You're about to waste at least a year and millions of dollars on enterprise AI projects. Here's how to avoid it.

Here’s a new article on enterprise AI from me and a co-worker. As with most maturity models, it’s 2/3 prescriptive and 1/3 “here’s some ideas that might help.” A bit of map and territory.

With AI, we’re seeing a familiar anti-pattern, but this time flavor-injected with generative AI: the board charters a tiger team, gives them budget. The team hires consultants because their developers don’t know Python. Consultants identify AI use cases, write some Python apps, and many of the pilots show good business value.

AI-related spending is set to reach $371.6 billion by 2027 as businesses shift from experimentation to targeted AI investments. IDC, February, 2025.

Then the board says, “OK, now do that 300 more times.” But the team hasn’t built a sustainable strategy. They haven’t used that time and money to add new organizational capabilities. Your developers still aren’t skilled in AI. So now adding AI to those hundreds of other applications isn’t easy. And that’s also when they discover day two AI operations: you have to run this stuff! Now you’ve just got piles of Python without real in-house capability. Day two AI operations isn’t cheap, it’s often underestimated, if planned for at all.

More toil for developers

Developers already spend 60% of their time on non-application work.1 Add AI infrastructure, and, I’d guess, that’ll climb another 10–15%.

This is exactly what happened with Kubernetes. Based on Google-brand, excellent devrel’ing and keynotes, companies assumed it was easy - until developers were drowning in YAML. Perhaps we’ll hear an AI-centric quote like this in a few years:

Well, I don't know how many of you have built Kubernetes-based apps. But one of the key pieces of feedback that we get is that it's powerful. But it can be a little inscrutable for folks who haven't grown up with a distributed systems background. The initial experience, that 'wall of yaml,' as we like to say, when you configure your first application can be a little bit daunting. And, I'm sorry about that. We never really intended folks to interact directly with that subsystem. It's more or less developed a life of its own over time. Craig McLuckie, SpringOne 2021

AI is heading the same way. You’ll spend 12 months building your own platform. It’ll barely work - if at all - and cost ~$2 million in staffing. Developers won’t use it like you expected. There’s no ROI, it delivers a third of what you promised, and you’re not really sure how to run and manage it long-term. And, yet, it still costs a lot of money.

CIOs have found generative AI goals harder to attain than first anticipated. Technology leaders are blaming data challenges, technical debt and poor strategy or execution for slowing down progress. Nearly half of CIOs say AI has not yet met ROI expectations, according to Gartner research. Reporting from the Gartner Symposium, October, 2024.

That original team, now AI experts, will leave, either for to work at an AI tech company or to help do it over again at a new enterprise. You’ll be stuck with an unsupported, incomprehensible mess. No one else understands it. You’ve locked yourself into an opaque platform, wasted years, and landed back where you started - with a sprawl of shadow platforms and tech debt. But, don’t worry: the next CIO will launch a new initiative to fix it.

Two new challenges

With AI, you’ll have two more problems.

First, AI evolves monthly, sometimes weekly. New models, new techniques (“agentic AI”). You need to keep up, or you’ll be stuck on outdated tech, losing competitive advantage. The best way to handle this? Just like any other service you provide (e.g., databases). Centralize those AI services, then you can upgrade and certify once, enterprise-wide, and give developers a single, enterprise-y source for AI models. The alternative is to find all the AI models usage across your hundreds of applications and enterprise-y them up one by one. You’ll quickly fall behind - just look at the versions of software you’re currently running, I bet many of them are three, five versions behind…especially whatever Kubernetes stacks you built on your own.

Second, if you use the same models as everyone else, you’ll get the same answers. Asking the AI “How do I optimize my pipe-fitting supply chain?” will yeild the same response as your competitors. The real advantage is adding your own data. That’s hard work, needing serious budget and time. And once you figure it out, you’ll need to scale it across teams, which means centralizing AI access, just as we saw above in AI model usage and updating.

Enterprise AI needs a platform. And what we learned over the past decade is: building your own platform is a terrible idea. This is especially true in private cloud, which I reckon is where about 50% of the apps in the world run, probably much more in large organizations.

Instead, improve your existing stack. Don’t rip and replace it.

If you’re like most enterprises, you have a lot of Java developers using Spring. Use Spring AI as your AI framework to connect to models. The Spring AI developers have been working quickly over the past year to add in the baseline APIs you’d need and adapt to new innovations. For example, the agentic AI framework Model Context Protocol came out in November, and Spring AI is now the official Java implementation for MCP.

And if you’re like a lot of other larger organizations, you already have a strong platform in place, Cloud Foundry. You can add a full-on model service gateway to host or broker AI models. You can host those models your own if you’re freaked out about public cloud AI, use public cloud AI, or, more than likely, do both! Most importantly, you’ll be able to keep up with model innovations, providing them to all your developers as quickly as your enterprise red-tape will allow for.

Your platform team can manage AI services like any other - security, compliance, cost tracking. Since it serves OpenAI-compatible endpoints, you can even still use those Python apps, but now your operations team can secure and manage them instead of whatever consultant-built Python stack you got stuck with.

So, the plan: (1) det developers using Spring AI to start embedding AI into their apps. Work on integrating your own, secret/proprietary data to the AI app workflow. When they’re ready, add AI services to your platform so production deployment and management is seamless. You’ll have day two AI operations covered. And because it’s centralized in a platform, you can now roll it out to those 300 more apps the board asked for.

Then, you can execute a proven playbook: developers should be shipping AI-powered features every two weeks, experimenting and refining based on real usage. I outlined this approach in Monolithic Transformation, with more executive-level insights in Changing Mindset and The Business Bottleneck.

You know, as always - try to avoid repeating the anti-patterns of the past.

Wastebook

• Bathos, n. – The fine art of inspiring deep emotion, only to trip over a misplaced joke, a clumsy metaphor, or an unfortunate mention of flatulence.

• Alacrity, n. – The suspiciously eager willingness to do something, often confused with competence but more often a sign that the asker failed to mention how much work it actually involves. (Found in the The Economist, explained by the robot, previous as well.)

• Screwbernetes.

• “Forward Deployed Engineers.” SEs with commit access.

• “She glared at me from across the road and shooed me off because I couldn’t stop laughing.” Sting.

• “Tapestries around her neck.” RotL, #570.

• “Ms Adichie’s viral TED talk on feminism received an even more impressive accolade: Beyoncé sampled its lines.” The Economist World in Brief.

• “bantam self-confidence.” Tina Brown.

Relative to your interests

• Measuring Productivity: All Models Are Wrong, But Some Are Useful - “Measure Speed, Ease, and Quality Different facets of productivity warrant the use of different metrics. We typically think of productivity as a balance among speed, ease, and quality.”

• Buy an expensive “AI Gateway”? Thanks, we’ll just build and open-source it, says Bloomberg - Should we be calling these “gateways”? Probably better than “broker.” // Also crazy to think we’ll just be recycling the same old patterns. Maybe it’s because they work! But, you have to strip away all the marketing-talk from it. And: building your own commodity infrastructure is a bad idea. It’s great for the team that does it and then quits to found a startup or go work at a higher paying tech company, so, I guess: more like capitalists being accidental generous to employees.

• Four Marketing Principles That Redefine Markets from Klaviyo’s CMO - ‘“Creating fear never works, because in the immediate, you can probably prompt people to take action because they’re like, ‘Oh my! I must do something,’ but it leaves a negative perception in their mind.” “You don’t become a beloved brand over a period of time.”’

• Skype is dead. What happened? - Ode to Skype, and complaining about Microsoft having no imagination to evolve it. It’d be helpful to read a detailed analysis of how and why.

• What I learned from one month without social media - “Literally, it makes no difference what I abstain from, I will always find a way to procrastinate.”

• I’m Tired of Pretending Tech is Making the World Better - Try to avoid tools that require you to acquire new tools to use the first tools.

• America Voted For Chaos. The Markets Are Feeling the Punch. - Dumb disruption. Also: masters of the universe hubris.

• One “Bad Apple” Correct Interpretation - On the use and mis-use of “bad apples.”

• Why Skyscrapers Became Glass Boxes - ”Ultimately, it was economics (or at least perceived economics) that drove developers to embrace this style. Glass curtain wall buildings were cheaper to erect than their masonry predecessors, and they allowed developers to squeeze more rentable space from the same building footprint. Ornate, detailed exteriors were increasingly seen as something tenants didn’t particularly care about, making it harder to justify spending money on them. And once this style had taken hold, rational risk aversion encouraged developers and builders to keep using it.”

• Yoon Suin and Orientalism - an example of a “woke” look at a D&D setting.

• The Lights of My Life - Accent lighting and lamps used by one photographer.

• The difficulty level of children - “It also runs the other direction. If you have two kids, and one kid is away (with a grandparent), it feels like having zero kids.”

Read by the robot

I don’t read everything, sometimes I have the robot read it for me. Beware that the robot sometimes makes things up. Summaries are for entertainment purposes only.

AI firms raced to shrink large models into cheaper, faster alternatives, ensuring even small companies can now afford to hallucinate at scale. IBM expanded its AI strategy, embedding intelligence into products and unleashing 160,000 consultants to generate AI-powered assistants. AI, once set to replace lawyers, now helps them work faster—though not quite enough to stop them from citing fake cases in court.

The fragrance industry, once built on the power of scent, now thrives on TikTok, where influencers sell perfumes their followers will never have to smell - arguably the best possible scenario for everyone involved.

Is neoliberalism truly in decline? Despite its failures - rising inequality, social fragmentation - it remains the dominant economic framework with no clear replacement. Meanwhile: it doesn't matter if you saw a rabbit, a vase, or an old woman because a study debunked the idea that optical illusions reveal personality traits.

Conferences

Events I’ll either be speaking at or just attending.

VMUG NL, Den Bosch, March 12th, speaking. SREday London, March 27th to 28th, speaking. Monki Gras, London, March 27th to 28th, speaking. CF Day US, Palo Alto, CA, May 14th. NDC Oslo, May 21st to 23rd, speaking.

Discounts: 10% off SREDay London with the code LDN10.

Logoff

I’ve been running the above, uh, screed in my mind for a few weeks now. Perhaps I’ll use it as the basis for my VMUG Netherlands talk next week. It’s not exactly the topic of the talk, but good talks, as delivered, often are.