If the answers are "small," "short," "narrowly," and "only problems," treat the conclusions as what they are: a narrow look at a specific cohort's challenges, not a definitive statement about an entire industry.

This article is built on data you can trust. Data that looks at the whole picture, not just the failures.

Now let's get back to what that data actually reveals.

The Human Layer AI Can't Replace (Yet)

Here's the part that should make you uncomfortable. The companies successfully deploying AI at scale aren't doing it by removing humans. They're adding a specific type of human back into the process.

Forward Deployed Engineers.

OpenAI is hiring them. Anthropic is hiring them. Every major AI company with enterprise customers is hiring them. These aren't your typical software engineers. They embed with customer teams for weeks or months, living inside the client's operations, understanding their workflows, translating business logic into AI-executable processes, and building production-ready systems that actually work in messy enterprise environments.

The role was pioneered by Palantir years ago and dismissed by many as an artifact of an older, less scalable business model. Consultants were supposed to be what AI replaced, not what AI required.

Turns out, wrong.

The Forward Deployed Engineer bridges the gap between what a model can theoretically do and what an organization can actually operationalize. They're not there to make the AI smarter. They're there to make the deployment real.

Cognizant's partnership with Anthropic exemplifies this. The company isn't just licensing Claude. It's combining frontier models with deep domain expertise and implementation capabilities to help clients move from pilot to production. The value isn't in the model alone. It's in the orchestration layer, the integration with existing systems, the workflow redesign, the governance frameworks, the human-in-the-loop controls that turn a powerful tool into a trusted production system.

This is the dirty secret enterprise AI vendors don't advertise. The bottleneck isn't model quality. It's implementation infrastructure. And implementation infrastructure requires people who understand both the technology and the business reality where it needs to operate.

The Protocol Layer You're Not Watching (But Should Be)

While everyone obsesses over the latest model release, something quieter but potentially more important is taking shape.

The Model Context Protocol.

Anthropic announced MCP in November 2024 as an open standard for connecting AI systems to external data sources and tools. Think of it as USB-C for AI. One protocol, universal connections, standardized interfaces.

Before MCP, every AI implementation required custom integrations for each data source. Want your agent to access Google Drive? Custom connector. Slack? Custom connector. GitHub, Postgres, internal databases? Custom connectors for each. The N×M problem, where N data sources and M AI applications create N×M integration efforts.

MCP collapses that into a single implementation. Build an MCP server once, and any MCP-compatible AI system can use it. OpenAI adopted it. Google DeepMind adopted it. Zed, Replit, Codeium, Sourcegraph integrated it.

This matters more than another benchmark improvement.

For agentic AI to scale (systems that plan, reason, and execute multi-step tasks autonomously), secure, standardized connections to tools and data are infrastructure requirements. MCP provides the registry, the compliance framework, the permission scopes, and the server identity verification that make agent actions auditable and controllable.

It's not sexy. It's plumbing. But plumbing determines what you can build.

Cognizant's deployment strategy explicitly incorporates MCP to give agents access to developer tools, enterprise systems, and data sources with proper governance. That's not a side detail. That's the implementation layer that makes 350,000-person rollouts possible without creating an ungovernable mess.

The companies winning at AI implementation aren't the ones with the best models. They're the ones with the best infrastructure for making models usable.

The Cost Wall Everyone's Hitting

Let's talk about what's actually killing AI projects.

AI implementation costs range from $50,000 for basic models to over $500,000 for complex enterprise solutions. That's just the initial deployment. Ongoing maintenance, model training, computational resources, compliance, and infrastructure upgrades add continuous expenses that many organizations didn't anticipate.

Cloud costs aren't just rising. They're becoming a barrier to successful scaling. The shift from experimentation to production means workloads that were tolerable as pilots become unsustainable at volume.

Forrester's prediction that 25% of enterprise AI spending will be deferred isn't about lost confidence in AI. It's about CFOs demanding financial rigor. The era of "let's try AI and see what happens" is over. The era of "show me the ROI in the first 90 days or the project gets cut" has begun.

Every executive that IBM surveyed has canceled at least one AI initiative due to cost concerns. That's not a few struggling companies. That's everyone.

The correction mechanism is already visible. Organizations are shifting from large general-purpose models to smaller, domain-specific ones that deliver better results at a fraction of the cost. They're optimizing cloud usage, using spot instances, implementing LLM routing to direct queries to the most cost-effective model, and employing quantization to reduce memory requirements.

The companies that figure out the cost structure early will have runway. Those that don't will burn through budgets before they reach production.

What the Implementation Crisis Reveals

Carnegie Mellon University found AI agents succeed only 30 to 35% of the time on multi-step tasks. Gartner predicts 40% of agentic AI projects will be canceled by 2027.

These aren't distant forecasts, but present realities being acknowledged.

The gap between what's technically possible in a demo and what's operationally reliable in production is wider than the AI industry wants to admit. Vendors oversell. Consultancies pitch transformation. Headlines hype revolution. Then reality arrives in the form of models that drift, compliance audits that fail, integrations that break, and costs that spiral.

The Bank of England warned of a potential "sudden correction" in equity markets driven by inflated AI valuations. The IMF described AI investment trends as reminiscent of the internet frenzy before the 2000 crash.

You can dismiss these as cautious institutions being cautious. Or you can recognize the pattern. Massive capital deployment. Uncertain monetization paths. Expectations disconnected from current capabilities. A rush to production before the infrastructure is ready.

We've been here before.

The Winners and the Losers

This correction isn't catastrophic, just clarifying.

The companies that win will be those with three things: platforms that orchestrate agents with governance and observability, protocols that standardize tool and data access (MCP or equivalent), and embedded implementation teams that turn capability into outcomes.

Cognizant's strategy embodies this. They're not betting on Claude being better than GPT or Gemini. They're betting on implementation infrastructure. The combination of frontier models, domain expertise, engineering platforms, and forward-deployed capability to move clients from pilot to production with measurable results.

Salesforce's Agentforce, despite mixed early adoption challenges, follows the same playbook. Multi-agent orchestration, policy controls, human-in-the-loop approvals, integration with existing Salesforce ecosystems. The value proposition isn't the agent. It's the governed, auditable, enterprise-ready deployment framework.

The losers will be organizations that keep chasing the next model release thinking better AI solves an implementation problem. They'll pilot endlessly, never reaching production. Or they'll deploy without proper governance and face security incidents, compliance failures, or operational breakdowns that set them back further than if they'd never started.

What This Means for You

If you're a business leader, stop asking "which model should we use?" Start asking "who can actually implement this in our environment, with our constraints, and deliver measurable outcomes in 90 days?"

The constraint isn't access to AI. Everyone has access. The constraint is implementation capability. Companies with embedded teams, platform infrastructure, and a disciplined approach to moving from pilot to production are pulling ahead. Those treating AI as a procurement decision are falling behind.

If you're building with AI, recognize that product-market fit now includes implementation-market fit. Can customers actually deploy your solution in their environment? Do you have the human infrastructure to help them? Can you prove ROI in their first quarter, not their third year?

The honeymoon phase where enterprises would pay for pilots that went nowhere is over. Financial rigor is here. The companies that survive the correction will be those that can ship, measure, and prove value quickly.

If you're watching the space, understand that 2026 is the year implementation capability becomes more valuable than model access. The cutting edge moves from "who has the best model" to "who can deploy at scale with proper governance."

The Uncomfortable Truth

AI is not failing. AI is succeeding past the point where success can be achieved with hype alone.

The bottleneck isn't intelligence, but it's the plumbing, governance, cost management, integration with legacy systems, training for teams that don't understand the technology, workflows that need redesigning, compliance frameworks that need updating, and the human expertise to navigate all of it.

The companies spending billions on AI infrastructure aren't wrong. They're betting on a future where this gets solved. But between here and there sits an implementation gap that's widening, not shrinking.

Cognizant deploying to 350,000 people matters because it's a live test of whether enterprise AI can scale past pilots. Forward Deployed Engineers matter because they're the human layer bridging the gap between capability and deployment. MCP matters because standardized infrastructure is how you move from bespoke integrations to scalable systems.

And the spending deferrals, the cost pressures, the canceled projects? They matter because they're the market telling the truth that vendors won't say out loud.

Making AI work in production is harder than making AI work in demos. And for the next 18 months, the companies that figure out implementation will win.

The rest will keep piloting.

What to Watch

Cognizant's 350,000-person rollout. This is the largest enterprise AI deployment announced to date. If it delivers measurable productivity gains with proper governance, it becomes the reference case every other enterprise studies. If it struggles, it proves the implementation gap is even wider than feared.

Forward Deployed Engineer hiring velocity. If OpenAI, Anthropic, and others keep expanding these teams, it signals the market recognizes implementation as the scarce resource. If hiring slows, it means they found another path to scale (or gave up on enterprise).

MCP adoption beyond the early cohort. Registry launch, compliance testing suites, enterprise deployments using MCP as infrastructure. If this becomes the standard, agent interoperability accelerates. If it fragments, we're back to custom integrations.

2026 Q1 earnings calls. Watch for CFOs asking about AI ROI. The language will shift from "we're investing in AI" to "here's the return we're seeing." Companies that can answer will get more runway. Those that can't will see budgets cut.

Project cancellation rates. Gartner's 40% figure for agentic AI projects is a prediction. The actual data will emerge over the next 12 months. If it's higher, the correction deepens. If it's lower, implementation is improving faster than expected.

The Blunt Assessment

Why This Isn't a Bubble (Yet)

You've heard the warnings. Bank of England officials talk about "sudden correction." IMF researchers compare AI spending to internet frenzy before the 2000 crash. LinkedIn is full of people convinced we're in a bubble about to pop.

They're not entirely wrong. They're just not entirely right either.

Let's separate the signals from the noise.

This Isn't a Bubble. It's Something Else.

In a bubble, you get the 2000 crash. Companies vanish. Entire markets collapse. There's no floor.

In a revenue realization crisis, you get market discipline. Companies that can monetize AI through implementation infrastructure pull ahead. Valuations reset from "growth at any cost" to "show me the ROI." Winners emerge. Losers consolidate or exit. The correction is painful for those not positioned for it. But it's not catastrophic for the underlying technology or the companies executing well.

We're in a revenue realization crisis. Not a bubble.

Cognizant betting 350,000 employees on Claude isn't a sign of desperation. It's a sign of scale. Forward Deployed Engineers hiring up across OpenAI and Anthropic isn't a desperation play. It's the market recognizing that implementation is now the constraint, not capability. MCP adoption expanding beyond the early cohort isn't hype. It's infrastructure.

The Correction Is Selective, Not Catastrophic

Here's what won't happen: Entire AI sector collapses. Models stop working. Enterprises abandon AI wholesale. The technology was over-hyped for months, but it's delivering.

Here's what will happen: 25% of planned enterprise AI spending gets deferred (it already has, per Forrester). Projects that can't show ROI in 90 days get cut. Valuations of startups without implementation credibility reset lower. The cloud bills for unoptimized deployments get shocking attention, forcing cost discipline. Companies without Forward Deployed Engineers or platform governance frameworks struggle to move past pilots. The winners pull further ahead.

This is correction. It's painful for companies without implementation infrastructure. It's healthy for the industry. It's not a bubble pop.

Within 18 Months

Can you deliver? Can you move from pilot to production? Who can prove ROI in Q1, not the 3rd year? Who has the people, platforms, and protocols to make AI work reliably in enterprise environments?

This is what matters. It's the differentiator. Are you positioned for what comes after?

References

Ropes & Gray. (2025, November 4). Artificial intelligence Q3 2025 global report. Retrieved from https://www.ropesgray.com

Bloomberry. (2025, November 8). The state of AI adoption in large orgs (analysis of 76k companies). Retrieved from https://www.bloomberry.com

Computing.co.uk. (2025, November 2). AI spending boom to cool as enterprises hit pause until 2027. Retrieved from https://www.computing.co.uk

The Register. (2025, October 27). AI bubble to deflate as enterprises defer spending to 2027. Retrieved from https://www.theregister.com

IBM Institute for Business Value. (2024, October 13). The hidden costs of AI: How generative models are reshaping enterprise computing. Retrieved from https://www.ibm.com

PR Newswire. (2025, November 3). Cognizant adopts Anthropic's Claude to accelerate enterprise AI adoption at scale. Retrieved from https://www.prnewswire.com

News Outsource Accelerator. (2025, November 6). Cognizant, Anthropic partner to scale enterprise AI. Retrieved from https://news.outsourceaccelerator.com

Wall Street Journal. (2025, October 13). AI startups have a new (old) secret weapon: Forward deployed engineers. Retrieved from https://www.wsj.com

Multimodal.dev. (2025, July 28). How our forward deployed engineers operationalize agentic AI. Retrieved from https://www.multimodal.dev

SSO Network. (2025, September 9). Forward deployed engineer: Turning AI promise into progress? Retrieved from https://www.ssonetwork.com

Anthropic. (2024, November 24). Introducing the Model Context Protocol. Retrieved from https://www.anthropic.com

Model Context Protocol. What is the Model Context Protocol (MCP)? Retrieved from https://www.modelcontextprotocol.io

Anthropic. (2025, November 3). Code execution with MCP: Building more efficient AI agents. Retrieved from https://www.anthropic.com

LinkedIn. (2025, November 5). AI reality check: What CTOs learned from 2025's AI implementations. Retrieved from https://www.linkedin.com

PromptLayer Blog. (2025, November 6). The GenAI divide: State of AI in business 2025. Retrieved from https://blog.promptlayer.com

Writer. (2025, November 9). A correction is coming to the AI market. Are you ready? Retrieved from https://www.writer.com

LinkedIn. (2025, June 15). AI implementation costs in 2025: Comprehensive breakdown. Retrieved from https://www.linkedin.com

Future Processing. (2025, November 6). AI pricing: How much does artificial intelligence cost? Retrieved from https://www.future-processing.com

Forbes. (2025, August 25). MIT finds 95% of GenAI pilots fail because companies avoid friction. Retrieved from https://www.forbes.com

Fortune. (2025, August 17). MIT report: 95% of generative AI pilots at companies are failing. Retrieved from https://www.fortune.com

Complex Discovery. (2025, August 21). Why 95% of corporate AI projects fail: Lessons from MIT's 2025 study. Retrieved from https://www.complexdiscovery.com

DataPro News. (2025, August 26). The MIT 95% AI enterprise failure rate report: True or false? Retrieved from https://www.datapro.news

MIT Sloan Management Review. (2025, September 30). AI implementation strategies: 4 insights from MIT Sloan management review. Retrieved from https://www.mitsloan.mit.edu

UC Berkeley Executive Education. (2025, September 17). Beyond ROI: Are we using the wrong metric in measuring AI success? Retrieved from https://exec-ed.berkeley.edu

Brookings Institution. (2025, August 11). Is there an AI bubble? https://www.brookings.edu

WIRED. (2025, November 5). The AI data center boom is warping the US economy. https://www.wired.com

CNBC. (2025, October 21). Are we in an AI bubble? Here's what analysts and experts are saying. https://www.cnbc.com

World Economic Forum. (2025, September 24). This is less an AI 'bubble' and more of a fragile 'boom'. https://www.weforum.org

Van Eck. (2025, October 14). Is AI a bubble? The dot-com bubble vs. today’s AI revolution. https://www.vaneck.com

Yahoo Finance. (2025, October 23). US AI revenue in 2025 nearly double closest competitor – analysis. https://finance.yahoo.com

Derek Thompson. (2025, October 1). This is how the AI bubble will pop. https://www.derekthompson.org

Fortune. (2025, September 27). Everyone's wondering if, and when, the AI bubble will pop. https://www.fortune.com

Bain & Company. (2025, November 6). $2 trillion in new revenue needed to fund AI's scaling trend. https://www.bain.com

Goldman Sachs. (2025, October 21). AI: In a bubble? https://www.goldmansachs.com

Business Insider. (2025, November 9). Stock market bubble: 5 warnings from dot-com boom to watch. https://www.businessinsider.com

Bessemer Venture Partners. (2025, August 14). The state of AI 2025. https://www.bvp.com

DW. (2025, November 10). AI bubble about to pop as returns on investment fall short? https://www.dw.com

iShares. (2025, November 5). Are AI stocks in a bubble? Why this isn't a dot-com redux. https://www.ishares.com

Reuters. (2025, October 9). US companies' profit growth seen softer, spotlight on AI spending. https://www.reuters.com

Understanding AI. (2025, October 26). 16 charts that explain the AI boom. https://www.understandingai.org

Aventis Advisors. (2025, October 2). AI valuation multiples in 2025. https://www.aventis-advisors.com