RAG is suited to changing factual information, while fine-tuning helps a model reproduce the decision process behind a person's work.
2
Small language models can absorb process knowledge with less compute, then act through tools such as Slack, Gmail, calendars, and browsers.
3
Enterprise agents should work with people through shared plans, permissions, and review steps instead of receiving unrestricted autonomy.
Summary
Jaipal Singh Goud distinguishes between fast data and slow data. RAG can retrieve changing factual information, while fine-tuned models can learn the reasoning and preferences behind repeated decisions. He describes small language models as easier to train, cheaper to run, and useful for capturing process knowledge from a person's work. His proposed agent has an intelligence layer, a memory layer, and actions. Prem AI's Prem Studio is presented as a way to turn collected work data into synthetic training data, choose a model, run fine-tuning and reinforcement learning, and evaluate the result. The discussion then moves to deployment. Agents may start as sidekicks across Slack, email, and voice, then gain control of browser and desktop tools. Goud is cautious about full autonomy in creative or information-sparse work. He prefers plans that humans and agents can revise together, with review points. He expects an operating-system layer to load specialised models for particular applications as tasks change.
RAG retrieves changing facts while fine-tuning captures decision habits
Jaipal Singh Goud describes fine-tuning and RAG as complementary. RAG is useful when a system needs factual information that changes often, such as customer-support content or an expanding database. Fine-tuning fits slower information: the considerations and patterns behind a decision. A fine-tuned model can decide what questions to ask the RAG system, while RAG supplies current context. Goud compares this with personal notes. Notes provide remembered information about vaccines, while a person's considered opinion reflects knowledge gathered and thought through over time. The two approaches work best together when the model needs both current facts and a repeatable way of making decisions.
An agent combines intelligence, memory, and actions
Goud defines an agent as three parts: an intelligence layer that makes decisions, a memory layer that learns and remembers, and an action layer that can read, write, or perform tasks. Small language models belong in the intelligence layer. He describes models between 1 and 7 billion parameters as relatively easy to train and useful for absorbing organisational knowledge. A company can fine-tune one on how a person performs a process, including the thinking between moving data from one place to another. This process knowledge is difficult to obtain through prompting alone. The model is intended to predict the next step given a situation and its context.
Collecting work data requires control over observation and retention
The proposed virtual workforce depends on data collection, model fine-tuning, and action. Data might come from Google Meet, Slack, Notion, notes, screenshots, or screen observation. Goud says the resulting data set should capture how a particular person performs a role and why they make decisions, rather than being a generic data set for an entire day. He also treats constant observation as a serious privacy problem. People need authority over where their data goes and how it is used. He says Prem AI keeps training data in its ecosystem, exposes it only to hosted open-source models, and deletes data after a seven-day period once the models are trained.
Prem Studio turns role data into a model that predicts the next action
Goud describes Prem Studio as a product that takes collected work data and helps prepare it for fine-tuning. It creates synthetic data, selects a small language model, runs training jobs, provisions GPUs, monitors the process, and can run reinforcement-learning experiments. The trained models are evaluated against held-out test data. The training objective is specific: given a situation and its context, the model should predict the next step. The intended result is a digital sidekick for a person's role, such as a marketer's way of working. The model can then be connected to tools such as a calendar, Gmail, Slack, or GitHub, with permissions controlling what it may do.
Agents need interaction design that people can trust
The conversation treats enterprise agents as a design problem as well as a model problem. Goud says AI systems are low-trust systems placed in high-trust environments. People may reject a document when it looks machine-generated, even if substantial work went into it. He argues that an agent should behave in ways people recognise from working with a colleague, including communicating through familiar channels and giving enough context. The agent might first review a document or answer a question through Slack, email, or voice. Later it could operate inside a browser or desktop workspace. The interaction needs to feel understandable and useful before the system is given more control.
Repeated workflows can constrain an agent's choices
When a person's behaviour follows a narrow pattern, Goud suggests analysing the collected data to find that pattern and reduce the model's inference space. A process with little variation may be better represented by a more predictable workflow. He compares this with a Markov chain, where events connect to likely next events. A social-media task may contain many possible paths, while a Google Ads task may follow a tight sequence. Memory systems and confidence scores could help ask whether the person has taken a particular next step before. That information could then guide retrieval and constrain the answer or action space during inference.
Creative work calls for planned collaboration instead of full autonomy
Goud says agents have less training data and face more creative decisions when they work across custom software. Video editing illustrates the problem: a system must choose titles, timing, colour grading, and other small decisions while moving between tools. He expects human-augmented agents in these settings. A person could provide a plan, let the agent execute several steps, and require a review before continuing. The plan itself could be revised together, with the agent suggesting changes. Goud compares this with coding agents, which can create a pull request but still need human review. He rejects the idea of delegating a complex task and simply hoping it finishes correctly.
Specialised models may be loaded for each application at runtime
Goud expects an intermediate orchestration layer at the operating-system level. Given a goal, it would plan which models and tools are needed, open an application, load the relevant model, and unload it when the task changes. A Premiere Pro model could handle video editing, followed by a browser-use model for publishing to YouTube. Each model could also be fine-tuned to an individual's style, such as how they write titles. This structure begins with one model when data is limited, then splits into specialists as the data grows. Goud predicts a large number of models tuned for particular people and applications, with models swapped in at runtime.
"It's going to be this family of models and we're 100% confident that it's going to be a billion models fine-tuned for specific people and applications that are going to be live out there."Jaipal Singh Goud36:08
Who should watch
You are deciding whether a workflow needs current retrieval, a fine-tuned model, or both, and want a practical distinction between those choices.
You are building an enterprise agent that may observe employees or access workplace tools, and need to think through permissions, retention, and trust.
You are experimenting with browser-use or computer-use agents and want a case for plans, specialised models, and human review.
