Why AI initiatives get stuck in pilot

AI initiatives often stall at pilot stage because organisations try to scale before the foundations are ready. In part one of this two-part series, we explore three common barriers: 

1. Data that is not AI-ready
2. Limited internal AI expertise
3. Integration debt that prevents AI from connecting with the wider business.

The takeaway for technology leaders is that moving from pilot to production requires sustained investment in data, capability and connected technology foundations.

Why are so many technology leaders struggling to move AI from pilot to production?

Many organisations have launched AI pilots, are testing use cases and are busy identifying opportunities to improve efficiency, customer experience and decision-making.

But despite growing investment in time, energy and budget, many organisations are struggling to move beyond proof-of-concept and are getting stuck in pilot purgatory.

Across industries, technology leaders are becoming increasingly frustrated with having powerful tools yet feeling powerless to make significant change.

There is more to say on this topic than one blog can do justice to, so we have split it into a two-part series. In both articles, we draw on NashTech survey data, real-world patterns and client experiences to explore the reasons AI initiatives stall in pilot or proof-of-concept phases, and what technology leaders can do to avoid those traps.

You can read part two here for the remaining reasons and practical next steps.

1. Your data isn’t ‘AI-ready’

AI is only as effective as the data that powers it.

When it comes to AI projects, neglected data is often the biggest stumbling block, which is why this challenge is first on our list.

In fact, roughly 70% of AI failures are tied directly to data and training challenges.

No matter how advanced the model, platform or codebase, an AI system built on inconsistent, incomplete or poorly governed data will not produce dependable results at scale.

Yet this is exactly where many organisations appear to be underinvesting.

When NashTech asked 1,000 technology leaders which layers of the evolving AI technology stack they were prioritising for investment, only 20% selected the data layer, including data lakes, vector databases and semantic or contextual layers. Instead, leaders seem to be prioritising the model layer, even though weak data foundations are among the most common reasons AI pilots fail to progress to production.

The consequences of this potential underinvestment are already visible in NashTech’s data:

• 36% of technology leaders say inconsistent data handling practices across development teams are a major security and compliance challenge.
• 34% say they have a lack of clarity around data ownership and responsibility.
• 42% report uncertainty around data governance and compliance as a challenge for agentic AI, while 40% say the same for generative AI.

This matters because the more autonomous and embedded AI becomes, the more important it is to know where the data comes from, who owns it, how it is used, and whether it can be trusted.

How to get your data ‘AI ready’

Before you start your AI initiative, look at your data foundation.

Audit your basic data tasks, including sorting, labelling, removing duplicates, annotating and refining before your modelling begins. Don’t assume the AI model will ‘figure it out’; if you do, you will discover that your model has simply learned your historical noise, bias and inconsistencies.

Understandably, some organisations want to skip this stage. Sorting out data foundations can be labour-intensive, costly and, frankly, less exciting than building new AI capabilities. And it is not a one-off exercise. Data management needs continuous attention, much like looking after a garden. A single data cleanse is like mowing the lawn once and expecting the garden to stay perfect all summer.

“If your data is chaotic, your AI will be chaotic. There’s no shortcut.”

Nigel Phillips, CDL

Because data management is not always visible to the Board, it can be difficult to demonstrate progress when you are asked what has been achieved. This is where a clearer budgeting conversation helps.

Organisations building or heavily customising their own AI models will naturally need to allocate more budget to model development, engineering and experimentation. But for those using commercial off-the-shelf LLMs, the bigger investment often needs to sit elsewhere: in data quality, governance, integration, monitoring and the workflows that make AI reliable in practice.

For the latter, smart budgeting in AI often starts with the data pipeline. If data is currently an area of weakness, you might think about putting the largest share of time, budget and effort into getting the data right.

Here’s a practical breakdown of how your smart AI budget could look:

• 5 – 75%: Data collection, cleaning and annotation
• 10 – 20%: Monitoring and training
• 10 – 15%: Model development and coding

Many organisations still apply an 80:20 mindset, putting most spend into software development and code, and only a small proportion into data work. For AI, that balance needs to shift.

“If your current AI budget shows 'data work' as a small line item and 'development' as the bulk, you are almost certainly under‑investing in the very thing that drives model quality.”

David Thorley, NashTech

And remember, this is not a one-off task. AI needs continuous monitoring, evaluation and maintenance to stay accurate and useful. Depending on the use case, that may involve retraining, updating prompts, improving retrieval sources, refining guardrails or changing the workflow. Without ongoing investment, models can drift and lose accuracy over time, undermining confidence and value.

If this is an area where you need support, NashTech’s data services and human‑in‑the‑loop (HITL) approach can help capture real‑world errors, correct them and feed those improvements back into the training data, so the model improves with every cycle rather than only at scheduled release windows.

2.  You lack internal AI expertise  

As demand for AI transformation grows, operational complexity is growing with it. Many teams are being asked to move faster, scale new capabilities and support more AI-enabled use cases, often without the specialist skills, infrastructure or operating model needed to keep pace.

This trap is evident in NashTech’s data:

• More than 43% of technology leaders report a lack of internal expertise in AI architecture and engineering
• 40% say they have limited access to scalable infrastructure or MLOps capabilities
• Only 28% are exploring partnerships with third-party providers to accelerate AI adoption

The issue is not just technical. It is organisational.

As AI begins to support more routine decisions and workflows, teams will increasingly shift towards exception handling, strategic planning and relationship management. That may sound like a major future change, but in many businesses, it is already happening informally, without the supporting structures, training or governance to make it sustainable.

AI capability also needs to extend beyond specialist technical teams. Board members need to understand the strategic implications, employees need the confidence and skills to use AI responsibly and procurement teams need to know how to assess the capabilities of external partners.

“Most companies want transformation, but very few have the operating model to sustain it.”

George Lynch, NashTech

That means capability-building cannot be treated as a side activity or a one-off training exercise. Organisations need role-specific upskilling, clear ownership for adoption, psychological safety for teams to question and learn from AI outputs, and the external support that builds internal confidence rather than creating a long-term dependency.

Small expert teams can make a pilot work in a controlled setting, but scaling AI requires many more people to understand how AI changes their role, where human judgement is still needed, how decisions are governed and who is accountable when the system behaves unexpectedly. Without that broader capability base, AI remains dependent on a few specialists and struggles to become part of everyday business operations. A scalable AI operating model needs clear ownership across the lifecycle: use case selection, data readiness, build, validation, deployment, monitoring, adoption and benefits tracking. Without this, the pilot team may prove the concept, but no one owns the transition into business-as-usual.

Integration is one of the biggest barriers to moving AI from pilot to production. A proof of concept can often work in isolation, but production AI needs to connect with the systems, processes and data that run the business.

That is where many organisations are struggling. NashTech’s data shows:

• Over 42% of technology leaders are having difficulty integrating AI with legacy systems
• Over 80% say their current technology stack is not fully equipped for AI

The challenge is not just that technology estates are old. It is that they are complex, fragmented and often only partially connected. When integration is weak, AI remains a side tool rather than part of the flow of work. Users may have to copy information between systems, decisions may not trigger downstream actions, and the organisation may struggle to measure whether the AI is improving the process at all.

Most enterprises, 89% according to recent NashTech Leaders Lab survey data, operate across a hybrid environment of:

• Legacy systems
• Multiple clouds
• SaaS platforms
• Home-grown applications

For AI to create value at scale, data needs to be accessible across the business to the people and systems that need it, while remaining protected by the right governance and controls.

Yet NashTech Leaders Lab data suggests this is still a major weakness, with 89% of technology leaders saying their integration is ad hoc and needs work, and that they rate their organisation’s integration maturity at just 2.78 out of 5.

Most organisations recognise integration as strategically important, but many still operate with data silos, fragmented processes and immature integration practices.

The most commonly cited hurdles to effective integration are:

In practical terms, that means AI teams are often trying to build intelligent systems on top of disconnected foundations.

And when integration is weak, the consequences show up quickly.

This is why some organisations are beginning to rethink the foundations of their technology strategy.

Here, integration is important; customers are more frequently turning to custom solutions rather than COTS due to configuration flexibility and integration, which in turn enables more AI overlay of these systems. NashTech’s data shows that 21% of technology leaders are more likely to use custom-built solutions enabled by AI tools, and a quarter are actively re-architecting their technology stack to support these capabilities.

“Connecting new AI tools with legacy systems is rarely simple. AI often requires more effort than initially planned, with incremental rollouts needed to identify issues early and build confidence. When teams lack organised knowledge transfer, problems take longer to resolve, project timelines extend and costs increase.”

Anon, Leaders Lab member

Before organisations can scale AI, they need to understand how work flows across teams, systems and decisions. Integration therefore needs to be high on the boardroom agenda.

For more practical guidance on reframing AI as a business transformation programme, building shared ownership and creating the conditions for measurable value, download 'Closing the gap between AI ambition and execution'.

What you can do differently

To move beyond pilot, leaders should ask:

• Is the data reliable, governed and accessible?
• Do we have clear ownership across business, technology, data and risk?
• Can the AI capability integrate into existing workflows?
• Have users been trained, supported and given confidence to adopt it?
• Do we have the people and capacity to build, manage and support it?

Getting the support you need

While 43% of technology leaders say they need custom development partners with AI capabilities, only 28% plan to explore partnerships with custom software development partners to accelerate AI adoption. Those that do may find themselves with a competitive advantage. 

Without the right expertise, operating model and delivery support, many organisations risk trying to scale AI with the same constraints that kept their initiatives in pilot in the first place.

To explore the remaining reasons AI initiatives get stuck in pilot, read part two of this series where we look at the next set of barriers and the practical steps leaders can take to move from experimentation to scalable business value.

Who is NashTech?

NashTech is a global IT consulting and technology services partner, helping organisations solve complex challenges through custom software development, AI and cloud solutions and digital transformation services that deliver real, measurable value, fast.

If you are considering how to get the best out of your AI strategy, we can help.