Last week, Future #1 made a specific claim: enterprises will hire agents before they know how to manage them, and the management layer is already being built. And now, Future #2: Once you have agents operating at scale, the next question is what data they can actually touch.
Future #2: a future where the enterprise securely leverages best-in-class general models, tuned to their needs and their proprietary data, in a seamless and unified experience.
The next wave of trillion-dollar companies will be built by combining foundation models with proprietary data. The bottleneck for enterprise AI is not the model itself; it is the trust scaffolding around the model: evaluation, enrichment, integration, and deployment. The clearest early illustration of that stack at work is a data engineering company that has spent the past year solving the hardest layer of all.
TensorStax: the data engineering proof case
TensorStax builds deterministic AI agents for data engineering. The core product autonomously designs, builds, and deploys ETL/ELT pipelines and data warehouse transformations through a proprietary LLM Compiler layer that imposes structure on what is otherwise a non-deterministic process.
The company’s founding insight is precise. Aria Attar, CEO, puts it directly: “Data engineering is a lot more rigid than software engineering; the level of precision you need is extremely high, much higher than most other technical agent use cases.” The LLM Compiler exists because general agents, left to their own devices, fail at exactly this kind of constrained, exacting work.
The numbers bear that out. TensorStax’s internal benchmarks show agentic task success rising from 40-50% with a base agent to 85-90% with the LLM Compiler layer in place. That gap separates an interesting demo from a production system.
Attar frames the market bottleneck this way: there is roughly one data engineer for every 250 software engineers. Every modern company is drowning in data it cannot pipeline fast enough to use. The models exist; the pipelines don’t. TensorStax’s second core insight is that the path forward runs through existing infrastructure rather than around it: “if we can get these systems to work with whatever existing tech stack a company has today, tools like Airflow, dbt, Spark, and so on, it ends up driving way more value.”
Bee Partners co-invested in TensorStax’s $5M seed, announced in May 2025. The enrichment layer is the crux of Future #2: you cannot tune a general model to proprietary data if you cannot first get that data into a structured, pipeline-ready state.
TensorStax is the enrichment layer. But the stack is broader than one company.
The infrastructure that makes general models trustworthy with private data
The enterprise cannot simply point a foundation model at its proprietary data and trust the results. Three other layers have to work before that trust is earned: evaluation, integration, and deployment.
The evaluation layer is the gating mechanism. Before an enterprise can trust that a model-plus-proprietary-data system is working, it needs continuous measurement: does the output correspond to the underlying data, at the level of precision the business requires? Without a CI/CD-native evaluation gate, LLM work stays an experiment. It never becomes infrastructure.
Okareo is building that gate. The platform provides automated LLM evaluation and observability for foundation models, RAG pipelines, and agents: synthetic data generation, runtime telemetry, and evaluation integrated directly into deployment pipelines. Matt Wyman and Boris Selitser, the founders, are building eval as a first-class engineering practice rather than an afterthought.
The integration layer is where a foundation model meets structured enterprise data. This is the problem that most enterprises think they can solve with retrieval-augmented generation, and where they consistently find the limits of a generative approach. RAG surfaces relevant text; it does not execute precise queries against structured databases.
Infactory is approaching this differently. The platform pre-compiles query logic so that retrieval executes directly against databases, without live LLM involvement at the retrieval step. Hallucination cannot happen at a layer where the model isn’t generating. Ken Kocienda, CTO, who invented iPhone autocorrect at Apple, has described the goal plainly: the system should refuse to answer when it can’t, rather than make something up. Brooke Hartley Moy and Kocienda are presently working through alpha with design partners.
The deployment layer is the one most enterprises overlook entirely. Most proprietary data doesn’t live in a clean cloud database. It lives on industrial sensors, embedded hardware, factory floors, and regulated environments where data cannot leave the device. The eval, enrichment, and integration layers mean nothing if they cannot reach where the data actually originates.
Atym is the substrate that makes the rest of the stack deployable at the edge. The platform brings WebAssembly-based container orchestration to resource-constrained hardware: environments with as little as 1MB of RAM. Bee Partners invested because without Atym, the rest of the proprietary-data stack cannot reach the environments where enterprise proprietary data is generated.
The trust gap
In mid-2025, the market is broadly focused on fine-tuning as the path to enterprise differentiation. The thesis is that companies with enough proprietary data can fine-tune foundation models into specialized tools, creating durable moats through data network effects.
That thesis understates the problem. Fine-tuning is a capability question; trust is a deployment question. The enterprise cannot use a fine-tuned model against live proprietary data without first knowing the outputs are reliable (evaluation), the data is structured and current (enrichment), the query layer is accurate (integration), and the model can reach the data where it lives (deployment). Those four layers carry the work that turns proof-of-concept into production. Without all four working in concert, the enterprise stays in pilot mode.
The evidence that this gap is real comes from the usage patterns themselves. A Microsoft and LinkedIn study last year found that 75% of knowledge workers were already using AI, but roughly 78% of them weren’t telling their employers. The pattern has a name: shadow usage. Workers are using general models to interact with private data, ungoverned and untested, because the trust infrastructure to do it properly doesn’t yet exist at scale. Enterprises are stuck in perpetual proof-of-concept loops not because the models aren’t capable but because the scaffolding around the models hasn’t been built.
Bee invested in the trust scaffolding before the market framed it that way. TensorStax at the enrichment layer, Okareo at the evaluation gate, Infactory at the integration point, and Atym at the deployment substrate: together they represent our thesis that trust, not foundation-model choice, is the durable moat in enterprise AI. Enterprises will buy that scaffolding from the companies already building it.
That is Future #2.