For every query an enterprise AI application processes, every subject matter expert improves its output – that interaction is training data. Most organizations are not capturing this. The production workflows already created by companies are generating a constant signal that improves AI models, and it is disappearing.
San Francisco-based Impromptu AI launches Alchemy Models on Thursday with a straight base: The AI applications enterprises are already building are generating training data, and much of it is being wasted. The platform automatically captures that signal, feeding validated output from subject matter experts back into a fine-tuning pipeline that improves the model over time. Enterprises retain full ownership of the resulting load.
This sits in a different area from both RAG and traditional fine-tuning. The RAG model retrieves the external context at inference time without modifying the weights. Traditional fine-tuning changes the weights but requires separately assembled labeled datasets and a dedicated ML pipeline. Using enterprise applications as the data source, Alchemy does the latter consistently.
Companies adopting the Foundation Model API face three complex barriers: inference costs that increase with use, no ownership of the models their data is effectively training, and limited ability to customize behavior for domain-specific tasks. Shania Leven, CEO of Empromptu, says those barriers are widely felt but rarely addressed.
"Every customer, everyone I talk to, says how am I not going to be interrupted? How will I protect my business? And they don’t see the way," Leven told VentureBeat in an exclusive interview.
How Alchemy creates a model from a running application
Most custom model training approaches require companies to separately collect, clean, and label the data before any fine-tuning begins. Alchemy takes a different approach: The enterprise application generates and cleans the training data itself.
Tantra Runs Through Impromptu golden data pipeline Infrastructure in two stages Before building an app, enterprise data is cleaned, extracted and enriched so that the application starts with structured input. Once it’s up and running, each output it generates goes back through a pipeline, where subject matter experts inside the organization review it and correct it. That validated output becomes the training data for the next fine-tuning run.
"The app, an AI application that customers are already building, cleans the data," Leven said.
The resulting streamlined models are what Impromptu experts call nano models: small, task-specific models optimized for a particular workflow rather than general-purpose logic. Assessment, guardrails and compliance controls run within the same pipeline, so governance moves along the training process. Customers keep the weight of the model outright. Empromptu guesses and runs on its own infrastructure, but the weights are portable and exportable for a fee. The platform is model agnostic, supporting Llama, Quench and other base models.
The biggest hurdle is the amount of data. The initial deployment runs on the base model while the application accumulates enough production data to trigger useful fine-tuning runs. Levene accepted the timeline without exaggerating. "It will take time to train the model," He said.
Alchemy differs from managed fine-tuning in terms 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 prepared training datasets and manage the fine-tuning process outside their application stack. The burden of data curation and model evaluation rests on the customer’s ML team.
The differentiation process of alchemy is integration. The training data is generated by the enterprise application itself, so there is no separate data preparation step and no ML expertise required. Application workflow is pipelined.
"Do I need Bedrock to fine-tune a model and figure out all that infrastructure and hire another ML team? No, anyone can do it now," Leven said.
The tradeoff is platform dependency. Alchemy only works in impromptu environments. Enterprises that want similar results on existing infrastructure will need to replicate the data capture, validation, and fine-tuning pipeline themselves.
A behavioral health company reduced session documentation time by 87% using Alchemy
Empromptu is targeting regulated and data-intensive verticals first: healthcare, financial services, legal technology, retail and revenue forecasting. These are the areas where general purpose model outputs have the greatest risk of mismatch and proprietary workflow data is most concentrated.
Among the early users is behavioral health company Ascent Autism, which uses Alchemy to automate session documentation and parent communications.
Facilitators use learner session recordings, transcripts, session notes, and behavior metrics to generate structured notes and personalized parent updates. That workflow previously required one to two hours of writing per session. With Alchemy training on the same data, it now takes 10 to 15 minutes.
"Relying solely on an API-based model can quickly become expensive," Faraz Fadawi, co-founder and CTO of Ascent Autism, told VentureBeat. "Alchemy gave us a way to structure the workflow, train models on our data, and reduce costs while improving output quality over time."
Fadawi said the company saw usable output accelerate with continued improvements to the system. The evaluation criteria went beyond accuracy to include traceability of session data and output consistency with the company’s clinical voice.
"We wanted a system that could learn our workflow and generate output tailored to how we actually work – not just summarize text," He said. Practical test: How much does the facilitator need to edit, does the output match their voice and does it meaningfully reduce the time spent. Facilitators have shifted from rewriting the generated notes to editing and quality-checking them.
What does this mean for enterprises
The data flywheel is real – but so is platform lock-in:
Every workflow is a training opportunity. Enterprises that capture and validate outputs from their production AI applications will reap that benefit over time. More usage generates more training signals, which produces more accurate domain-specific models, which produces better outputs, which produces cleaner training data in the next cycle.
Levene positions alchemy as the third architectural choice. Enterprises have spent the last two years choosing between RAGs for domain knowledge access and fine-tuning for model expertise. Workflow-driven model training is a third option, which combines the ongoing improvement of fine-tuning with the operational simplicity of building inside a managed platform.
"Having that data moat is the most valuable currency," Leven said.
<a href