After the AI pilot, who keeps it running?

AI implementation challenges and the pivotal question to solve them

With over 20 years of experience in managed services and global delivery operations, I see a pattern standing out in every artificial intelligence (AI) conversation: businesses are investing heavily in deployment and barely thinking about what happens next. AI isn't just a technology decision. It's an operating model decision, and most organizations aren't treating it that way.

Every enterprise leader is asking some version of the same question about AI: “What should we do with it?” There’s no reticence about the technology; full steam ahead with AI implementation in business. They’re evaluating use cases, running pilots, attending vendor demos, building innovation teams, but still running into AI implementation challenges. The energy is real, and the intent is genuine, but something is incomplete.

The question separating organizations making real progress from those burning through budget on experiments is, “Once we deploy AI, who keeps it running?”

AI lifecycle management is not a glamorous thing to ponder, and it won’t show up in keynote presentations or vendor pitch decks. But AI implementation challenges are hardly uncommon, and it’s this question that determines whether AI becomes an operational asset or an expensive science project.

The market data paints an unambiguous picture. ModelOp’s 2026 benchmark report shows that 67% of enterprises now have over 100 proposed AI use cases—but fewer than 25 are in production. MIT’s 2026 enterprise study puts it more starkly: 95% of generative AI pilots fail to reach production.¹ Not because the models don’t work, but rather, the data foundations, application environments, and governance frameworks weren’t built to support them. And for the small percentage that do make it to production, there’s a largely unanswered question: who operates and supports them from that point forward?

That gap between deployment and sustained operation is why AI projects fail.

AI implementation challenges based on AI model type

Not all AI runs the same way, so they can’t all be supported the same way. AI lifecycle management looks different depending on the models being used. This operational gap is so dangerous because enterprises treat AI implementation in business as a single thing. The reality is, they’re deploying three fundamentally different categories of technology, each with different operational demands.

Classic machine learning (ML) models

Classic ML models that power demand forecasting, fraud detection, predictive maintenance, and quality scoring have been in production at enterprises for years. But there’s a big difference between “in production” and “on autopilot.” These models run on structured data, require ongoing re-training as business conditions shift, and depend on continuous monitoring for model drift. That's an ongoing operational responsibility, not a one-time data science project.

Generative AI

Generative AI introduces a different operational profile entirely. Large language models that specialize in content generation, summarization, and unstructured data processing bring entirely different infrastructure demands and present different AI implementation challenges. There’s token cost management and retrieval-augmented generation (RAG) pipeline tuning to prompt version control and production guardrails to prevent AI hallucinations and bias. The support model looks nothing like classic ML, even though both are AI.

Agentic AI

The autonomous agents executing multi-step workflows, interacting with external tools, and making decisions with minimal human oversight add yet another layer. These systems require orchestration frameworks, expanded security monitoring, and governance models that most organizations haven't designed yet. This is especially important to consider as agents access more systems and data than any single application ever would. Autonomous systems need governance, monitoring, and operational support at a scale most IT organizations haven’t even begun to plan for.

And across all three categories, there’s a shared operational burden that’s easy to underestimate: Machine Learning Operations (MLOps). Model versioning, A/B testing in production, rollback procedures, compliance logging, and bias monitoring all require sustained attention. Responsible AI also requires consistent attention to ensure the models remain explainable, fair, and aligned with evolving regulatory requirements, like the EU AI Act.²

These aren’t one-time setup tasks; they’re ongoing AI lifecycle management responsibilities that grow as the AI portfolio expands. Each category demands different skills, infrastructure, monitoring, and operational support models and risk profiles. If no one has defined who owns that operational layer, this lack of cohesion will create AI implementation challenges that organizations will struggle to overcome.

In AI lifecycle management, deployment is only the beginning

A pattern is becoming clear across AI implementation in business logistics, retail, and services. AI deployment is underway, but innovation or growth is stalling.

For instance, a global food services company has AI in production today (computer vision and sensor-based analytics evaluating trainee performance in real time). The technology works, but they soon realized that their current support arrangement for AI will create operational silos without a clear path to scale. Instead, they’re asking their primary managed services provider to take on the task. With support managed under their services provider, the AI capability can be easily integrated with other applications that their users already depend on.

This also opens the door to something bigger: a 360-degree performance evaluation framework, following trainees from the classroom to on-the-job performance, once they graduate from the program. That’s the kind of business value that only emerges when AI model deployment isn’t siloed from the rest of the application ecosystem.

This isn’t an isolated case. Similarly, a major logistics company invested billions in AI-powered network optimization, facility automation, and predictive routing. But their own strategic analysis acknowledges that legacy technology and data silos must be dismantled before AI can deliver on its full promise. In another example, a regional grocery retailer is layering AI into inventory management, in-store analytics, and personalized digital experiences, but the underlying IT infrastructure varies wildly across hundreds of locations. This infrastructure variation presents a major roadblock to widespread adoption.

In all of these cases, the common thread is deployment happened or is underway, but the operational backbone to AI lifecycle management is either missing or improvised.

It’s not just the AI, it’s the entire workflow

Most discussions about AI support and operations focus narrowly on the model monitoring, re-training, and managing infrastructure, but AI doesn’t run in a vacuum. An AI agent sits inside a business workflow that touches enterprise resource planning (ERP), customer relationship management (CRM), supply chain platforms, databases, application programming interfaces (APIs), and custom applications. The agent is one component (often the most visible one), but the surrounding application ecosystem is what makes it work.

When something goes wrong, the issue might not be the AI model at all. It could be an upstream data feed from a legacy application that changed format, or a broken API integration with a connected system. Diagnosing and resolving these issues requires someone who sees beyond the AI layer to include every application it touches and understands how they interact.

Any AI implementation challenges are compounded when legacy architectures become the bottleneck. Technical debt (that was manageable when applications operated independently) becomes a critical point of failure when an AI agent depends on those same systems for inputs, outputs, and decision logic. This fundamentally changes what AI model deployment and support mean.

To avoid common AI implementation challenges, it’s imperative to understand you’re maintaining an interconnected suite of applications, where AI is embedded in the workflow, alongside enterprise systems that were built in different eras, on different platforms, with different architectures. The talent and expertise required to support this AI lifecycle management doesn’t exist in any single team today. No matter what the ownership for implementation looked like, once AI is embedded in business workflows that span across enterprise systems, the ownership model must change.

These changes look like:

• Internal IT needs to own the infrastructure and application landscape that the AI depends on
• Business owners own the workflow outcomes and escalation paths
• The data team owns quality, lineage, and governance of data feeding the models
• Risk and compliance own the regulatory and ethical guardrails
• A managed services partner (that already understands the application ecosystem) provides the operational backbone and day-to-day continuity that holds it all together

This approach requires building a shared operating model, where every stakeholder knows their role; the handoffs are defined, and someone has accountability for the end-to-end workflow in production.

Improper sequencing is the root of AI implementation challenges

If there’s one takeaway from what's happening across logistics, retail, utility services, and corporate services, it’s this: AI model deployment is about building the operational foundation in the right order.

You can’t run reliable AI on fragmented data. You can’t scale AI on legacy applications that weren’t designed for integration. You can’t secure AI if you haven’t addressed the expanded attack surface that comes with agents interacting across systems. And you can’t sustain AI without an operational model that accounts for the full lifecycle, not just the pilot. Without the proper sequencing, AI implementation challenges will stall innovation and create a ripple effect of inefficiency.

The proper sequencing looks like:

1. Modernize the application environment to create a stable, integrable foundation.
2. Establish data governance so AI models have clean, trustworthy inputs.
3. Harden the cybersecurity posture to protect an environment where AI is accessing more data and making more decisions.
4. Build the shared operational framework across internal IT, business operations, data, risk, and operational partners that will support the full suite of applications in production.

Organizations that treat AI lifecycle management as a design requirement from day one consistently achieve lower total cost of ownership than those that improvise it after deployment. This makes sense because retrofitting governance, support models, and integration frameworks after the fact is always more expensive than building them in.

This requires alignment across numerous parts of the business. That alignment isn’t as exciting as a keynote about agentic AI transforming the enterprise, but it’s a necessary component of what makes the transformation stick.

AI implementation challenges can be solved with the right approach

The enterprises that win with AI implementation in business over the next 2-3 years will be the ones that built the shared operational muscle to sustain AI in production. When internal IT, business owners, data teams, risk and compliance, and managed services partners each own their part of keeping the models accurate, integrations stable, workflows running, users supported, and the risks managed, innovation can truly unfold.

The question isn’t whether your organization should use AI; that’s already decided. The question is whether you’ve defined who owns what happens after the demo is over—and whether every stakeholder at the table knows the part they must play in a successful AI adoption.

To learn more about how your organization can harness the power of data and AI, fill out the form below.