Silicon Valley loves a clean narrative. Find a young dropout, stick him in front of a whiteboard, give him a catchy mission statement about democratizing intelligence, and watch the venture capital floodgates open. Aidan Gomez fits the casting call too well. Co-author of the seminal attention mechanism paper, founder of Cohere, and standard-bearer for enterprise language models. The media treats him like an oracle prophesying a future where every Fortune 500 company runs on customized neural networks.
It is a comforting bedtime story for executives terrified of becoming obsolete.
It is also completely wrong.
I have watched traditional enterprises set millions of dollars on fire trying to shoehorn massive transformer models into messy, legacy supply chains. They hire expensive consultants, buy into the hype of bespoke corporate brains, and end up with an expensive chatbot that hallucinates inventory numbers. The lazy consensus of the tech press is that Gomez and his peers are building the infrastructure of modern commerce. The reality is much uglier. Enterprise AI as it is currently pitched is a solution looking for a problem, propped up by circular marketing and a profound misunderstanding of how actual businesses create value.
Let us dismantle the mythology brick by brick.
The Enterprise Customization Fallacy
The core pitch from Cohere and similar shops is simple. Consumer models like OpenAI are generalists. They know everything about poetry and nothing about your balance sheet. Therefore, enterprises need private, secure, customizable models trained or fine-tuned on proprietary corporate data.
Sounds logical on a slide deck. In practice, it is a structural dead end.
Data is not an asset just because it exists in a database. For most companies, internal data is a toxic waste dump of duplicate spreadsheets, outdated PDF policies, and Slack messages arguing about lunch. Fine-tuning a multi-million-parameter model on corporate detritus does not create an oracle. It creates a very expensive echo chamber that reproduces institutional blind spots with high confidence.
Consider what happens when a company tries to deploy these systems into critical workflows. If you build a model on last year's procurement logs, you are encoding historical inefficiencies into an automated decision engine. You are not modernizing operations. You are calcifying past mistakes into code.
Gomez speaks eloquently about data sovereignty and privacy. Enterprises want their proprietary intelligence kept behind firewall boundaries. Fair enough. But treating a language model as a secure vault for corporate IP misses the fundamental mechanical nature of these systems. Transformers do not store facts in a secure register. They compress statistical correlations. When you ask an enterprise model a specific question about internal compliance, it is not looking up a rulebook. It is guessing the next most likely token based on a probability curve shaped by messy text files.
That is not enterprise software. That is a probabilistic magic eight-ball with high electricity bills.
The Commoditization Trap
Let us look at the economic reality. Cohere positions itself against the consumer giants by focusing strictly on business use cases—search, retrieval, text generation for internal workflows.
This is a dangerous middle ground.
OpenAI and Anthropic are pouring billions into foundational capabilities that improve every month. Google and Microsoft are bundling functional AI tools directly into the enterprise operating systems companies already pay for. When your office suite comes with native intelligence baked into the spreadsheet cells and document editors, buying a separate subscription for a standalone enterprise platform starts looking like an unnecessary luxury.
The moat is drying up before it is even finished.
If foundational model capabilities are rapidly commoditizing—and they are—then selling APIs that wrap those models in enterprise packaging is a race to the bottom. Margins compress. Customer acquisition costs skyrocket. The companies winning right now are not the ones building slightly better retrieval-augmented generation pipelines for insurance companies. They are the ones owning the workflow, the data pipelines, and the user interface where work actually happens.
Gomez is betting that enterprises will always need a specialized provider for their cognitive plumbing. But plumbing is only valuable if the water is clean and the pipes don't leak. Right now, enterprise AI plumbing is leaking hallucinations all over the executive boardroom.
What You Should Be Doing Instead
If you run technology strategy for a mid-market or enterprise organization, stop listening to the prophets of general intelligence. Put away the fantasy of training your own custom transformer.
Focus on deterministic software first. 90 percent of the business problems companies try to solve with expensive language models can be handled by standard relational databases, robust APIs, and simple deterministic logic scripts. If a script can check a compliance rule in two milliseconds with zero percent error, do not spend six figures deploying a probabilistic model that gets it right 94 percent of the time.
When you do use language models, treat them like untrusted interns, not infallible systems of record. Build strict validation layers around every single output. If the model generates code, run it through an automated test suite before it touches production. If it summarizes financial reports, force it to cite exact document IDs and line numbers, and build fallback triggers that kick execution back to a human the moment confidence dips.
The companies winning with AI are not the ones buying into the hype cycle of bespoke enterprise models. They are the ones ruthless enough to treat intelligence as a commodity tool rather than a corporate savior.
Stop buying the ghost. Build the machine.
