Why 80% of Enterprise AI Pilots Fail — And What Actually Works

Goldman Sachs published a research note on Q4 2025 S&P 500 earnings. The finding: no meaningful productivity gains from AI adoption. Billions spent. Nothing moved. Except software coding and customer service, which saw a 30% boost. Everything else sat on the shelf.

I read that and recognized the pattern immediately. I spent 18 months driving AI adoption at Marvell Technology, a semiconductor company with 7,000 employees. We went from single-digit AI tool usage to over 80% activation across Glean, Moveworks, GitHub Copilot, and M365 Copilot in under a year.

Most of the companies Goldman measured never got close. Here is why.

The Three Adoption Killers

1. Nobody owns it end-to-end

The most common setup: IT procures the tool, sends a training link, moves on. Maybe there is a Center of Excellence that meets monthly. Nobody in the room where real business decisions happen is accountable for whether AI actually changes how work gets done.

At Marvell, we embedded AI adoption into the CIO Chief of Staff function. Not IT support. Not a vendor relationship. An executive-level owner who sat in business reviews and could connect AI usage to actual outcomes. When the VP of Engineering asked "why should my team use Copilot," the answer came from someone who understood both the technology and the business context.

Without that ownership, AI tools become shelfware. They show up in procurement dashboards but not in workflows.

2. They measure the wrong thing

Goldman measured speed. Did tasks get faster? For most companies, the answer was no, because that is the wrong question.

When AI makes the first 80% of a task faster, people do not leave early. They expand scope. An analyst who used to spend four hours building a model now builds the model in one hour and spends three hours on scenario analysis that was never feasible before. The output is better, not faster.

If you measure cycle time and see no change, you conclude AI is not working. If you measure output quality and scope, you see a different story.

At Marvell, we tracked adoption rate (are people using it?), use case breadth (how many workflows?), and qualitative feedback (what are people doing differently?). Speed was one metric among many, not the headline number.

3. They skip the change management

This is the one that kills the most pilots. Companies treat AI deployment like a software rollout: install, configure, train, done. But AI adoption is a change management problem, not a technology problem.

People do not resist AI because they are Luddites. They resist it because their workflow is already optimized for the way things work today. Adding a new tool means temporarily getting worse at their job before they get better. That is a real cost, and if nobody acknowledges it, people quietly stop using the tool after the first week.

The framework that worked at Marvell:

Sequencing matters more than tool selection. We did not roll out four tools at once. We started with the one that had the lowest friction (Glean for search), proved value, then layered in the next. Each success built credibility for the next adoption push.

Champions over mandates. We identified early adopters in every business unit and gave them resources to experiment. Their results became the internal case studies that convinced the skeptics. Top-down mandates generate compliance. Peer examples generate adoption.

Measure adoption, not deployment. Deployment is IT's job. Adoption is a leadership job. We tracked weekly active users, not license counts. If a team had licenses but low usage, that was a signal to investigate, not a reason to buy more seats.

The Blueprint Is Already There

Goldman's finding that coding and customer service showed results is not a limitation. It is the blueprint. Both areas work because they are specific, measurable, and repetitive. The tasks are well-defined enough that AI can reliably improve them.

The mistake is thinking that other functions are fundamentally different. They are not. Finance, HR, legal, operations all have specific, measurable, repetitive tasks buried inside broader workflows. The companies that find those tasks and target AI at them specifically are the ones seeing results. The ones that deploy AI broadly and hope for the best are the ones Goldman measured.

The Aggregate Hides the Signal

There is one more thing Goldman's analysis misses. They are measuring the average company. The average company bought Copilot, emailed a training link, and never followed up. The companies that actually got results are statistical outliers. They are not in the aggregate. They are shipping.

If you are an executive reading Goldman's note and feeling validated that AI does not work, I would encourage you to ask a different question: Are you in the average, or are you one of the outliers? Because the gap between those two groups is widening every quarter.

What to Do About It

If your AI pilots are failing, or if you have not started because Goldman told you it does not work, here is where to start:

• •Assign an owner. Not IT. Not a vendor. Someone who understands both the technology and the business decisions it should inform.

• •Pick one workflow. Not "adopt AI across the organization." Find the most specific, measurable, repetitive task in your highest-value function and target that.

• •Measure output, not speed. Track what people are doing differently, not just whether they are doing it faster.

• •Sequence, do not flood. One tool, one function, one win. Then build on it.

• •Treat it as change management. Budget for adoption support the same way you budget for the software license.

The technology works. The question is whether your organization is set up to use it.

Not sure where your organization stands? Take the AI Readiness Scorecard at haagsman.ai/scorecard to find out in 5 minutes. If you run a smaller company, read the AI Quick-Start Guide for Small Business for a practical starting point.