Page 3 of 3

Posted on

Machine Learning: Backbone of Enterprise AI

Infographic titled 'ML: The Real Backbone of Enterprise AI' illustrating applications of machine learning in logistics, manufacturing, and rail inspection. Key elements include a risk matrix, supply chain monitoring, camera inspection systems, and rail scan portals.

If you follow the headlines, you’d be forgiven for thinking AI is all about sentient chatbots and sci-fi robots.

But talk to people actually running operations in logistics, manufacturing, retail, or healthcare, and you hear a different story. The AI that quietly creates value day to day is far more grounded: machine learning models that predict, classify, and optimize.

In other words, while the marketing conversation is about “AI”, the reality on the ground is that most deployed AI is machine learning.

Following on from my previous post, this post looks at why machine learning (ML) should be treated as the backbone of a practical AI strategy, using DHL, manufacturing quality control, and automated train inspection as concrete examples – and how Dell helps organizations make this shift from AI hype to ML powered outcomes.

From AI Hype to ML Reality (In Plain English)

“AI” is a broad label. Under the hood, three families of machine learning do most of the real work:

  • Prediction – forecasting demand, failure, or risk from historical data.
  • Language understanding – extracting meaning from emails, documents, and news.
  • Computer vision – recognizing patterns and defects in images and video.

These systems don’t “think” like humans. They learn patterns from data and use those patterns to make better decisions at scale.

For most organizations, an “AI strategy” that actually delivers is really a plan to:

  1. Find the business decisions and processes where better predictions matter.
  2. Get the data into shape.
  3. Apply the right ML techniques to improve outcomes.

Let’s see how that looks in the real world.

DHL: How a Logistics Giant Turns AI into Everyday ML

As one of the world’s largest logistics providers, DHL runs a global network of warehouses, hubs, trucks, and aircraft. It’s been very open about using AI to “crack logistics” – and when you dig in, the heavy lifting is done by machine learning.

1. Spotting Supply Chain Risk Early (ML + Language Models)

DHL’s Supply Watch solution extends its Resilience360 risk management platform using machine learning and natural language processing to provide early warning of potential supply chain disruptions.

What it does:

  • Scans signals across 140+ risk categories, including financial, environmental, and social factors, plus crime, labour breaches, quality defects, shortages, and delays.
  • Uses language models to read unstructured text (news, reports, online sources).
  • Applies trained ML models to decide which signals suggest a real risk.

Why it matters:

  • DHL can warn customers before a disruption hits, instead of apologizing after the fact.
  • Operations teams get machine generated alerts instead of trying to manually watch everything.

You don’t need to be a logistics giant to see the pattern: find the messy data your people are drowning in, and use ML to turn noise into early warning signals.

2. Smarter Routing and Fleet Operations

Logistics is a world of thin margins. Fuel, capacity, and delays add up quickly.

DHL applies machine learning to:

  • Forecast volumes and capacity needs.
  • Optimize routes to reduce distance, time, and emissions.
  • Explore innovations like “truck platooning” (smart caravans of trucks managed by ML models) and more automated fleets.

In simple terms, ML learns from what has happened on your network, then proposes better ways to move things around in future.

3. Computer Vision for Parcel Inspection

A particularly visible DHL use case is AI powered computer vision for parcel inspection and dimensioning in its sorting hubs.

The challenge:

  • E commerce has exploded parcel volumes.
  • Manually checking labels, measuring boxes, and spotting damage no longer scales.

DHL’s ML first solution looks like this:

  1. Cameras over conveyor belts capture images of every parcel from multiple angles.
  2. Computer vision models (a branch of ML) are trained on millions of images to:
    • Read labels and routing codes, even if partially damaged.
    • Measure size and shape to ensure accurate billing and trailer/container space usage.
    • Detect dents, tears, and leaks, flagging suspect parcels for manual review.
    • Check that items are sorted to the right chute or bin.
  3. The system runs close to the conveyor (edge computing) for speed and learns over time from mistakes and human overrides.

The result: more parcels processed per hour, fewer errors, and fewer disappointed customers.

DHL’s own public materials on computer vision in logistics reinforce these benefits – automated measurement, damage detection, label checking, and predictive maintenance for assets, all powered by ML based vision systems.

Manufacturing Quality Control: Where AI Meets the Production Line

The same pattern shows up in factories. If you walk the floor of a modern plant, you’ll see AI in action in surprisingly down to earth ways.

DHL’s AI Driven Computer Vision trend report summarizes how manufacturers worldwide are using ML driven vision systems today.

1. Automated Visual Inspection

For years, product quality was checked by trained humans standing at the line. That approach:

  • Is tiring and inconsistent.
  • Struggles with very small or subtle defects.
  • Doesn’t scale well when production speeds up.

Computer vision and deep learning change that.

The report notes that one of the most important existing uses of computer vision in manufacturing is automated quality inspection during production.

Real world examples from DHL’s AI Driven Computer Vision report include:

  • Novacura (Sweden) – installs specialized cameras where human inspectors once stood and uses computer vision to automate inline inspection. Their system can detect cracks and faults in metal components, including defects too small or subtle for the human eye to reliably catch, helping manufacturers keep quality high without slowing the line down.
  • Photoneo – provides 3D machine vision scanners designed for tough industrial tasks, such as precisely inspecting heat exchangers. These scanners deliver higher accuracy and throughput than manual inspection and generate rich data that can be used to optimize upstream processes and reduce rework.

The net effect is simple to explain to non technical readers: cameras watch every product coming off the line, and ML models instantly decide “pass” or “pull this one out,” thousands of times a minute. That’s AI – but under the hood, it’s machine learning.

2. Predictive Maintenance from Images and Video

The same report also highlights predictive maintenance powered by ML.

The idea:

  • Cameras and sensors watch key machines and tools.
  • Historical visual data is used to build a picture of what “normal” looks like.
  • ML models then spot subtle changes – a vibration pattern, a color shift, a new noise – that suggest a problem is coming.

SparkCognition, for example, applies ML algorithms to historical visual data from factory equipment to:

  • Learn a baseline of healthy operation.
  • Analyze live video in real time, flagging deviations.
  • Enable maintenance teams to fix issues before they cause breakdowns, scrap, or safety incidents.

This is another great example for general readers: AI isn’t just “thinking robots” – it’s watchful models that help you take care of your machines before something goes wrong.

Automated Train Inspection: Duos Technologies

The same principles extend beyond factories and logistics to heavy infrastructure. Duos Technologies, Inc. — a Jacksonville-based company specialising in AI-enabled inspection of fast-moving rail vehicles — is a real-world example of ML driven computer vision at scale.

With over 1.6 million railcars running on nearly 200,000 miles of track in North America, manual inspection of every component simply doesn’t scale. Duos addresses this with its Railcar Inspection Portal (RIP®): an AI-enabled imaging system that captures high-resolution, 360-degree images of railcars — including the undercarriage — as they move at track speeds.

The ML layer is what turns images into intelligence. Professionally developed AI models automatically identify a wide variety of specific defects and anomalies, flagging them for inspectors before they become safety incidents. Optional add-ons extend this further — the apis® system handles pantograph inspection on electrified vehicles, while the obliquevue® system uses multiple cameras at precise angles to image the leading and trailing surfaces of undercarriages in detail.

The results mirror what we see in manufacturing quality control:

  • Higher safety – defects are caught earlier and more consistently than sporadic manual checks allow.
  • Less downtime – inspections happen at track speed rather than taking rolling stock out of service.
  • Better use of skilled staff – inspectors focus on resolving flagged issues rather than performing repetitive visual scans.

Just like on a factory production line, the “AI” here is a set of computer vision and ML models, wrapped in an inspection workflow and deployed on edge infrastructure close to the track — the same foundational pattern, applied to a different domain.

So Where Does Dell Fit In?

Most organizations aren’t short of ideas for AI. They’re short of foundations:

  • Getting data off lines, cameras, and sensors.
  • Storing and processing it efficiently.
  • Training and running ML models reliably – at the edge, in the data center, and in the cloud.
  • Doing all of that securely, at reasonable cost, and without needing a research lab.

That’s where Dell comes in.

1. Data and Infrastructure Built for ML, Not Just Storage

ML systems like the ones at DHL, in advanced factories, and in automated rail inspection rely on:

  • High volume image and sensor data.
  • Low latency processing near the action (e.g., in a warehouse, plant, or rail yard).
  • Scalable, resilient infrastructure to train, monitor, and update models.

Dell helps customers with:

  • Modern storage and data platforms to ingest and manage video, image, and telemetry data that ML models need.
  • High performance compute – from GPU enabled servers to optimized architectures – to train and run ML and computer vision workloads efficiently.
  • Edge solutions that bring compute and storage closer to where data is created (e.g., in a factory, distribution center, or beside the track), ideal for real time inspection, predictive maintenance, and on the spot decisioning where every millisecond counts.

In short: if you want to do what DHL is doing with parcel inspection, what manufacturers are doing with inline quality control, or what Duos Technologies is doing with automated railcar inspection, you need the right data and infrastructure foundations, not just an algorithm.

2. An Open, Partner Friendly AI and ML Ecosystem

No single vendor has all the answers. Dell’s approach is to provide open, flexible platforms that work with the ML tools and frameworks you already use or want to use – whether that’s computer vision for quality control, NLP for risk signals, or time series models for maintenance.

That means:

  • You can bring your own models, tools, or partners into a Dell powered environment.
  • You’re not locked into one cloud or one proprietary stack.
  • You can run workloads where it makes most sense – on prem, at the edge, or in a hybrid model.

For business leaders, the key point is: you can scale machine learning in a way that fits your existing operations and risk posture, instead of having to rip and replace everything.

3. Turning Use Cases into Roadmaps

Many organizations know they want “AI in logistics” or “AI in manufacturing quality” but struggle to decide where to start.

The most successful journeys tend to:

  1. Pick one or two high impact ML use cases – for example:
    • Visual inspection of a critical, high value product line.
    • Predictive maintenance on a bottleneck machine center.
    • Automated inspection of safety critical assets like railcars.
    • Anomaly detection on shipment or sensor data.
  2. Stand them up on robust infrastructure so they perform reliably, not just in a pilot.
  3. Use the learnings (data requirements, skills, governance) to build a repeatable pattern you can apply to the next set of use cases.

Dell works with customers at exactly this level: aligning infrastructure, data, and ML capabilities with the specific outcomes they care about – higher throughput, better quality, fewer disruptions, lower cost.

Conclusion: Call It an AI Strategy, Build It as an ML Strategy

Looking at DHL, manufacturing quality control, and Duos Technologies’ railcar inspection, a clear pattern emerges: the impressive AI stories are really machine learning stories in disguise. The value comes from well chosen, narrow problems — detect this defect, predict that failure, flag that risk.

If you’re shaping an AI strategy, it helps to be explicit:

  • You’re going to be a machine learning first organization.
  • You’ll focus on clear decisions and processes where better predictions matter, good data and solid infrastructure, and practical ML techniques — prediction, language, vision — that can be reused across problems.

Generative AI and large language models absolutely have a role to play — making insights easier to access, helping people interact with systems in natural language, and unlocking value in unstructured data. But for most organizations, the bigger near-term opportunity lies in applying proven ML techniques to the operational problems that cost real money every day.

With the right foundation — the kind Dell helps customers build — that backbone can be strong, scalable, and ready for whatever AI innovation comes next.

References

Posted on

Machine Learning: The Quiet Workhorse of AI

The geek side of my brain is excited. Last year I studied Machine Learning with Imperial College, this year I get to study it again as a module for my Masters.

ML is the coming together of math, statistics, and probability. The stuff we ever loved or hated at school. My math teacher did a good job.

ML is the quiet AI just doing its job. It’s not exciting and rarely makes the headlines. But ML is worth knowing because it covers up to 70% of all AI deployments.

There is a growing disconnect at the heart of enterprise AI — and most organisations are not talking about it openly.

On one side, boardrooms and budget committees are pouring money into Generative AI at a pace that has no precedent in enterprise software history. On the other side, the systems quietly running your fraud detection, your demand forecasting, your customer churn models, and your infrastructure monitoring are overwhelmingly powered by something far less glamorous: traditional Machine Learning.

This is not a post arguing that GenAI is overhyped. It is a post grounded in what the data actually shows about where AI is deployed, where investment is flowing, and — most importantly — what the gap between the two means for your organisation.

The Numbers Don’t Lie

When you look at enterprise AI production deployments today, the picture is clear. Traditional ML still accounts for roughly 65% of all production AI workloads. Generative AI sits at around 28%, with Agentic AI — the newest frontier — making up the remaining 7%.

Now look at where the money is going. GenAI commands approximately 68% of AI budgets. Traditional ML, despite doing the majority of the heavy lifting in production, receives only around 25% of investment. Agentic AI, still early-stage, takes the remaining slice.

[Insert Blueprint Visual: Production Deployments vs Budget Share]

That inversion — the gap between where AI actually runs and where the money flows — is the central tension every data leader needs to understand in 2025.

To put the investment scale in context: enterprises spent $37 billion on Generative AI in 2025, up from $11.5 billion in 2024. That is a 3.2× year-on-year increase — a growth rate that makes even the most aggressive SaaS expansions look conservative. (Source: Menlo Ventures, State of Generative AI in the Enterprise, 2025)

The Lab Is Full. The Production Floor Is Not.

Here is where the story gets uncomfortable.

Despite the frenzy of investment and experimentation, 89% of organisations are piloting or exploring GenAI workflows. That sounds like broad adoption — until you follow the funnel.

Only 37% of those organisations have any GenAI in genuine production. Just 15% have achieved enterprise-wide deployment. And the number that have crossed the finish line — generating real, measurable revenue impact from GenAI — sits at a striking 5%.

MIT research puts it bluntly: 95% of enterprise GenAI implementations fail to meet expectations. (Source: MIT NANDA, State of AI in Business, 2025)

This is not a technology failure. The models are capable. The failure is almost always in the gap between a compelling demo and a production-grade system that actually handles edge cases, integrates cleanly with existing data pipelines, respects governance requirements, and improves over time.

Why Traditional ML Is Still Winning in Production

Traditional ML has a 20-year head start on operationalisation. The tooling, the MLOps frameworks, the deployment patterns, the monitoring practices — these are mature. A gradient boosted model predicting server failure in a data centre does not hallucinate. A random forest classifying network anomalies does not need a prompt engineer.

There is also an uncomfortable skills reality sitting underneath these numbers. 50% of organisations report that their teams lack sufficient AI and ML expertise — and that figure has not changed year-on-year. (Source: Capgemini World Quality Report, 2025)

That skills gap has two effects. First, it slows GenAI from reaching production because teams lack the engineering depth to build reliable pipelines around LLMs. Second, it means that traditional ML — the systems already running — often goes unmaintained and under-optimised because attention and budget have shifted elsewhere.

The Agentic AI Signal Worth Watching

While the ML vs GenAI debate dominates the conversation, a quieter signal is building in the data. Agentic AI — systems where AI autonomously plans, decides, and acts across multi-step workflows — is moving faster than most predicted.

Deloitte forecast that 25% of enterprises using GenAI would deploy AI agents by end of 2025, growing to 50% by 2027. (Source: Deloitte TMT Predictions, 2025) Early deployments span customer service automation, supply chain management, and knowledge management workflows.

The important architectural point: Agentic AI is not a replacement for ML or GenAI. It is a consumption layer that sits on top of both — orchestrating models, tools, and data to complete complex tasks with minimal human intervention. Which means the 95% pilot failure rate for GenAI becomes an even more critical problem to solve before Agentic AI can scale. You cannot build a reliable agent on an unreliable foundation.

What This Means for Data Leaders

The data tells a clear story with three practical implications:

1. Protect your ML estate while you experiment with GenAI. The systems generating real business value today are mostly traditional ML. In the rush to build GenAI pilots, it is easy to starve these systems of the maintenance, retraining, and infrastructure investment they need. That is a risk hiding in plain sight.

2. The pilot-to-production gap is an engineering problem, not a model problem. Most GenAI initiatives stall not because the model is wrong, but because the organisation lacks the data pipelines, governance frameworks, and integration depth to operationalise it. Closing that gap requires treating GenAI deployment as a software engineering discipline — not a research experiment.

3. Budget and production reality need to rebalance. A 68% GenAI budget share driving 28% of production workloads is not a sustainable position long-term. Organisations that succeed will align investment with demonstrated production value — using pilot performance data, not vendor roadmaps, to drive allocation decisions.

The Bottom Line

The hype is real. The investment is real. But so is the gap.

Traditional ML is the quiet workhorse of enterprise AI — underinvested, under-discussed, and still doing most of the work. Generative AI is the loudest conversation in every boardroom — over-invested relative to its current production footprint, but with genuine transformative potential for organisations that can bridge the engineering gap.

The 5% who are generating real revenue from GenAI are not smarter than the other 95%. They are more disciplined. They treated production readiness as the goal from day one — not the afterthought.

That is the AI reality worth building toward.

Sources: Capgemini World Quality Report 2025 · Menlo Ventures State of Generative AI in the Enterprise 2025 · MIT NANDA State of AI in Business 2025 · Deloitte Global TMT Predictions 2025 · Deloitte State of AI in the Enterprise 2026

Posted on

In Praise of Boring, Everyday AI: A tale from Manchester Dell Heroes

Yesterday in Manchester, at a Dell Heroes event, I set out with a simple mission: talk about AI reality, not AI hype. In a room full of technical practitioners, I said: “Real AI is the everyday boring AI that just works in the background. You know, the silent AI.”

You could feel the room lean in. Heads nodded. People smiled. For a technical audience that spends their days building, fixing, and maintaining systems, that line hit home. This blog is really an extension of that moment.

The Manchester Moment

The event was packed with conversations about backup platforms, infrastructure, data, and AI. But as we talked through architectures and use cases, one theme kept surfacing: Most of the AI that really matters doesn’t look like a demo. It looks like data pipelines and information flows being processed to becoming knowledge – plumbing. So when I said, “Real AI is the everyday boring AI that just works in the background,” it wasn’t a throwaway line. It was a recognition of the work Data Engineers, Infrastructure Architects and Data Scientists do each day:

  • Architecting and maintaining data centre infrastrucutre
  • Building data pipelines that never get a slide in the keynote
  • Maintaining models that no one outside the team even knows exist
  • Integrating “intelligent” components into systems where the business only cares that it “just works”

The resonance in that room convinced me this story needed to be told more widely.

Headlines vs. Reality

We’re surrounded by AI headlines:

  • New models with more parameters
  • Viral screenshots of chatbots and image generators
  • Predictions that AI will either save or destroy everything

But that’s not what most technical people are dealing with day‑to‑day.

The reality looks more like this:

  • A model quietly reducing fraud by a couple of percentage points
  • An optimization algorithm saving a few milliseconds on a call that happens millions of times a day
  • A recommender system nudging customer behaviour just enough to move the needle
  • A forecasting model helping a business stock the right things at the right time

None of that is glamorous. All of it is real value.

And none of it works without the unglamorous effort behind the scenes.

The AI That Just Works (and Never Gets Thanked)

In Manchester, I asked the room to think about how many AI‑driven systems they rely on that nobody ever talks about:

  • Routing and logistics that adapt intelligently to changing demand
  • Monitoring and anomaly detection that flag issues before humans even notice
  • Capacity planning and autoscaling that keep services available under load
  • Security and fraud detection that run 24/7 in the background

These systems don’t win awards for “cool factor,” but the business would feel their absence in minutes.

Good AI in production is like good infrastructure:

  • When it’s working, you hardly think about it
  • When it fails, everyone suddenly realises just how important it was

The Silent Engineers Behind the Magic

There was another reason that line resonated in Manchester: the audience recognised themselves in it.

Behind every “boring” AI that quietly delivers value, there are people:

  • Data engineers wrestling with messy, fragmented, inconsistent data to build pipelines that are robust, reliable, and governed
  • Data analysts who understand the business context, define the questions worth answering, and interpret what the models are actually telling us
  • Machine learning engineers and MLOps teams wiring models into production, monitoring their behaviour, and keeping them healthy over time
  • Platform and infrastructure engineers ensuring there’s a secure, scalable foundation to run all of this on

None of these roles are “headline jobs.” You won’t see a press release about a beautifully designed data pipeline. But without them, the whole AI story collapses.

In the room in Manchester, I could see this land. For once, the narrative wasn’t “AI will replace you.” It was: AI depends on you.

Reliability Is the Real Innovation

We tend to celebrate AI for novelty:

  • “Look what this model can do!”
  • “Look at this amazing demo!”

But the practitioners in that Dell Heroes audience know something different:

Reliability is the real innovation.

  • A system that quietly does the right thing millions of times a day is more impressive than a one‑off demo
  • A model that improves a business metric by a small percentage, consistently, is more valuable than a viral moment
  • A solution that can be observed, governed, audited, and trusted is more important than one that simply looks clever

That mindset shift—from spectacle to stability—is what will determine which organisations actually win with AI.

The Discipline of Responsible Boredom

One thing that doesn’t get talked about enough is how often responsible AI work ends in the word “no.”

In Manchester, several conversations after my talk turned to the same themes:

  • “We could do this, but the data quality isn’t there yet.”
  • “We can’t deploy that; we don’t fully understand the bias.”
  • “It looks good in the lab, but it’s not robust enough for production.”

Saying no isn’t anti‑innovation; it’s the foundation of trust.

The “boring AI” that just works in the background only exists because teams were disciplined enough to:

  • Clean the data
  • Engineer the features
  • Monitor the models
  • Design for failure
  • Respect governance and policy

It’s not exciting. But it’s exactly what separates a toy from a system the business can depend on.

A Shout‑Out from Manchester to the Quiet Heroes

So this blog is, in many ways, a written version of that shout‑out I gave in Manchester. To everyone who saw themselves in that line: “Real AI is the everyday boring AI that just works in the background.” This is for you.

  • To the data engineers whose success is measured in the number of issues that never occur
  • To the data analysts who keep bringing the conversation back to “does this actually help the business?”
  • To the ML and MLOps teams who lose sleep over drift, latency, and edge cases so users don’t have to
  • To the platform and infra teams who keep the foundation secure, scalable, and compliant

You are the reason AI can move from slideware to reality.

From Hype to Habitat

If there was a single takeaway from that Dell Heroes session in Manchester, it’s this: AI will matter most when it stops being a headline and becomes silent habitat. It should be something we live and build inside:

  • Embedded in systems
  • Integrated into workflows
  • Governed, observed, and trusted
  • Invisible most of the time—until you look at the outcomes

The future of AI isn’t about constant “wow” moments. It’s about a steady, compounding “of course it works like that” reality. And that future belongs to the boring AI in the background—and to the quiet experts who make it possible.

From Manchester, here’s to the boring and the silent.
You’re the real heroes of AI reality.

Posted on

Getting AI Right First Time: From Hype to Durable Capability

Every week, we see a new headline promising that artificial intelligence will transform our business overnight. Yet when you strip away the hype, most successful AI journeys look surprisingly similar to any business systems development and deployment. AI is not different. AI is not magic. And AI for business is definitely not instant. Be prepared for structure and strategy.

In my experience, the organisations that really unlock value from AI tend to follow five strategic steps.

Step 1 – Gain Awareness: Confront Your AI Reality

The starting point isn’t technology – it’s conversation.

Bring business, IT and data teams into the same room, often with a trusted Partner in the mix, and ask a simple question: where could AI create meaningful value in our business? That might be reducing manual effort in operations, accelerating customer support, or spotting patterns in financial or sensor data that humans miss.

Those conversations do three important things:

  • Expose opportunities: You’ll quickly hear recurring pain points and ideas for automation or augmentation.
  • Surface constraints: Skills gaps, fragmented data, legacy systems and budget limitations all come into focus.
  • Build early buy-in: People feel consulted, not “done to”, which matters later when you need adoption.

From there, you can assess your skills, infrastructure and budget, and sketch a lightweight AI roadmap. At this stage, you’re aiming for awareness and alignment – not a 200-page strategy. A one-pager with 3–5 candidate use cases and some honest assumptions is often enough.

Step 2 – Experiment and Prove Value: Earn the Right to Scale

Once you have a shortlist of candidates, resist the urge to boil the ocean. Pick small, high-impact projects where you can clearly measure success.

The most effective teams:

  • Form a joint delivery team that blends business experts, internal IT/data talent and partner support to fill gaps.
  • Focus on use cases where they can move a real KPI – for example, cutting call handling time, reducing scrap, or improving forecast accuracy.
  • Define clear, measurable outcomes and time-box the work. If the pilot can’t be measured, it’s a science project, not a business initiative.

You should absolutely expect some failures here. That’s healthy. The goal isn’t perfection; it’s proving that AI can pay its way in your environment. A handful of well-designed experiments that deliver tangible ROI will do more to build credibility than any slide deck.

Step 3 – Make Scalable, Measurable Impact: Escape the Pilot Graveyard

After a few successful pilots, a new question appears: how do we scale this responsibly?

This is where many organisations stumble. They have one or two great proofs of concept, but they’re stuck in what I call the “pilot graveyard” – isolated successes that never become part of how the business actually works.

To move beyond that point, you need to:

  • Check your data foundations: Are the sources trustworthy, governed and repeatable? Can other teams reuse the same data without rebuilding everything from scratch?
  • Address explainability and trust: Can you show users and stakeholders how the model works, what its limits are, and how decisions are made?
  • Maintain executive sponsorship: Scaling AI often touches process, policy and sometimes jobs. You need senior backing to navigate that change.

At this stage, you should be thinking in terms of re-usable data products and pipelines, not one-off experiments. That might mean standard feature sets, modular components, and shared services rather than bespoke, project-by-project builds.

Step 4 – Scale Capabilities: Build Your AI Operating Model

Once you’ve proven that AI can work for you and you’ve escaped the pilot graveyard, it’s time to invest more seriously.

This typically involves three themes:

  1. Strengthen your data and AI platforms
    Whether you’re primarily on-premises, in the cloud, or running a hybrid strategy, you need platforms that can reliably support model training, deployment and monitoring at scale. Performance, security, cost efficiency and data locality all matter here.
  2. Formalise your AI strategy and governance
    This is where you move from a handful of projects to a portfolio of AI initiatives aligned to business priorities, with clear standards around ethics, compliance and risk.
  3. Democratise data and AI usage
    Push data literacy and AI awareness into the wider organisation, not just the data science team. Business–tech “translators” are especially valuable at this stage: people who can connect business problems to technical solutions, and vice versa.

You’ll also want some central coordination so you don’t end up with dozens of disconnected AI efforts that duplicate work and create inconsistent experiences for customers and employees.

Step 5 – Leverage Full Capabilities of Data & AI: Make AI Boring

By this point, you’ve moved beyond experiments. AI is starting to touch multiple processes and teams. The next step is consolidation.

For many customers, that means:

  • Setting up an AI Center of Excellence (CoE) to provide shared expertise, standards and best practices.
  • Putting robust MLOps and governance in place to handle lifecycle management: data pipelines, model versioning, deployment, monitoring, retraining and decommissioning.
  • Standardising how models are built and reused, with clear patterns and templates so teams don’t reinvent the wheel each time.

The real goal here is to make AI boring and repeatable – something that’s just part of the operating model, not a special project needing a steering committee every time. When AI looks more like a utility and less like an experiment, you know you’ve built a durable capability.

The Thread Through All Five Steps

Across all five steps, the pattern is the same:

  1. Start small – with honest conversations, pragmatic scoping and focused pilots.
  2. Prove value – in language the business cares about: revenue, cost, risk, speed, experience.
  3. Industrialise – by investing in platforms, governance and people so you can repeat success, not chase one-offs.

Partners play a crucial role in that journey. They bring hard-won experience of what works (and what doesn’t) across different industries and architectures. And platforms like Dell’s AI Data Platform provide the bedrock – scalable data infrastructure and data capabilities – that help customers move from AI “science projects” to AI as a durable, enterprise-wide capability. Data first, GPUs second is my advice.

Getting AI right first time isn’t about avoiding all mistakes. It’s about designing your journey so that each step builds confidence, creates value, and lays foundations for the next. When you do that, AI stops being a buzzword and starts being how you run your business.

Posted on

Why AI Projects Fail: The Importance of Architectural Discipline

Architecture at the Gate: Decisions, Drift, and the Cost of Getting it Wrong

I was speaking at an event yesterday. At a break, a guy from the audience asked, ‘Why do so many AI projects fail?’ I thought about this for a moment. And I got to thinking about AI technical debt on shiny new hardware. Then I thought, ‘This is nothing new, software and IT infrastructure projects have been failing since forever.’ But with AI, the problem is amplified. Failures are real loud because the promise of mysterious AI magic has failed.

AI is not magic. And AI should not be treated as a black box. It’s still data, software, algorithms and data centre hardware. What seems to be missing here is architecture best practice — that’s the magic.

Every deployment decision is a bet. The question isn’t whether risk exists — it always does — but whether you’ve looked at it honestly, documented what you found, and made a conscious choice rather than a convenient one.

The Go/No-Go Moment

When a large-scale deployment arrives at the gate with partial compliance coverage, the pressure to approve is real. Time-to-market feels urgent. Stakeholders are watching. The system mostly works. That’s precisely when architectural discipline matters most.

A go decision requires evidence, not confidence. Evidence that availability targets are met, that security controls are implemented and tested, that recovery procedures work under realistic conditions. Confidence without evidence isn’t readiness — it’s optimism wearing a hard hat.

A no-go is the right call when critical risks are unaddressed, when evidence is absent, or when the business impact of a production failure exceeds the cost of delay. That calculation is rarely comfortable, but it’s always honest.

The conditional go is where mature architectural judgement lives. It acknowledges that real systems rarely arrive at deployment gates perfect. Minor gaps, known residual risks, conditions with named owners and defined timelines — these are manageable. The conditional go converts vague concern into accountable commitment. It keeps delivery moving without pretending the gaps don’t exist.

What separates a sound conditional go from wishful thinking is specificity. Vague conditions — “security will be reviewed post-launch” — are not conditions. They are deferrals wearing governance clothing. Every condition needs an owner, a timeline, and a definition of done.

The Drift Problem

Approving a deployment is not the end of the architectural story. It may be the beginning of its most dangerous chapter.

Organisations change faster than their success criteria. SLOs defined at project inception drift out of alignment with business priorities that have shifted beneath them. KPIs that once reflected genuine outcomes become targets that teams optimise for rather than indicators of real performance. Architectural decisions made under one set of assumptions quietly become load-bearing under a completely different set.

The mechanisms that prevent this aren’t complicated, but they require discipline. Observability data should be reviewed against SLOs on a regular cadence — not to confirm things are fine, but to detect early signals that alignment is eroding. Decision registers should carry version history, because a decision that made sense in one context may be actively harmful in another. Stakeholder reviews shouldn’t wait for crises — scheduled, structured conversations about whether the architecture still serves the business are cheaper than the conversations that happen after it doesn’t.

The organisations that handle this well share one characteristic: they treat architectural knowledge as a living organisational asset rather than a set of diagrams produced at project start and forgotten at project end.

The Real Cost

So, my answer to why AI projects fail… The AI adoption wave of the last two years produced a generation of systems deployed without honest go/no-go governance, without versioned decisions, without success criteria anyone could measure. The technical debt was predictable. The security debt was avoidable. The governance debt is still compounding.

Architecture evaluation isn’t bureaucracy. It’s the mechanism by which organisations stay in control of the systems they build — and accountable for the ones they don’t.

ISO 42010 makes sense. Better to have Conway’s Law than Murphy’s Law.