How enterprises can train AI models without ML teams or labeled data

Enterprises are discovering that their existing AI workflows already generate the data needed to train custom models—but most organizations fail to capture this resource. San Francisco-based startup Empromptu AI is changing that with Alchemy Models, a platform that transforms real-time AI outputs and expert corrections into continuous model improvements. The system eliminates the need for separate ML teams or labeled datasets, allowing companies to refine AI behavior directly from production workflows.

Turning wasted AI interactions into training data

Every time an enterprise AI application processes a query or receives expert feedback on its output, that exchange represents untapped training potential. Most organizations let this data slip away, losing the chance to refine models for their specific use cases. Empromptu’s Alchemy Model system captures these interactions automatically by routing validated outputs from subject matter experts back into a fine-tuning pipeline.

The platform distinguishes itself from traditional fine-tuning and retrieval-augmented generation (RAG) approaches. Traditional fine-tuning requires curated datasets and separate ML infrastructure, while RAG retrieves external context without altering model weights. Alchemy integrates directly into existing enterprise applications, using the workflow itself as the data source to generate, clean, and refine training data continuously.

Empromptu CEO Shanea Leven emphasizes the operational challenges companies face when relying on third-party models. "Every customer I speak with asks the same question: How will I avoid disruption? How can I take control of my business model?" she explains. "The path forward isn’t clear when you’re dependent on external systems you don’t own."

From production workflows to specialized AI models

Alchemy’s process begins before an application is even built. Empromptu’s infrastructure, called Golden Data Pipelines, cleans, extracts, and enriches enterprise data to create structured inputs for the AI system. Once deployed, the application’s outputs are fed back through the pipeline, where internal experts review and correct them. These validated corrections then serve as the foundation for the next fine-tuning iteration.

The result is what Empromptu calls Expert Nano Models—compact, task-specific models optimized for a single workflow rather than general-purpose reasoning. The platform handles evaluations, guardrails, and compliance controls within the same pipeline, ensuring governance remains consistent. Customers retain full ownership of the model weights, though Empromptu hosts and runs inference on its infrastructure. For a fee, enterprises can export their models for use elsewhere.

The primary limitation is data volume. Early deployments operate on base models until sufficient production data accumulates for meaningful fine-tuning. "Training the model will take time," Leven notes. "But the data is already there—we’re just ensuring it’s put to use."

Why Alchemy simplifies fine-tuning for non-ML teams

Competitors like OpenAI’s fine-tuning API and AWS Bedrock Custom Models require organizations to prepare datasets externally and manage fine-tuning outside their application stack. This process demands dedicated ML expertise and infrastructure, creating a barrier for most enterprises.

Alchemy removes these hurdles by integrating the training pipeline directly into the application workflow. "Do I need to spin up yet another ML team to fine-tune a model? No," Leven says. "Anyone can do it now." The tradeoff is platform dependency: Alchemy only functions within Empromptu’s environment, meaning enterprises would need to replicate the entire pipeline if they want similar results on existing infrastructure.

A real-world example: 87% faster documentation in behavioral health

Empromptu is targeting highly regulated, data-heavy industries like healthcare, financial services, legal technology, retail, and revenue forecasting—sectors where generic AI outputs often miss the mark. Behavioral health provider Ascent Autism is among the early adopters using Alchemy to automate session documentation and parent communications.

Facilitators previously spent one to two hours manually writing notes after each learner session. By training Alchemy on session recordings, transcripts, behavioral metrics, and internal notes, the process now takes just 10 to 15 minutes. The system generates structured session notes and personalized parent updates while aligning outputs with Ascent Autism’s clinical voice.

Faraz Fadavi, co-founder and CTO of Ascent Autism, highlights the cost and quality benefits. "API-based models can become prohibitively expensive at scale," he says. "Alchemy let us structure our workflow, train on our own data, and reduce costs while improving output quality over time." The company prioritized traceability to session data and consistency with its clinical tone, ensuring the AI model learns to operate like their human experts—not just summarize text.

As enterprises seek ways to reduce dependency on third-party AI systems, platforms like Alchemy offer a practical path forward. By leveraging existing workflows and internal expertise, companies can build AI models tailored to their needs without the overhead of traditional ML pipelines. The next phase of enterprise AI may depend less on external tools and more on the data organizations already produce every day.