A Pharma Data Platform's AI Agent Was Working. We Rebuilt It Anyway.

What Changed

The first engagement ended with the platform answering questions across 300+ columns of regulatory, HTA, and reimbursement data, in every jurisdiction in the dataset, and with regular use at 55% of platform users. The agent did what it was built for: it took a self-contained question and returned a grounded answer.

Then users started asking differently. A landscape question led to a narrowing. The narrowing led to a drill into the evidence behind one decision. That led to the assessment document explaining it. Questions carried more conditions, crossed more of the dataset, and assumed the agent remembered what it had just shown. The first generation answered questions. Users had begun to hold conversations.

Meanwhile the technology moved. Agentic AI, systems that reason in steps, examine intermediate results, and correct themselves, went from research to production viability. The knowledge foundation from the first engagement, the metadata layer and domain-organized views, was exactly what such a system needs to reason against.

The company had a choice most organizations never make deliberately. The system was working. Adoption was growing. Nothing was forcing a change. Prosperaize proposed modernizing anyway, on the grounds that the users had moved, the technology had moved, and the foundation was ready. Engineering leadership made the decision to do it from strength rather than wait until the system fell behind.

What Prosperaize Delivered

The modernization kept the foundation and renewed the engine. The metadata layer and domain-organized views carried over unchanged. On top of them, Prosperaize built a new reasoning agent. What follows is what the platform's users can now do.

A Working Dialogue with the Data

Follow-ups operate on exactly the results the user was just shown. "Of those decisions, which received the higher benefit ratings?" narrows the actual set from the previous answer, not a fresh approximation of it. References like "the most recent one" resolve against what the user saw. Each turn can go deeper, from a landscape to a subset to one decision's trial evidence to the assessment document behind it, without starting over.

In one conversation, a market access user can ask for the HTA landscape for an indication in France and the current first-line options, then narrow to the decisions with the stronger benefit ratings, then ask which clinical trials supported the most recent of those and what kind of studies they were. Three turns. Before, each of those was a separate investigation: filter, export, cross-reference, open documents.

Questions with Many Moving Parts, Answered in One Pass

An indication, a set of jurisdictions, a treatment line, a date window, evidence characteristics: combined in a single question and answered as one. This is the long tail beyond any template.

Self-Correction

The agent corrects itself. When a first attempt does not hold against the data, it recovers and tries again rather than returning the wrong answer confidently. This is the property that separates an agent that handles the unexpected question from one that handles only the anticipated one.

Knowing the Edge of the Data

The agent says what the platform holds and what it doesn't, so users can trust a "no" as much as an answer.

Answers that show their work

Every answer arrives with the records it came from, linked back to the platform, and with the interpretation choices it made, such as "recent" read as a date window, granted rather than requested ratings. The answer is grounded in the complete result of the question, not a sample of it. For teams whose output feeds regulatory and pricing strategy, an answer they can audit is worth more than an answer alone.

How the Engagement Ran

AI Reliability & Optimization, ongoing

This is the fourth stage of how we work: the stage where we keep improving the live AI system to increase its business value.

The modernization ran the way the original build did: evidence first. Prosperaize developed the new agent against the same knowledge foundation and evaluated it side by side with the running system, same questions, same data, answers and behavior compared directly. It went through an adversarial testing campaign before promotion, then rolled out through the company's environments in stages while the existing system kept serving users. No big-bang cutover. No user disruption. The running system stayed up until its successor had proven itself against it.

Results

• Questions that previously decomposed into filters, exports, and spreadsheets are answered conversationally, in one pass, in under a minute.
• Follow-up chains work on exactly the results shown: landscape, subset, evidence, document, in one conversation.
• The platform's full regulatory, HTA, and reimbursement dataset, including the evidence behind each decision, is conversationally answerable.
• The hardest questions, multi-condition, cross-source, drill-down, no longer escalate to in-house experts and data engineers. They are self-serve, in the session.
• The knowledge layer from the first engagement transferred unchanged. The reasoning engine on top of it was renewed at a fraction of the original effort.

What They Learned

• The durable asset is the knowledge layer, not the model on top of it. The metadata and domain views built in the first engagement outlived their first engine and carried the second without change. That was the plan, and it is what "compounding AI assets" means in practice.
• Modernize while the system is succeeding. The company did not wait for adoption to stall or for users to route around the agent. Keeping a live AI product current is a standing practice with a concrete payoff, not an emergency response.
• Validate the successor against the incumbent, not against a demo. Same questions, same data, answers compared directly. The running system set the bar, and the new one had to clear it before taking over.
• Conversation changes what kind of product a data platform is. A platform whose users compose filters is a reference tool. One whose users converse with the data works the way analysts work: iteratively, each answer shaping the next question.

If This Sounds Like Your Company

Your AI system works, and nobody is actively improving it

The system delivers results, adoption is steady, and no one is asking what the next generation of the technology would do on your data. Value that is not growing is quietly falling behind.

Your users have outgrown what the system was scoped for

They ask follow-ups it cannot hold, combine conditions it was not built to combine, or expect it to remember the last answer. The scope was right when it shipped. The users moved.

The technology has moved since you shipped

Agentic systems that reason in steps and correct themselves are now viable in production. If your AI was built before that, the gap between what it does and what is possible on the same data is widening.

You own a knowledge layer worth more than the model on top of it

If your data is well described and well organized, the intelligence layer can be renewed at a fraction of the original effort. The foundation is the investment. The engine is replaceable.

Where to Start

If you run an AI system that is working and want to know what modernizing it would produce, an AI Reliability & Optimization engagement assesses the system against the current state of the technology and your users' actual questions, and shows what a renewed engine on your existing foundation would do. The working system keeps running while you find out.