What everyone gets wrong about that MIT study on AI

There’s a headline that’s rattling leaders in boardrooms and LinkedIn feeds: “95% of generative AI pilots are failing.”

It’s a statistic from a recent MIT report, and it’s leading to a great deal of unease. Executives are questioning whether AI is fulfilling its promise to transform the productivity, effectiveness, and scale of our companies. Some are concluding that while AI may sound impressive, it doesn’t work in reality.

But there’s a problem with this narrative: It misinterprets the data. If we look closer, the data actually points to the blueprint for successfully adopting AI and experiencing impact.

The lesson of this study isn’t that “AI doesn’t work yet.” It’s that generic AI struggles where specialized AI excels:

MIT’s research indicated that while generic projects only succeed 5% of the time, projects with specialized vendors succeed 67% of the time.

For the 87% of sales leaders who report being under immense pressure to implement AI, misreading this report isn’t just a minor error; it’s a missed opportunity to capture market share.

Recently, I had a great conversation with Graham Miller, Gong’s Head of GTM AI Strategy & Enablement, that explored the reality of why projects fail so often, and what it takes to make them successful.

In our conversation, Graham used an analogy that helps clarify what’s behind the 95% failure rate.

Imagine you want to build a high-quality mahogany desk. You go to the hardware store and buy a hammer. It’s a powerful tool, but on its own it won’t get you there. To achieve that outcome, you’ll need a system of tools: Blueprints, saws, drills, and finishing equipment that all work together.

The same distinction applies to AI.

Generic large language models (LLMs) are built to be broadly capable across many tasks. They excel at generating content, summarizing information, and answering questions. But they operate mainly as prompt-based tools. They aren’t designed to carry context through workflows, across teams, stages, and systems.

In contrast, purpose-built vertical AI is designed around specific workflows and is embedded where work happens. It’s able to offer guidance in the moment and support coordinated execution across roles.

This is where so many AI projects break down. When revenue organizations invest in generic AI and expect their reps to be more productive, they’re handing out hammers and hoping for furniture. But success depends on shared workflows that align the entire team on an operating rhythm: Managing deals, reviewing pipeline, forecasting, coaching, and executing consistently. It’s not that LLMs aren’t effective. It’s that they’re most powerful when layered with purpose-built systems designed for execution within the workflows and context of the job to be done.

There’s a second reason many pilots fail: A lack of the right data to deeply understand the full context of your particular situation.

Remember that any AI is only as good as the data it can reason over.

That requirement applies to all AI systems, not just LLMs. Whether built on generic foundation models, CRM-based analytics, or other approaches that rely on partial or manually entered inputs, if the underlying context is shallow, the guidance will be generic.

If you ask an AI how to close a deal without understanding your buyer or your sales motion, the guidance will sound reasonable but vague. It’s like asking a stranger for relationship advice. They can offer platitudes or quote a self-help book, but they don’t know you, your history or what actually matters in your situation.

Your best friend does. They understand your backstory, the patterns, and the unspoken dynamics. This is what allows them to offer advice that’s tuned for you and your context.

For revenue teams, that context lives in day-to-day interactions: calls, emails, meetings, objections raised, deals stalled, and patterns that repeat over time. Yet many AI systems only see fragments of that reality. CRM-based systems depend on whatever reps manually log, which is typically sparse, biased, or inaccurate. Broad, horizontal AI can draw on vast general knowledge, but it still isn’t connected to how customers behave or how deals unfold.

The Gong Revenue AI Operating System (OS) is designed to close that gap. By automatically capturing and analyzing real interactions across calls, emails, meetings, and CRM activity, Gong develops a shared understanding of how deals truly progress. It knows where buyers disengage, how top performers handle objections, and which behaviors correlate with success.

That depth of context is what allows AI to move beyond generic suggestions and deliver guidance that is specifically targeted to what works for you and your business. It’s advice teams can trust because it’s grounded in their own reality.

Given MIT’s data showing how often AI efforts fail to make it into production, how can you ensure that your revenue team ends up among the organizations that see real, sustained impact?

Graham and I mapped out four pillars that differentiate failed experiments from successful AI-powered transformations. They are:

1. Purpose-built, not general

You can’t rely on a general-purpose solution to solve specific revenue problems. To get results, AI has to be purpose-built for revenue use cases and tuned to the realities of your business.

First, effective revenue AI is verticalized by design. It’s trained on sales and revenue-specific data and knows the behaviors that lead to success. Second, it learns from your own data over time. Beyond understanding revenue in general, it must adapt to what works in your organization.

This is where the adage, “Garbage in, garbage out” still applies. Effective AI is grounded in relevant, high-quality data and continuously refined based on real outcomes. This way, guidance improves as the system learns what success actually looks like for your team.

2. Embedded in workflows, not a sidecar

The most effective AI is embedded directly into your team’s revenue workflows, not tucked away in a separate destination. That might look like AI automatically highlighting which deals are drifting away from best practices and guiding reps toward the next best action. Or, it might involve defining top performers and delivering insights and coaching feedback that help scale what works. It can also continuously assess deal health and forecast risk, helping teams understand what to change to influence positive outcomes.

When AI is integrated this way, it doesn’t ask teams to change how they work. It strengthens the workflows they already rely on.

3. Outcomes-driven, not just “productivity”

Most failed pilots have vague goals like “make our reps faster,” or “give our reps better information about buyers.” Those are too surface-level to succeed, much less measure.

Instead, align your AI strategy to clear, tangible metrics that deliver ROI. Go after outcomes like measurable increases in deal size, win rates, and forecast accuracy so you can track the revenue impact. If you aren’t doing that, you’ve got little more than a new toy for your team.

4. An end-to-end operating system with no blind spots

For AI to fulfill its promise of being your operating system and guiding your team effectively with all of the insights and best actions to take, it needs to have a holistic understanding of your sales process. AI becomes effective and trustworthy for teams when it has the complete picture, deeply understands all aspects of the sales motion, and guides the organization with precision to the execution of a tailored process based on best practices.

The narrative that AI is failing is dangerous because it gives leaders the wrong conclusion: That they need to sit on the sidelines while their competitors adopt specialized, purpose-built platforms. Once your competition does that, they’re not just “trying out AI.” They’re getting real results, like higher close rates and faster ramp times, and moving past the pilot phase.

Don’t be fooled by the 95% stat everyone’s talking about, and don’t let it scare you. Instead, pay attention to the overlooked stat — the one that tells you that AI projects using specific solutions work 67% of the time.

The Gong Revenue AI OS is built to make a direct impact on your bottom line, ensuring your AI strategy is a revenue engine, not a tech project. It lets you bypass failed experiments and drive real, measurable business outcomes.