Why Your Failed AI Initiative Is Actually An Organizational Problem

When talking with executives, questions have moved from “should we use AI?” (the vast majority of business leaders say yes) to "why isn't ours working?"

Seen this song and dance time and again: a vendor’s demos impressed someone who championed this internally and enough that a contract got signed with an AI pilot launched.

Aaaaaand then somewhere between "this is promising" and "this is operational," everything stalled.

I've spent more than three decades in enterprise data and technology, and I can tell you with confidence your technology is rarely the reason. Building AI agents today is commoditized because you can connect to a model, configure a workflow, and have something running in mere days.

But what determines if AI actually delivers value is everything that has to exist around the technology:

• Data architecture

• Accountability and ownership structure

• Observability and capability to ship updates

• Organizational readiness to treat it as a living capability (vs. a one-off tech go-live)

This is why I strongly believe AI failing is an organizational problem, and those leading (and hiring for) AI initiatives need to understand to full data-AI picture before making hiring decisions.

The numbers tell a story most haven't read yet

The research on AI failure rates is almost mind-numbing in how consistent different reports show the same challenges.

RAND Corporation found that over 80% of AI projects fail to deliver significant business value, more than double the failure rate of non-AI technology projects. IDC, in research conducted with Lenovo across hundreds of organizations, discovered 88% of AI proofs of concept never make it to production.

And S&P Global reported 42% of companies scrapped most of their AI initiatives in 2025, up 17%.

Why? Organizational readiness.

Gartner found 63% of organizations lack confidence in their AI data management practices, and predicted organizations would abandon 60% of AI projects that lack AI-ready data by 2026 – I can say that one was pretty on the mark.

Hiring patterns reveal another AI gap

One of the clearest signals of this problem shows up in how organizations are actually staffing their AI efforts.

We analyzed nearly 190,000 U.S. job postings and found companies posting 46% more AI and machine learning roles than data infrastructure positions while paying those AI specialists $15,680 more per year.

They’re hiring people who can build AI capabilities while leaving unfilled the roles responsible for the data quality, governance, and infrastructure that make AI viable - this is the enterprise equivalent of buying a Formula 1 car and fueling it with cooking oil!

Patterns are instructive; sectors showing the widest gaps - sales, legal, engineering – have significant competitive pressure and the fear of being left behind.

More disciplined sectors like finance and healthcare have historically real, immediate, and regulated consequences when data’s poor.

Finance is posting 240% more data infrastructure roles than AI roles. Healthcare? 164%. These industries have compliance requirements that force building data foundations before experimenting with new capabilities … and that discipline protects them from failure rates plaguing less-regulated sectors.

Separate analysis of U.S. Census and Bureau of Labor Statistics data reinforces a related gap: around one in five U.S. businesses now use AI, but only 1.3% have hired anyone trained to run it. When asked what changed after starting to use AI, the most common answer was effectively … nothing!

Two-thirds made no real operational changes, and most of those that did simply asked existing staff to take it on alongside their current responsibilities. A technology this transformative is adopted mostly by people doing it on the side; say hello to a structural talent gap hurting adoption.

A data problem underneath this AI problem

When AI initiatives fail, there's almost always a data architecture problem underneath the surface. For enterprises, that’s a massive layer to unravel.

Our research estimates the total annual cost of poor data quality to the U.S. economy at $617 billion, a.k.a. roughly 2% of the United States’ GDP. That figure was built by scaling Gartner's enterprise baseline across all 8.36 million U.S. businesses and 139.8 million employees, weighted by industry-specific data intensity.

AI doesn't just operate on dirty data, though … it amplifies it!

Clean, connected data produces genuine insight at machine speed. The fragmented, duplicated mess I see most organizations actually live with everyday produces confident-sounding nonsense at the same machine speed and often with no human in the loop to catch it.

This is why AI pilots that looked impressive in a controlled demo collapse in production, by the way … any demo uses clean, curated inputs while your live agent draws from your actual data environment with all its contradictions, gaps, missing fields, and competing definitions across platforms.

I like to ask this is a gut check: if I ask ten people in your business to define what a "customer" is, how many different answers would I get?

Usually the silence that follows tells me everything I need to know.

“But wait, a customer is someone who buys what you sell,” I hear you thinking. Obviously, but in practice it’s far different:

• Sales sees a customer at contract signature

• Finance sees a customer when an invoice is paid

• Service, the number of support cases

• IT, the integration(s) needed to support the client

• Product, a feedback loop helping determine what to build next

So while yes, a customer is exactly “who buys what we sell” these same customers can be viewed very differently by teams across your organization…

…and those definitions are baked into separate systems (from contract management to invoicing to order fulfillment) that have never been reconciled.

And any AI agent working across these same systems behaves unpredictably because the organizational foundation beneath it was never aligned.

Why AI pilots stall

When I talk to executives about AI failures, the conversation usually starts with the technology. Which model? Which platform? Which vendor?

This tells me everything I need to know.

Because three decades around data has borne out four distinct reasons why projects fail, and technology isn’t the reason (tech builds on top of your foundations, it doesn’t create them – no matter what a vendor’s promising while trying to sell you their tool).

In short, the four reasons are:

No defined outcome. The most common failure mode is a pilot that starts with "let's try AI" instead of "here's the specific decision we're trying to improve." If you can't name the decision you’d made based on what you get, you're not ready to automate this.

No clean data path. I’m not just talking about data quality, but the underlying architecture. Which system is the source of truth? How are objects defined across departments? How does data move between systems, and where does it break down? Unclear answers = chaos.

No cross-functional alignment. AI agents don't respect org charts. If an agent touches customers, pricing, contracts, or policy, it's speaking for the entire organization (not just who built it!). One team's perfectly logical AI output can contradict another team's definition of the same situation.

No governance or observability. Language models don't return the same answer twice. They adapt. Behavior shifts over time as usage patterns evolve and context changes. How do you know (from user feedback, for instance) when the agent’s drifting and how can you ship a fix quickly?

AI is not software. Please stop treating it that way.

AI agents behave far more like employees than applications – I’m not saying start treating AI as some living thing, but your approach would benefit from an HR-centric employee onboarding approach.

When you hire someone, you don't give them access to systems and hope for the best, right?

There’s orientation, culture alignment, system access, training, Q&A, feedback sessions and 1-on-1s, reviews, etc. When an employee struggles, you help.

AI agents require the same approach:

• Onboarding grounds the agent in the right data and information architecture.

• Training leads to refinement, feedback loops, and reinforcement over time.

• Performance management is observability, monitoring, and ongoing tuning.

• Governance defines what the agent can say, what it can’t, and what you do when something goes awry (and it will, this is normal)

You would never hire an employee and then stop checking their work! So why are many, in the scramble to out-adopt their competitors, doing the same with AI? That’s why 80% failure rates are making headlines.

What this means for people leaders

If you’re accountable for human capital investments, you should be alarmed.

Data’s clear that the AI talent shortage is more reflective of an architecture problem masquerading as a talent gap, and global organizations are spending heavily to hire AI specialists into environments where the underlying data foundation dooms them from day one.

No AI expert wants to spend countless weeks on manual data cleanup work that should have been handled by proper infrastructure investment. Any expert will know right away if they’re set to succeed or fail, and I suspect 1-2 year turnover will be rough for those who’ve rushed into AI-first projects built on 10+ years’ of bad data architecture.

More broadly, asking who "owns" data quality inside an organization is almost never answered clearly. Some have Chief Data Officers or centralized Ops functions; these are better than most. Many assume it’s an IT issue because data reflects business decisions and business language … in other words, it’s assumed to be technical.

AI has a way of making that ambiguity (technical vs. business) very expensive, very quickly.

Start with architecture before activation

Your data foundation has to come before the model and before you race ahead hiring for AI roles that aren’t set to succeed.

And you absolutely must have clarity around “what decisions will we make based on what this AI provides us?”

Start here: define who’s accountable for each domain of data – there’s eight of them, part of our 8-4-4 approach to data - and establish shared definitions across functions.

Run each category through the Four Rs test (what’s revealing, reliable, reusable, and relevant?) … this should take weeks, not months of years despite what some might claim. Any data that passes at least three Rs is worth keeping and building upon, anything else is ripe for deletion.

Then, build governance processes that treat data quality as an ongoing operational discipline. Sorry, but this only starts when an agent launches – so you must build accordingly.

Most importantly, I want you to take this mindset: AI isn’t a tech launch like a CRM or data lake. It’s a loud, noisy amplifier of how your business operates today, and if you want to be a leader then you have to do the foundational work first.

Otherwise, you’re racing towards becoming another statistic.

---

Andy Boettcher is the Chief Innovation Officer at DoubleTrack, a data and AI consulting firm that helps mid-market and enterprise companies by fixing their data architecture and AI foundations first. He has spent more than three decades working in enterprise data, technology, and AI readiness.