Skip to content
Fast Company Executive Board

The Fast Company Executive Board is a private, fee-based network of influential leaders, experts, executives, and entrepreneurs who share their insights with our audience.

Why your AI pilots aren’t scaling and what fixes them

Clean, usable data is the foundation of successful AI implementation.

[Photo: Getty Images]

Profile photo of Muamer Cisija

Co-Founder and President · Symphony

By now, you know the pattern. Big AI announcement. Executive buy-in, expensive consultants, and aggressive timelines for the proof-of-concept (POC). Six months later, there’s a decent-looking POC. Then nothing happens.

A year later? Still sitting there as a POC. Eighteen months later, someone finally says what everyone’s thinking—why hasn’t this scaled?

Here’s what nobody wants to say out loud: AI wasn’t the issue. The underlying data foundation was never there to begin with.

I’ve seen this scenario multiple times—healthcare organizations, manufacturing operations, professional services firms, etc. They all want the same stuff. GenAI chatbots that revolutionize customer service, predictive analytics that see around corners, and automated workflows that cut headcount.

Within three weeks, every single conversation ends up in the same place, “We need to rebuild our data platform first.”

WHEN AI COMES BEFORE DATA

The problem is straightforward—companies are trying to run sophisticated machine learning (ML) models on data infrastructure that can’t handle basic business intelligence.

We’re talking about manufacturing companies that don’t have a data warehouse. Healthcare providers with patient data scattered across 12+ different systems and no unified view of anything. Professional services firms that cannot answer simple questions like, “Did we make money on this project,” because billing lives in one tool, time tracking in another, and project management somewhere else.

Proofs of concept have their place and seem like the first step forward to implementing AI. They help you test what’s technically feasible without betting on the farm. You can experiment with small user groups, see what breaks, and build internal excitement about what is possible. The issue is running POCs without simultaneously fixing the underlying infrastructure to support them.

Smart companies are experimenting with AI use cases while centralizing their data at the same time by implementing governance and building proper warehouses. The POC teaches them what they need to build. The infrastructure work is what lets them ship it.

HOW THE ORDER GOT FLIPPED

The AI hype created warped incentives. Executives saw ChatGPT and immediately needed an AI strategy. Vendors started aggressively pushing their platforms. Everybody rushed straight to talking about models and automation, competitive advantage, and disruption.

  • Nobody wanted to slow down and discuss the boring parts:
  • Where does your data live?
  • Is it usable?
  • Who maintains it?
  • Who’s responsible when it breaks?
  • Does anyone understand what these fields mean?

Those are the details AI depends. While talking with a CTO at a PE-backed healthcare company, he shared, “We spent six months and probably $500,000 trying to build AI solutions before someone finally asked how many active patients we have across all our facilities. Turns out we couldn’t answer that question definitively. Couldn’t even get close.”

THE PARADOX

Here’s what’s paradoxical about the whole situation. AI needs solid data infrastructure to function properly. But once you have that infrastructure in place, AI becomes one of the fastest ways to improve it even further.

Modern tools can use AI to automatically classify your data, suggest metadata tags, and identify quality issues before they cascade. AI-powered data catalogs can map relationships between datasets that would take humans months to document manually. Machine learning models can detect anomalies in your pipelines and predict where quality problems are about to emerge.

There are three steps I always recommend to organizations preparing to implement AI.

1. Build infrastructure before trying to scale. If you’re planning to roll out a customer support bot or predictive maintenance model, your data needs to live somewhere modern and be usable first.

2. Add structure to data early in the process. Decide who owns what. Agree on basic quality expectations, what constitutes “good enough,” and set simple classification rules. This feels like unnecessary overhead when you’re trying to move fast and ship demos, but this is usually the exact moment where teams either graduate from POC or stay stuck in it forever.

3. Let AI do the tedious work. You don’t need humans manually tagging metadata, monitoring data quality, or documenting relationships anymore. The right tools can handle most of that automatically—faster and more consistently than any human could.

    Only after that foundation exists does it make sense to really push hard on use cases. With a solid base, teams can move incredibly quickly. Start with problems that genuinely matter to the business but aren’t overly complex. Show some real returns, build trust internally with skeptical stakeholders, and then double down hard on what’s working.

    The timing of this has changed more than most people realize. Building a legitimate data foundation used to take 18-24 months minimum if you wanted to do it right. With current tooling and AI-assisted workflows, companies are doing it in four or six months now. Sometimes faster if they’re aggressive about it.

    THE WINDOW WON’T LAST FOREVER

    The companies that successfully implement AI won’t be the ones that launched the flashiest POCs first or made the biggest announcements. They’ll be the ones that started experimenting early while simultaneously building the infrastructure to scale what they learned.

    There’s a window right now where doing this well creates a genuine competitive advantage. Eventually the whole market catches on and data-first becomes table stakes. The question is whether you’re ahead of that curve or behind it when it becomes standard.

    Muamer Cisija is the cofounder and president of Symphony.