Most AI projects sound useful long before anyone can use them.
The proposal explains what an agent could do. The date for a working system remains vague.
This project started from the opposite direction: one defined job, the company’s real data, and a working R&D agent built in one month.
The client is a manufacturing company whose researchers develop formulations against customer specifications and budgets. The agent was not designed to “do R&D.” It was designed to improve the starting point for a specific part of that work.
Call this the Working-First Build: connect the agent to the real workflow before making claims about its impact.
The work before the experiment
Formulation work does not begin at the mixing bench.
Before physical testing starts, a researcher needs to review older formulation records, check which raw materials are available or approved, and examine proprietary notes about what worked or failed before.
That research helps determine where to begin. But information can sit across different records, making the first pass a job of retrieval and comparison.
The opportunity was not to replace formulation expertise. It was to support the research that happens before that expertise is applied at the bench.
What the agent does
The agent queries the company’s own data: customer specifications, formulation history, raw-material information, and its proprietary knowledge base.
It can return a practical starting point. That might be the existing formulation closest to a new specification, an explanation of how a material substitution could affect cost and properties, or a first draft of a new formulation shaped around the customer’s brief.
The output is not a finished product. It gives the researcher a place to start evaluating.
An AI agent becomes useful when connected to the information required for a defined job—not when it merely produces a convincing answer.
Read [internal link: what an AI agent actually is] for a fuller explanation of that difference.
What the first month included
The one-month build covered requirements, connections to the company’s real data sources, and a version the R&D team could use with its existing records.
It was not a disconnected proof of concept built around sample data. The goal was to make the agent usable inside the actual research process.
The first release focused on the initial research pass because that was the clearest job to define, connect, and evaluate.
What remains human
The agent does not mix materials, assess them physically, or confirm that a formulation works.
Every proposed starting point still goes through a human researcher and physical testing. The agent supports where the search begins. It does not replace the part of the work that requires specialist judgment and real-world validation.
This boundary is part of the product, not a limitation to hide. A useful agent needs a clear handoff to the person accountable for the result.
What happens next
Dihardja delivered a working tool in one month, but the evidence stops there for now. The team has not yet recorded a verified usage count, time-saved figure, or testimonial.
Those results should be measured before this becomes a full performance case study. Until then, the honest claim is narrower: the agent is connected to real company data and is in active use by the R&D team.
Explore the broader delivery approach on our [internal link: services page].
Do not judge an AI project by how convincing the strategy sounds. Ask what people can actually use—and what still needs to be proven.
Have an AI project
in mind?
Let's discuss how AI can transform your business.
