Every query an enterprise AI application processes, every edit a subject matter expert makes to his or her output—that interaction is training data. Most organizations do not capture this. The production workflows that companies have already built create a continuous signal that improves AI models and it disappears.
San Francisco-based Impromptu AI launched Alchemy Models on Thursday with a straightforward introduction: The AI applications that enterprises are already building are generating training data, much of which will go to waste. The platform automatically receives that signal, feeding validated results from subject matter experts into a fine-tuning pipeline that improves the model over time. Enterprises directly own the resulting weights.
It is located in a different area than both RAG and traditional fine tuning. RAG captures the external context during inference without changing the model weights. Traditional fine-tuning varies the weights, but requires separately assembled labeled datasets and a dedicated ML pipeline. Alchemy implements the latter continuously, using the enterprise software itself as a data source.
Companies adopting model-based APIs face three complex limitations: inference costs that scale with usage, ownership of the models on which their data is effectively trained, and limited ability to customize behavior for domain-specific tasks. Shanea Leven, CEO of Empromptu, says these limitations are widely felt but rarely considered.
"Every client, every person I talk to is like, how am I not going to break? How will I protect my business? And they just don’t see the way," Leven told VentureBeat in an exclusive interview.
How Alchemy builds a model from a running program
Most custom model training approaches require companies to collect, clean, and label data separately before any fine-tuning begins. Alchemy takes a different approach: the enterprise application itself creates and cleans the training data.
Mechanism goes through Impromptu Golden Data Pipelines infrastructure in two phases. Before the application is built, enterprise data is cleaned, extracted and enriched so that the application starts with structured inputs. After it runs, each output it generates goes back through the pipeline, where experts within the organization review and fix it. This validated result becomes the training data for the next fine-tuning operation.
"The app cleans the AI application data that customers have already created," Leven said.
The resulting fine-tuned models are what Empromptu calls Expert Nano Models: small, task-specific models optimized for specific workflows rather than general-purpose reasoning. Assessments, guardrails and compliance controls work within the same pipeline, so management goes hand in hand with the training process. Customers have full ownership of their model weights. Empromptu infers and runs on its own infrastructure, but weights are portable and exportable for a fee. The platform is model agnostic, supporting Llama, Gwen and other major models.
A hard limitation is data volume. Initial deployments run on the base model while the application collects enough production data to trigger useful fine-tuning. Levene admitted without sugarcoating the timeline. "Modeling will just take time," he said.
Alchemy differs from the controlled fine-tuning of who does the work
OpenAI’s fine-tuning API and AWS Bedrock custom models both offer enterprise fine-tuning. Both require organizations to bring in separately developed training datasets and manage the fine-tuning process outside of the application stack. The burden of building the data and evaluating the model rests with the client’s ML team.
Alchemy’s differentiation is process integration. The training data is generated by the enterprise application itself, so there is no need for a separate data preparation phase and ML expertise. An application is a workflow pipeline.
"Do I have to have Bedrock and create another ML team to figure out how to tune a model and understand all the infrastructure? No, now anyone can do it," Leven said.
Exchange is platform dependent. Alchemy only works in the Impromptu environment. Businesses looking to achieve the same results on existing infrastructure will have to replicate the data collection, validation and fine-tuning pipeline themselves.
Behavioral health company reduced session documentation time by 87% using Alchemy
Empromptu primarily targets regulated and data-intensive verticals: healthcare, financial services, legal technology, retail and revenue forecasting. These are the sectors where general-purpose model results carry the highest risk of inconsistency and where specific workflow data is most concentrated.
Early adopters include behavioral health company Ascent Autism, which uses Alchemy to automate session documentation and parent communication.
Facilitators use student session notes, transcripts, session notes and behavioral indicators to create structured notes and personalized parent updates. This workflow used to require one to two hours of writing per session. It now takes 10-15 minutes with Alchemy training on the same information.
"Relying solely on API-based models can quickly become expensive," Faraz Fadavi, co-founder and CTO of Ascent Autism, told VentureBeat. "Alchemy gave us a way to structure our workflow, build models based on our data, and reduce costs by improving output quality over time."
Fadavi said the company quickly saw usable results with continuous improvement as the system improved. Evaluation criteria went beyond accuracy and included tracking of session data and consistency of output with the company’s clinical voice.
"We wanted a system that could learn our workflow and produce results based on how we actually work—not just summarizing text," he said. Practical test: how much editing the facilitators need to do, whether the speech matches their voice and significantly reduces the time spent. Facilitators moved from rewriting generated notes to editing and quality checking them.
What this means for businesses
The data flywheel is real, but so is platform locking:
Every workflow is a learning opportunity. Enterprises that capture and validate the results of manufacturing AI applications will increase this advantage over time. More usage generates more training signals, which generate more accurate domain-specific models, better results, and cleaner training data in the next cycle.
Leven introduces Alchemy as a third architecture option. Enterprises have been choosing between RAG for access to domain knowledge and fine-tuning for model specialization over the past two years. Workflow-based model training is a third option that combines the continuous improvement of fine-tuning with the operational simplicity of setup within a managed platform.
"The fact that this data mine is the most valuable currency," Leven said.
