Depending on which headlines you read, AI is either transforming everything or draining enterprise budgets with little to show for it. One viral stat last year — MIT’s much-misinterpreted “95% of GenAI pilots fail” — spooked boards and vendors alike. But the truth is less cinematic: Models are advancing far more quickly than enterprise operating disciplines, leaving companies scrambling to catch up.

Amid bold bets, failed AI rollouts, and a few secrets worth stealing, separating hype from real insight isn’t easy. Looking back over the past 12 months, here’s the download on what broke, what actually worked, and the plays worth building on as we move forward into 2026.

The top 3 potholes

1. Platform before problem

Too many enterprises have started with “We need more AI” instead of “We need to solve this problem.” That mindset inevitably ends with impressive demos and underwhelming business impact.

Deloitte’s enterprise AI research, for instance, found that many early generative AI initiatives lacked the basics: defined KPIs, clear ownership, and a direct line to business outcomes. At Fortune Brainstorm Tech 2025, Deloitte CTO Bill Briggs shared the firm’s observations: Teams launched a flurry of ideas, but none were anchored to a measurable P&L target. This meant no direction, no accountability, and ultimately no path to scale.

Without an owner, projects drifted between groups. With loose scopes, prototypes never matured. Fuzzy data requirements and governance that arrived too late meant that adoption quickly evaporated. The result: Pilots that looked great in a demo but were impossible to justify when budget discussions rolled around.

2. Bad pipes, bad outcomes

The cause of most failed enterprise AI rollouts traces back to one thing: Infrastructure that isn’t ready for AI. Data quality, middleware, access rules, connectors, and permissions all matter far more than enterprises expected. When the underlying “pipes” are weak, even the best models falter.

For example, when American Express rolled out the first version of its internal AI assistant, Travel Counselor Assist, it buckled almost immediately. At Fortune Brainstorm Tech, American Express CIO Ravi Radhakrishnan openly discussed the challenges. The core issue wasn’t the model; it was the company’s unstructured, inconsistent, and outdated data sitting beneath it. Without a clean substrate to reason over, the system returned slow and sometimes incorrect answers.

That painful moment forced Amex to pause and rebuild its data layer before doing anything else. This unglamorous move was the critical unlock. Today, thousands of Amex travel counselors across 19 markets use AI tools daily — a level of enterprise AI adoption made possible only because the company made the decision to reset the foundation. It had to go slow at the start in order to move quickly later.

AI is the great revealer of neglected infrastructure. If you’re still deferring data and middleware investment, AI will quickly and brutally expose that gap.

3. Communication as a side quest

People rarely reject AI because the tech is poor. They reject it when the rollout feels threatening or confusing. Duolingo learned this the hard way when its early “replace humans” messaging landed as a threat, not a vision.

Across organizations, I’ve seen the same pattern: When leaders talk about automation without clarifying roles, expectations, or incentives, employees brace for cuts. The result isn’t adoption — it’s hesitation and failed rollouts.

The top 3 plays that actually work

1. Metric first, model second

The most successful enterprise AI teams write their scope before they create a single line of code. That turns “Let’s jump on the AI bandwagon” into “Let’s reduce claim cycle time for Ops by 20% by March 31,” a tangible goal a VP can fund and a team can ship.

Making this shift helped Deloitte turn its scattered pilot portfolio into focused workstreams across code review, claims triage, KYC checks, and internal search use cases. To help scope projects, here’s a one-line scope template to put in place before tackling any build:

To move [metric] by [X%/$] for [team] by [date], we will [use case], measured from [baseline] to [target].

If you can’t fill in all the blanks, you’re still building a prototype, not a product.

2. Fix the pipes before the pilot

If we learned one thing about enterprise AI over the past year, it was that the system around the model mattered more than the model itself.

Amex’s turnaround is a master case study. Once the company rebuilt data pipelines, normalized sources, enforced accurate permissions, and elevated middleware from an IT cost center to a strategic enabler, adoption flipped. Travel Counselor Assist is now used daily by thousands of employees not because the model got dramatically better, but because the environment around it did.

3. Guardrails, not handcuffs

McKinsey’s 2025 enterprise AI report underscored a consistent theme: Top performers kept governance centralized while letting innovation live at the edges.

At Fortune Brainstorm Tech, Nike CTO Muge Erdirik-Dogan described how their centralized AI lab, which boasted high-quality talent and strong governance standards, sat too far from real-world problems. Once business-line teams owned their use cases (within clear, centralized guardrails), adoption accelerated.

Guardrails create clarity. Decentralization creates relevance. Scalable AI requires both.

Three bets to double down on for 2026

1. Small wins → Big ROI

In enterprise AI adoption, boring often beats brilliant. Across finance, insurance, and product organizations, the highest-ROI cases are simple and straightforward timesavers: reconciliation helpers, claim summarizers, research agents. Blackstone’s DocAI, for instance — a search and summarization assistant for internal documents — returned hours back to employees. Compounding efficiency outperformed moonshot ambition.

2. AI literacy > AI enthusiasm

Most enterprises underestimate a critical bottleneck: AI literacy. This is the basic fluency necessary to use AI tools responsibly — to interpret outputs, escalate errors, and redesign workflows around new capabilities.

I’ve seen organizations gain real traction by approaching AI literacy differently:

• They train managers first, not last. Managers are the translators between AI outputs and real operational decisions. Without that layer of understanding, adoption stalls.
• They establish lightweight AI competency baselines by role. Teams learn when to trust outputs, when to verify them, and when to override them, reducing both misuse and overreliance.
• They embed AI fluency into onboarding, the same way enterprises once institutionalized cybersecurity and compliance training.

These efforts aren’t about turning employees into data scientists. They’re about equipping them with the judgment, vocabulary, and mental models needed to work alongside AI effectively.

In 2026, AI literacy will be as fundamental as spreadsheet skills. Enterprises that invest early will see smoother rollouts, faster adoption, and fewer downstream failures. Those that treat it as optional will keep reliving last year’s pilot purgatory.

3. Every company needs an “AI supply chain”

One of 2025’s biggest shifts was the realization that AI isn’t a single system — it’s a supply chain. Data → models → tools → workflow → compliance → feedback loops.

The companies that advanced quickly treated this chain like a product:

• They mapped dependencies.
• They instrumented quality at every step.
• They built reusable components (prompts, agents, connectors, governance packs) rather than one-off experiments.
• They created shared services that business units could build on top of.

This approach turned AI from a series of disconnected pilots into a scalable platform, which reduced duplication, increased reliability, and accelerated subsequent deployments.

Enterprises that codify their “AI supply chain” will ship faster, safer, and cheaper than those still running bespoke builds.

2025’s central lesson: Don’t bolt a rocket to a horse cart

Last year, most enterprises tried to attach breakthrough AI onto legacy workflows. When the engines thrusted, the frame collapsed. The leaders who won — Deloitte, Blackstone, Nike — didn’t chase bigger rockets. They rebuilt the cart:

• They defined real problems.
• They fixed the pipes.
• They broadcasted clear guardrails.
• They empowered teams closest to the work.

In 2026, the AI rockets will only get more powerful. The question is whether the cart — your people, your processes, your infrastructure — is ready for the speed.