Prosperaize

A pharma data platform’s AI agent was working. We rebuilt it anyway.

A live AI system on a pharmaceutical market access platform was rebuilt on its own knowledge foundation, and single answers became a working dialogue with the data.

Single answers became a dialogue with the data
Client
Pharma data platform
Industry
Life Sciences
Asked by
Engineering leadership
What we did
RunCompound
How it shipped
Side by side with the live system, staged, no cutover
Built with
A ReAct agent on the existing domain views and metadata
Published
3 November 2026

Results at a glance

What the rebuild changed, on the same data.

<1

minute for complex, multi-part questions, answered as a conversation.

Production use

300+

columns of regulatory, HTA and reimbursement data, asked about in conversation.

Data coverage

7

domain views carried over unchanged from the first build. Only the engine was renewed.

Architecture

0

cutovers. The live system kept serving users until the new agent had proven itself.

Rollout

01 · 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, with regular use at 55% of platform users. The agent took a self-contained question and returned a grounded answer.

Then users started asking differently. A landscape question led to a narrowing, the narrowing to the evidence behind one decision, and that to the assessment document explaining it. The first generation answered questions. Users had begun to hold conversations.

Meanwhile, agentic AI, systems that reason in steps and correct their own course, became viable in production. The system was working and nothing forced a change. We proposed modernizing anyway, and engineering leadership chose to do it from strength.

“Modernizing from strength is cheaper than modernizing from failure, and the only way to never need the latter.”

One conversation, three turns

Before, each step was a separate investigation: filter, export, cross-reference, open documents.

TURN 1The landscapeTURN 2A subset of itTURN 3The evidenceBEHIND ITThe document
  1. Turn 1LandscapeThe HTA landscape for an indication in France, and the current first-line options.
  2. Turn 2SubsetNarrowed to the decisions with the stronger benefit ratings, on exactly those results.
  3. Turn 3EvidenceWhich clinical trials supported the most recent of those, and what kind of studies they were.
  4. Behind itThe documentThe assessment document that explains the decision, linked back to the platform.
Follow-ups work on exactly the results the user was just shown.

02 · What we delivered

Same foundation, a new engine.

The metadata layer and the domain views carried over unchanged. On top of them we built a new reasoning agent.

01
A working dialogue with the dataFollow-ups operate on exactly the results the user was just shown, and references like “the most recent one” resolve against them.
02Many moving parts, one passAn indication, jurisdictions, a treatment line, a date window and evidence characteristics, combined in one question and answered as one.
03Self-correctionWhen a first attempt does not hold against the data, the agent recovers and tries again instead of answering wrong with confidence.
04Knowing the edge of the dataThe agent says what the platform holds and what it does not, so users can trust a “no” as much as an answer.
05Answers that show their workEvery answer comes with the records behind it and the interpretation choices it made, grounded in the complete result.

03 · How it ran

Evidence first, no cutover.

We developed the new agent against the same knowledge foundation and ran it side by side with the live system: same questions, same data, answers compared directly. It went through an adversarial testing campaign, then rolled out through the company’s environments in stages while the existing system kept serving users.

The running system stayed up until its successor had proven itself against it. No big-bang cutover, no disruption to users.

Side by side · before the rollout

Same questions, same data

The readout: the new agent against the live system.

Compared onThe live systemThe new agent
Questions completed without an error91%100%
Time on the same questionsBaseline10% quicker
Follow-ups on the results just shownSingle answersHolds the conversation
Adversarial testing campaignPassed
VerdictServed users until replacedGo: staged rollout

The running system stayed up until its successor had proven itself against it. No big-bang cutover, no disruption to users.

Measured side by side on the same 100 questions and 30+ conversation flows.

04 · Results

Complex, multi-part questions come back in under a minute, as a conversation.

01

Questions that used to break into filters, exports and spreadsheets are answered conversationally, in one pass, in under a minute.

02

Follow-up chains work on exactly the results shown: landscape, subset, evidence, document, in one conversation.

03

The full regulatory, HTA and reimbursement dataset, including the evidence behind each decision, can be asked about in conversation.

04

The hardest questions, multi-condition, cross-source and drill-down, no longer go to in-house experts and data engineers.

05

The knowledge layer from the first engagement carried over unchanged. The engine on top was renewed at a fraction of the original effort.

05 · What they learned

The first build made the data answerable. The modernization made it conversational.

The same knowledge foundation carried both.

The durable asset is the knowledge layerThe metadata and domain views outlived their first engine and carried the second without change.
Modernize while the system is succeedingKeeping a live AI product current is a standing practice, not an emergency response.
Validate against the incumbent, not a demoSame questions, same data. The running system set the bar the new one had to clear.
Conversation changes the productA platform whose users converse with the data works the way analysts work: each answer shaping the next question.

If this sounds like your company

A working system can still be falling behind.

Your AI system works, and nobody is improving itValue that is not growing is quietly falling behind.
Your users have outgrown its scopeThey ask follow-ups it cannot hold and expect it to remember the last answer.
The technology has moved since you shippedAgents that reason in steps and correct themselves are now viable in production.
You own a knowledge layer worth more than the modelWith well-described data, the intelligence layer can be renewed at a fraction of the original effort.

Is your AI working, and standing still?

Schedule a call with a partnerFree until 11 December