Six years ago I sent an email no one on the team expected. At the end of January 2020, just back from a transatlantic trip, I had noticed something I couldn't shake — small signals, barely visible, but sharper by the day.
As someone who was never a fan of working-from-home policies, I told my team they should stay home if they didn't feel like coming in. This was weeks before the lockdown, while the rest of the region was still mainly oblivious.
Not clairvoyance, but a gut feeling combined with the conviction that as a business leader you sometimes have to make a drastic decision before anyone else sees the need. Driven by a sense of civic responsibility. In what then was 12 years of company history, that was the first time I wrote a message like this one.
This is the second time. 18 years in business.
Last Friday I communicated another unexpected decision: we are banning all OpenAI products. Full stop. Not because of technological shortcomings — ChatGPT works and integrates excellently — but because what has been unfolding in AI over the past days and weeks.
We are on the eve of something big, just like in March 2020, and the signs are there for anyone willing to look. Claude Opus 4.6 and OAI's Codex 5.3 are stunning developers. Without yet being able to put our finger on it, you can feel something's up. The advancements with AI are pushing us perhaps to singularity. There's a lot of opportunity but also some risk, and we all play a role when it comes to the latter.
Here at LeadFabric we're not walking away from AI — in fact, we're doubling down on it. In many ways. But with one change already in place: without OpenAI. Let's first take a step back to understand our position and why we're ditching OpenAI.
The cycle of hype and disillusionment
We're living a paradox. Boardrooms — especially in tech — have spent the last few years buzzing over AI demos. Impressed reactions, bold strategies, all built on the assumption that today's AI infrastructure is stable and reliable.
On the other hand, you can see the classic hype curve at work: the same cycle of inflated expectations followed by the trough of disillusionment.
The first wave hit tech companies themselves. Investor pressure led to heavy cuts in sales and marketing — a sobering correction. "AI will solve this" turned out to be a bit more nuanced than the pitch decks promised.
Then came a second wave. Sectors like banking are cutting headcount with AI efficiency as the justification. Where the tech sector has already sobered up, these industries are just entering the "this has to happen now, our competitors are doing it too" phase.
But now — and this is where it gets interesting — the next wave is coming. Recent stock prices, with SaaS shares from Adobe, HubSpot, and Salesforce losing between 40 and 70% in a single year, reflect a growing market belief that AI will soon write all software. Even enterprise platforms aren't being spared.
Aside from all of that, it also seems that workforce reductions are used proactively to force change against the inertia of existing teams that keep operating the old way. That this conveniently delivers a short-term positive impact on the bottom line is a welcome side effect.
This cyclical pattern — enthusiasm, correction, displacement into other sectors — only makes the underlying questions more urgent. Because while we're still figuring out what AI can and can't do, decisions with irreversible consequences are already being made.
And those decisions are being built on foundational infrastructure that far too few people are scrutinizing.
The paradox of legacy enterprise platform vendors in the AI era
The traditional enterprise software industry finds itself in a fundamental strategic contradiction. On one hand, their current subscription-based revenue streams are still dependent on static, process-oriented applications that rigidly lock in workflows. For the back-office applications of IT stack platform vendors, this is less of a problem — but their CX and front-office applications are in trouble.
Many software vendors — forced by market pressure and competition — acknowledge somehow reluctantly that their architecture and its underlying process layer is no longer sustainable in an environment where agentic AI systems facilitate flexibility and dynamic adaptation.
In the front-office domain, this disconnect is growing even faster: where customer expectations are evolving exponentially in all kind of directions — inspired by digital experiences outside the B2B context — marketing and sales platforms continue to innovate only linearly within the constraints of their legacy architecture.
This tension manifests in what we might describe as a "hybrid transition strategy": platform vendors are implementing AI agents as functional extensions on top of their existing infrastructure, rather than rethinking the entire play. This isn't a technology choice — it's an economic one: a complete overhaul would destabilize their current revenue model.
The result? Incomprehensible architecture diagrams. Let's zoom in for a minute into the area of sales and marketing tech. Talk to any buyer of marketing enterprise technology today and a clear pattern emerges — one that's hard to see as anything other than "confuse the customer," a proven tactic in enterprise IT.
Where SaaS and cloud once won by schmoozing directly with business users, circumventing IT and entering the enterprise through the back door, it now feels like IT has quietly taken back complete control. The business logic flowcharts and architecture powerpoint slides have become so complex that once again, platforms are being pushed on organizations that the business side never asked for. Technology decisions that precede process design, with IT teams closely aligned to their stack vendors who promise that fully integrated platforms offer the most future-proof solution. However, the "seamless integration" they tout frequently exists only at the UI level—in reality, many stack components are cobbled together through M&A activity rather than built as a cohesive whole.
If any of you have been following Salesforce's latest chapter with their new Marketing Cloud, you know exactly what I mean. It's a mess — as in "if you can't convince them, confuse them." I'm not even sure most of their sales reps can explain the proposition decently themselves. It's buzzword bingo all over the place. B2C and B2B? Account versus individual orchestration? No difference, apparently. And when it comes to very specific functional needs the platform today can't seem to handle, the answer is always "we do that via the agent layer." Yeah... sure you do.
Microsoft and Copilot? Same story. AI technology is being sold — or initially given away — before the use cases are defined. The ROI that Copilot users are currently seeing is surprisingly low. Only 3% of Microsoft 365 users have opted to pay for it, and even Satya Nadella admitted (reported via The Information) internally that the integrations "don't really work." That's what happens when technology trumps user and process needs.
Gartner: Predicted that over 40% of agentic AI projects will be canceled by end of 2027 due to escalating costs and unclear business value.
Strategic advisory in a fragmented ecosystem
Unlike the traditional system integrator model — characterized by IT vendor entanglement where commission structures routinely took priority over client interests — LeadFabric has consistently maintained a business-first approach for 18 years.
The pathologies of vendor-driven implementations are well known: feature-sales maximization regardless of client maturity.
We were working hands-on with North American marketing and sales technology before those vendors even had boots on the ground locally. That gave us the independence to right-size subscriptions based on what clients actually needed. We didn't sell software based on the price and feature list. We sold change that SaaS enabled. Rightsized to the client needs and maturity levels.
This orientation structurally clashes with vendors who are incentivized to sell as much as possible. And our relationship with enabling technology providers frequently collided after their acquisition by stack vendors, because we kept playing the client's card. The acquiring vendor's sales teams — often unhindered by any real knowledge of the processes or tools their company had just bolted onto the price list — were mostly there for one reason: upsell into the existing client base.
The current hybrid transition strategy only intensifies this tension. And the risk is that we'll see more of the same — except this time with bigger consequences. AI enabled change triggers secondary change: in infrastructure, in architecture, in how buyers and teams operate. Getting your process transformation advice from the entangled vendor/SI combination is more dangerous than ever. You just don't buy tech without knowing what to do with it, but that is exactly what we see happen. e.g. with copilot and others.
Organizations investing in monolithic single stack front-office platforms today risk a threefold cost:
- vendor lock-in to depreciating architecture,
- change management for systems of diminishing relevance
- parallel costs during the inevitable migration to agent-fabric models.
Our Fabric strategy anticipates this shift: composable workflows with variable lifecycles, orchestrating across legacy systems where necessary, but architecturally centered on agentic capabilities. This requires deep AI competence at the infrastructure and orchestration level — not merely presentation-layer integration.
But it also means we must be far more critical and deliberate about our AI infrastructure and its components. The ethical dimension of the providers matters too. There is too much potential impact.
Back to the reason for this post: OpenAI and growth at all costs — literally
So if migrating to or doubling down on all-in-one platform stacks including a native agent layer isn't the answer — and weaving your own fabric is — you need to be deliberate about the building blocks in that mesh. This brings me back to OpenAI. Are they a reasonable choice?
OpenAI's strategy is crystal clear: to become too big to fail as fast as possible, and ultimately sit inside every node of your fabric and that of everyone else's. That means cutting corners where it hurts and allowing anything as long as it feeds growth.
Consistency? Gone. Just two years ago, Sam Altman claimed OpenAI would never introduce advertising. It was positioned as a core differentiator. Now it's quietly being rolled out. The problem isn't the feature itself — it's the pattern of stated principles evaporating the moment the growth curve demands it.
Transparency? A myth. There is nothing "open" about OpenAI. Where competitors publish their preparedness frameworks and allow external audits, OpenAI continues to operate behind closed doors. Stanford's Graduate School of Business demonstrated political bias in outputs. Their lobbying against the EU AI Act was outright aggressive. The high employee turnover at OpenAI can frequently be traced back to conflicts between staff and the company's imposed strategic direction.
Safety guardrails? Retrofitted. Their Preparedness Framework shifted from pre-release testing for manipulation potential to "monitoring after the fact." For a general-purpose chatbot without strict constraints, that's reckless. Deepfakes, misinformation, propaganda — the model is neutral toward these use cases. All in service of growth.
Vendor lock-in? By design. Proprietary APIs, closed source, data strategies that make extraction easier than export. Once you're dependent, you play by their rules.
This applies to all proprietary LLMs, but OpenAI makes no effort whatsoever to mitigate it. Anthropic and Google take a different approach, and open-source models even more so.
And then there's the shady collaboration with defense contractors, the inadequate moderation of dual-use applications, and the fundamental questions about data privacy that are dodged with vague statements.
The evaporating early mover advantage of OpenAI
Here's where it gets interesting — and unsettling. OpenAI's technological lead is disappearing fast. Claude outperforms GPT-4 on complex reasoning, Gemini excels at multimodal tasks, Mistral is closing the gap with every release while staying open source, and DeepSeek demonstrated that comparable performance is achievable at a fraction of the cost. We're not saying we should go for Deepseek though, but that is another discussion.
We see it in our own benchmarks: where ChatGPT was still the standard for specific tasks six months ago, we now regularly find alternatives that deliver better outputs, respond faster, or align more precisely with our use cases.
OpenAI's market share is declining. Which isn't abnormal in itself and shouldn't be a problem either. Their early mover advantage — once a cornerstone of their valuation and influence — is evaporating faster than they can ship new features. According to Rutger Bregman, they spend three dollars for every dollar they make. So anything that can close that gap is critical to them.
And here's the crux: given their track record of erratic policy changes, disappearing principles, and a growth-at-all-costs mentality, what do you expect will happen as the pressure mounts?
A company that traded principles for growth while sitting at the top will not suddenly become more transparent and ethical now that competition is catching up. Quite the opposite. The temptation to operate even more opaquely, lobby even more aggressively, and cut even more corners only grows as market share shrinks and investors grow impatient.
That's not speculation — it's pattern recognition. We've seen it in tech history before: companies that lose their early mover advantage make increasingly erratic choices as the pressure builds.
What's happening beneath the presentation layer?
This is where the real point lies. Most AI discussions start from existing infrastructure and closed models. Many enterprises banned usage of all public LLMS and revert back to solutions where they only train the model using their own data. There's something to be said about that too. Especially when we are talking about content generation. But that is food for another blog post.
Others, usually midmarket players, that impose less restrictions using LLMs try to use it for answering questions like: "How can we use ChatGPT for analysing pipeline propensity?" "How can we integrate this into our outbound engagement tooling?", "How can we be using chatGPT via workflows and how in agent mode can it suck data from legacy applications to take next best action decisions?" etc. Well, these sort of questions can't be asked without considering the infrastructure side of things.
Before we get to the how part, we need to ask questions that should run much deeper.
Data sovereignty: What and whose data are we talking about? Where are your conversations being processed? Under which jurisdiction? Who has access? How long is data retained? Many enterprise contracts are vague on these points — until there's a breach.
Model governance: Is your vendor training on your data? Can you enforce an opt-out? How do you actually verify that? OpenAI's data policy has changed multiple times, often without announcement. Under government pressure they can't guarantee full privacy, yet their settings panel still gives you the impression you can. These aren't oversights — they just don't want anything getting in the way of usage.
Technological Independence / Sovereignty: Europe is finally waking up from its naivety about technological dependency. The collaboration between ASML and Mistral — far too little discussed — sends a signal. Why are we building critical infrastructure on someone else's closed-source models?
Ethical architecture and culture: AI ethics is not a checkbox exercise after the fact. It lives in model selection, training data, deployment strategy, and guardrails. A general-purpose chatbot without strict boundaries is a fundamentally different ethical choice than specialized, auditable agents.
Total cost of ownership: The monthly API costs are a fraction of the picture. The real costs? Vendor dependency, re-engineering when pricing changes, compliance risk, reputational damage in case of misuse, migration efforts when performance falls behind competitors, and the cost of exploring alternatives when your vendor makes a strategic pivot.
Performance stability: What if your workflows are built on GPT-4's capabilities, but a competitor delivers better results tomorrow? Do you have the flexibility to switch, or are you locked into an ecosystem?
Our Fabric course: building at a deeper layer
Our decision is not anti-AI. On the contrary. It's pro-fundamentals.
Since 2008 we have been running 100% cloud-native. Not because it was trendy, but for security and compliance reasons for our enterprise clients. In 2008 we were the very first Box enterprise customer in the Benelux. We've been running on Salesforce Enterprise just as long. Our PSA is Wrike Enterprise. And so on.
When it comes to AI: we are experimenting broadly today with Claude, Perplexity, local LLMs, and on-device offline deployment. Our team uses generative AI daily for data analysis, content, coding, and engagement. Until recently, often with OpenAI. No longer.
The enterprise infrastructure space is helping too, with AI-driven SecOps support solutions and AI-powered SIEM platforms, MCP servers etc. We're investing in building knowledge here as well. Just as we've long criticized SIs for blindly following outdated implementation manuals written by engineers rather than process experts — ignoring actual client needs in the front office — we believe the same principle applies deeper in the stack when it comes to implementing agentic AI. As future agentic AI fabric weavers, we need to bring our functional subject matter expertise all the way down to the infrastructure layer, from MCP to where SIEM queuring lives. This can't be left to IT and cybersecurity alone.
We keep experimenting — but with a critical eye. We are prioritizing European LLMs: Mistral and LeChat are getting more attention. We are intensifying our experiments with VPS-hosted internal models. We are developing single-purpose highly specialised agents instead of one-size-fits-all chatbots. We are building secure safe and walled garden LLM solutions for customers where there is a zero LLM policy.
Crucially, we are building competence in the layers beneath the model and the UI: tokenization strategies that guarantee data privacy by design, on-device deployment for sensitive use cases, proactive agent design with safety as a starting point rather than an afterthought, infrastructure ownership so we're not held hostage by any single vendor, hybrid architectures that combine the best of open and closed source under our control, and model-agnostic workflows that let us switch between providers based on performance rather than vendor lock-in.
This is all part of our Fabric strategy: entirely independent of application and IT platform providers, we want to help and support mid-market companies build their own commercial front-office fabric.
We can also offer outside the firewall services for large enterprise front office users currently locked in their own co-pilot bubble. A mesh of capabilities that we and they own — not a dependency on one glossy interface that may no longer be the best option tomorrow.
Back to the story about your new marketing infrastructure. We believe that the future is about building your own Fabric. A combo of legacy systems that won't go away easily, covered with a mesh or fabric of agentic AI workflows complementing missing functionality, where some flows have a longer shelf-life and others that are usually GTM campaign related and that only last for a few Months.
Where are we headed?
For a GTM consultancy/agency the coming years will not be defined by who works with the best chatbot interface. They will be defined by who understands client needs and acts as a true enabler. But also by who understands, controls (!!!), and can shape the underlying infrastructure according to their values and requirements.
That means investing in competencies, not just subscriptions. Building on diverse, auditable infrastructure. Treating ethics as an architectural principle, not a PR statement. Taking European technological independence seriously as we predict that some midmarket and enterprise clients will start looking for these competencies. Embedding safety and security by design, not by disclaimer. And maintaining the flexibility to switch between providers based on performance and principles.
Six years ago I sent my team home because I believed that even as a small player, we had a societal obligation. Today we are parting ways with OpenAI for the same reason.
The question is not whether AI is the future. The question is: whose AI, under what conditions, with what safeguards — and what happens when the vendor takes weird turns?
Lastly, we are open for business.
Take the recent hype around OpenClaw, or is it Moltbot, and similar proactive AI agents. A lot of questions surface here as well. Enough to justify banning its usage as well. Not least because of security reasons.
But let's face it, it's a development that can't be stopped. It's the next evolution. Those who are actively exploring it today, debating it, and experimenting with it in their labs — carefully, without deep connections to their own production data — are building a competitive advantage. We are not against new scary AI evolutions at all, we want to understand it better and we need to learn. But we want to work with parters and solutions that do things right.
The knowledge we gain now will prove critical later. The questions we're asking today will be asked again once the necessary guardrails have been developed. We want to be part of that. We want to help build those guardrails. And we want to make active choices — like the one about OpenAI — rather than being the ones who point fingers and wait
Contributors
-
Koen De Witte
LinkedIn