I never set out to spend a decade in AI; I just wanted to get stock options and an IPO so I could retire. Ten years later, I’m still here, selling AI to companies that are trying to make sense of it all. In those ten years, I’ve worn a lot of hats: sales engineer, consultant, go-to-market guy, and lately, the person who helps companies make sense of the chaos. I’ve worked at startups that got swallowed up, sold products that vanished, and watched tech I once championed get left behind overnight.

This isn’t a victory lap for me. It’s a confession, a mea culpa, ten years’ worth.

2016: The World Was Going to Be Automated

Back then, I was a sales engineer at RapidMiner. My job was to build and demo no-code AI pipelines for customers who wanted to predict churn or squeeze more out of their marketing. It felt like magic at the time: Drag a connector, drop a model, and suddenly, anyone could “do machine learning.” The demos were slick. Real life? Not so much. No one trusted no-code pipelines.

What eclipsed the no-code pipelines was AutoML, and DataRobot was the dominant startup. The pitch was simple: you don’t need a PhD, just point the tool at your data and let it work its magic. Python was taking off. Scikit-learn was the new standard. Jupyter notebooks were suddenly on every laptop. And then AlphaGo beat Lee Sedol. Suddenly, every executive wanted to talk about AI.

The hype was real back then, but this was the dirty little secret they wanted to keep: most companies buying these tools didn’t have clean data. They didn’t have a real business question; they didn’t have anyone to keep the thing running after we left.

We sold them a car, but nobody checked if there was a road to drive it on.

2017: Going It Alone

In mid-2017, I left RapidMiner and struck out on my own. I started a Data Science consultancy, and it felt like freedom for a while, but then it became hard - very hard. It was hard to make a living as a single data scientist looking for contract work, and I realized that it’s much easier to work for a startup than be one yourself. Startups with venture capital funding offered a steadier paycheck than chasing down my dwindling set of leads.

Then something happened that I didn’t catch. Facebook released PyTorch.

A group of researchers at Google published a paper called “Attention Is All You Need” that, at the time, was mostly discussed in academic circles. Looking back, that paper was a grenade with a very long fuse (Vaswani et al., 2017). It eventually changed the AI landscape for good.

2018: The AutoML Wars Heat Up

In late 2018, I joined H2O.ai as a sales engineer. They had built a competing AutoML product, and we went head-to-head against DataRobot. AutoML dominated conversations, and the hyperscalers like Google, Amazon, and Microsoft all raced to automate model building and become industry leaders.

While this was going on, the money pit called “Big Data” got the defibrillator turned on and shocked it back to some sort of life. Big Data was still the buzzword, but it felt like a walking zombie, and Spark was becoming a solution in search of a problem. Most customers didn’t have petabytes. They had messy, medium-sized data that no AutoML tool could fix. Still, we kept talking about scale as if it were the real issue.

Google released BERT that October, a language model trained on massive amounts of text that could understand context in ways previous models couldn’t. It quietly changed search forever. (Devlin et al., 2018) Most of us in the AutoML world didn’t fully register what that meant yet.

MLflow launched, and suddenly, the startup that made Spark sexy again, Databricks, took off. The solution looking for a problem was suddenly replaced with a solution looking to put models into production, problem or not (MLflow, 2018).

2019: The Year MLOps Became a Job Title

2019 was a blurry year for me. I spent all my time in airports, hotel bars, building pilots that rarely reached production. It felt like such a waste of time and money.

The AI industry shifted, and new buzzwords emerged: MLOps. Responsible AI. Explainable AI. Each of them pointed to a real problem, but each also became its own category of software to buy, consultant to hire, and conference to attend. Everything was always piecemeal, not comprehensive, and came with a disclaimer: something else vital was “sold separately.”

This was the year that I spent most of my time on explainability. SHAP, LIME, model cards - companies were worried about bias, especially after the Amazon recruiting mess. But “Responsible AI” usually meant buying a tool that spit out a report. It rarely changed how decisions got made because no one understood how it could be used to transform the organization.

While I was screwing around with explainability, OpenAI released GPT-2 with a PR campaign that framed it as dangerous to release. It generated reasonably coherent text, and everyone in the field was impressed. It felt like science fiction to me, and it was actually just a preview of what was to come (Whittaker, 2019).

Oh, I forgot. Kubernetes was suddenly in every architecture diagram. Most data scientists I knew had no idea what it actually did. (Kubernetes Established as the De Facto ‘Operating System’ for AI as Production Use Hits 82% in 2025 CNCF Annual Cloud Native Survey, 2026)

Then came 2020, a year that brought massive disruption: the world stopped, but AI did not.

2020 was rough, both personally and professionally. When COVID hit, flights stopped, events vanished, and the in-person sales game I’d built my career on turned into Zoom calls with blurry backgrounds overnight.

Once the reality set in that the world would be “closed” for a while, something shifted. The companies dragging their feet on digital transformation suddenly had no choice. Data science teams that used to beg for headcount were now fighting for leadership’s attention. It felt like a strange kind of vindication.

While everything was starting, their virtual meetings with “Can you hear me?” OpenAI released GPT-3 in June. I didn’t get around to it till September, and I was genuinely surprised how well it worked. It’s 100x better than GPT-2, and it made me rethink everything I thought I knew about this industry. The model could write, translate, code, and sustain conversations for a few minutes. New large language models were being created and dropped into Hugging Face at a rate I had never witnessed.

The era of the foundation model had arrived while we were still pitching AutoML as nothing had changed.

2021: The Money Found AI Before AI Found a Purpose

The money faucet was opened wide in 2021, and venture capital flooded AI startups in ways that would have seemed impossible just a few years before. GitHub Copilot landed and instantly changed how I thought about writing code. DALL-E started making images from text (GitHub Copilot, 2021). Meanwhile, the NFT and crypto circus was so loud that it drowned out real AI conversations. Can your AI be on the Blockchain? Those were the questions I was fielding.

MLOps tools multiplied faster than anyone could keep up. Weights & Biases, Neptune, Comet, Determined AI, the ecosystem was crowded and confusing. Every company I talked to had a different stack; none of them were happy because it was all piecemeal, once again.

Then I started noticing something new: companies began adding “AI” to job titles without changing the jobs. Chief AI Officers popped up at places with no real data strategy. It felt like the Chief Digital Officer fad all over again, just louder and powered by AI.

2022: The Image Changed Everything

Before November 2022, generative images weren’t even a thing, and then Stable Diffusion dropped as open source. DALL-E 2. Midjourney. Suddenly, AI wasn’t just a tool for analytics or developer productivity. It was making things. Creative things. Things that looked like art pissed off all the artists.

People saw what these models could do, and it hit hard. Designers panicked, stock photo sites panicked, and ethicists sounded alarms about training data. And suddenly, friends who’d never cared about AI before started asking me: Is this going to take my job?

Then, on November 30, ChatGPT launched. If GPT-3 was a grenade with a long fuse, ChatGPT was the explosion in a crowded room. A hundred million users in two months (Malik, 2023). Executives who’d nodded through AI briefings for years were suddenly texting me at midnight, asking what it meant for their business.

It meant this: everything you’d done in AI up to now was about to be measured against a new standard, and most of it wouldn’t make the cut. I saw the specter of layoffs hanging over many good people.

2023: The “Pivot”

I hate the word ‘pivot,’ but it appears to have happened. Everyone pivoted to AI, and every vendor, every startup, every PowerPoint deck was about LLMs, RAG, vector databases, LangChain, and agents. It’s like a whole new vocabulary showed up overnight.

OpenAI was the industry’s dominant force. It launched GPT-4 in March. Right after, Claude showed up, and Meta’s Llama models enabled local deployment. The AutoML companies that had raised hundreds of millions were scrambling to explain how they fit into this new world. Some tried to pivot by slapping “AI” on their marketing, but it didn’t change anything. The smarter ones just started looking for a buyer.

The startups I worked for, believed in, or sold for became footnotes in history. RapidMiner was sold, then sold again to Siemens. H2O pivoted to LLMs and generative AI, dumping its Driverless AI. DataRobot rebranded and is still hangin’ in there. Tools that were expected to democratize data science were now outpaced by something even larger. The simple, easy button, be all AI - ChatGPT.

It was humbling, scary, and a bit fascinating at the same time.

2024: The Hangover

Enterprise AI finally arrived in 2024. But it brought a new problem: everyone was doing AI, but almost nobody could say what they were getting out of it.

Claude 3, GPT-4o, Gemini 1.5, Llama 3 — the model race sped up so fast that today’s breakthrough was old news by next quarter (GPT-4o, 2024). AI coding tools were everywhere. Cursor, Copilot, and a dozen others changed how we write software faster than anything since Git.

And yet, in most of the companies I spoke with, the real problem wasn’t the model. It was still the data, the pipelines, and the change management. Sound familiar? It should. It was the same problem we had in 2016, but now we have fancier tools to “not solve” them with.

Then the EU AI Act has passed. (AI Act enters into force, 2024) AI safety became a mainstream boardroom conversation. The “move fast and break things” energy of 2022-2023 was colliding with regulation, liability, and the first real post-mortems on failed AI deployments.

2025: The Ground Shifted Again

In January, DeepSeek R1 landed like a cold shower. A Chinese lab built a model almost as good as GPT-4o, but cheaper and faster. The idea that American tech giants had an unbeatable lead vanished overnight. The new destroyer of worlds took stage: Agentic AI, aka AI Agents.

Stocks moved. Narratives fell apart.

The AI race was global, and nobody could ignore it anymore (Atalan, 2025).

Agentic AI moved from an experiment to an expectation. MCP, Anthropic’s Model Context Protocol, turned into infrastructure. AI wasn’t just answering questions anymore. It was taking action, running workflows, and connecting systems. The question wasn’t if AI could do the job. It was how much to hand over, and what to keep for ourselves. I had spent ten years in this field, and I was feeling strange; everything felt both familiar and completely foreign.

2026: Where We Are Now

The agents are real. The hype is still real. The gap between them? That’s where the interesting work lives. Now I’m at World Wide Technology, helping companies navigate the most confusing tech landscape I’ve ever seen. And what I keep finding is the same thing I found in 2016: the technology isn’t the problem.

The real problem? It’s the question nobody asked before buying the platform.

What We Actually Got Wrong

Let me be honest about what this industry got wrong. What I got wrong.

We sold tools as answers. AutoML, MLOps, LLM APIs — every wave promised the tech would fix everything. It never did. The real problem was always organizational: unclear ownership, messy data, misaligned incentives, and resistance to change. No model fixes that.

We promised democratization before we delivered it. “Anyone can build AI now” — that was the pitch in 2016, 2019, 2022, 2024. Every time, it was half true and mostly premature. The gap between a slick demo and a real, working system was always huge. We kept understating it.

We chased scale before we had clarity. Big Data was sold to companies without big data problems. Kubernetes got deployed by teams that didn’t need it. (Ponseel, 2021) Now LLMs are landing in organizations that still can’t agree on a data dictionary. The urge to build for tomorrow while yesterday’s basics are broken hasn’t gone away.

Responsible AI stayed on the slides. Every company I talked to had a Responsible AI initiative. Almost none changed how they made decisions. The gap between the ethics deck and the real deployment? That’s where the risk lived. Nobody looked there.

We confused the speed of the frontier with the speed of adoption. The models improved rapidly. Enterprise adoption didn’t. Most companies are still running on the same old data, the same old governance, the same old risk-averse culture. The gap between what AI can do and what most companies have actually deployed is bigger now than it was five years ago.

The Thing That Didn’t Go Wrong

Here’s the truth: the technology worked.

Not always the way we promised. Not always on the timeline we sold. Not always with the ROI we projected. But real, meaningful capability showed up in model quality, accessibility, and the range of problems it could address.

Ten years ago, I was showing people how to drag nodes in a GUI to build a decision tree. Now, I’m helping companies connect language models to real business processes in ways that would have sounded like science fiction in 2016.

We got a lot wrong, but the direction? That wasn’t one of them.

References

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L. & Polosukhin, I. (2017). Attention Is All You Need. NIPS 2017. https://doi.org/10.48550/arXiv.1706.03762

Devlin, J., Chang, M., Lee, K. & Toutanova, K. (2018). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv preprint arXiv:1810.04805. https://doi.org/10.48550/arXiv.1810.04805

(2018). MLflow. MLflow. https://mlflow.org/releases/

Whittaker, Z. (February 16, 2019). OpenAI built a text generator so good, it’s considered too dangerous to release. TechCrunch. https://techcrunch.com/2019/02/17/openai-text-generator-dangerous/

(January 19, 2026). Kubernetes Established as the De Facto ‘Operating System’ for AI as Production Use Hits 82% in 2025 CNCF Annual Cloud Native Survey. CNCF. https://www.cncf.io/announcements/2026/01/20/kubernetes-established-as-the-de-facto-operating-system-for-ai-as-production-use-hits-82-in-2025-cncf-annual-cloud-native-survey/

(2021). GitHub Copilot. GitHub. https://github.com/features/copilot/

(2024). GPT-4o. OpenAI. https://en.wikipedia.org/wiki/GPT-4o

Atalan, Y. (February 2, 2025). DeepSeek’s Latest Breakthrough Is Redefining AI Race. Center for Strategic and International Studies. https://www.csis.org/analysis/deepseeks-latest-breakthrough-redefining-ai-race

Malik, A. (November 5, 2023). OpenAI’s ChatGPT now has 100 million weekly active users. TechCrunch. https://techcrunch.com/2023/11/06/openais-chatgpt-now-has-100-million-weekly-active-users/

(July 31, 2024). AI Act enters into force. European Commission. https://commission.europa.eu/news/ai-act-enters-force-2024-08-01_en

Ponseel, M. (April 6, 2021). Kubernetes Adoption Rises But Challenges Still Remain. DEVOPSdigest. https://www.devopsdigest.com/kubernetes-adoption-rises-but-challenges-still-remain