A path-breaking adverse drug reaction prediction model with roughly 85% accuracy was shelved after 18 months. Not because the science failed. Not because the technology underperformed. Because the team could not produce an audit trail. The data lineage was opaque, the evidence was absent, and regulators had no way to verify what the model had actually done. That experience shaped everything Tarini Mohapatra built afterwards.
As founder of TFives, he has spent the years since solving the problem that shelved that program and the broader pattern the Massachusetts Institute of Technology’s Project NANDA documented in 2025: 95% of enterprise GenAI (generative artificial intelligence) pilots fail to deliver measurable returns or reach production at scale, a gap that runs deeper still in regulated life sciences. “Submission packaging is not archaeology,” Mohapatra states. “You need a replayable trail available when inspection or filing arrives, evidence you can reconstruct: same inputs, same rule version, same outcome.”
Evidence Built Into Every Workflow, Not Assembled at the End
The traditional approach to regulatory evidence worked when little changed between development and submission. In modern AI environments, data, models, and processes change continuously. By the time a team scrambles for screenshots and email threads at submission, the story is already broken. TFives treats evidence and compliance as part of every workflow from the first decision onward. Every meaningful action, a rule check, an AI suggestion, a human approval, a data handoff, lands in one Execution Integrity record. That record captures who did what, on which version of the rules, with what inputs, and what outcome was delivered.
The practical shift is significant. Inspections move from reconstruction to replay. What today requires hunting across spreadsheets, document stores, SharePoint, and dozens of disconnected applications converges into a governed export from one spine. Not a magic button, but a predictable path regulators and internal quality assurance (QA) can follow consistently. Reproducibility expectations under the European Union Artificial Intelligence Act 2024 and established good x practice (GxP) are addressed at the point of creation rather than engineered backward from a compliance deadline.
Since emerging from stealth in January 2026, TFives has completed 30-plus proofs of concept and pilots across commercial, quality, safety, and governance workflows. The FIVES Concordance score continuously compares approved content against what is shipped and what the public sees, bridging the deterministic rule engines regulators understand and the probabilistic nature of AI. Just as FICO and Moody’s gave markets a shared language for risk, Concordance aims to establish the same standard for in-market promotional and AI-assisted compliance truth.
Unified Governance Is the Infrastructure That Makes AI Production-Grade
The failure mode in pharmaceutical AI is not a shortage of policies or standard operating procedures. Every function has them. The failure is that each function interprets the same policy differently, using different tools, with different audit trails, producing fragmented governance that AI cannot operate across at a production scale. A McKinsey cross-industry survey, reported in Pharmaceutical Technology in 2022, found 52% of companies investing in AI, but no meaningful increase in AI risk mitigation since 2019 among firms already using AI in at least one function. Teams ship pilots effectively. They stall when delivery requires crossing 5 to 10 disconnected systems with no shared governance layer beneath.
The solution is not more policy documentation. It is a unified governance spine, largely shared across functions with a thin layer of line-of-business personalization, that every team executes against from the same versioned playbook simultaneously. Alignment comes from infrastructure, not from sending more documentation. TFives orchestrates and proves compliance across systems of record, including Veeva PromoMats, safety databases, eQMS, and submission tools. It does not ask enterprises to replace investments they have already made. It connects them under a shared compliance layer that makes the whole greater than the sum of its parts.
Composable Modules That Eliminate the Compliance Bottleneck
Regulatory scrutiny is intensifying on both sides of the Atlantic. The FDA’s Elsa platform and HALO data consolidation, spanning 2025 through 2026, give reviewers AI-assisted access to submission history. Today’s filings will be cross-checked against prior submissions for consistency in claims, specifications, and narrative. That is continuity of evidence across time, not honor-code compliance.
Mohapatra’s framework treats production AI as composable domain modules that plug into a shared spine with compliance, correlation, and replay built in from the start rather than bolted on at the end. Three layers define the architecture. The TFives spine covers Execution Integrity, dual-signal correlation, versioned rules exceeding 2,200, evidence replay, and FIVES Concordance scores. Domain modules span promotional, safety, manufacturing, clinical, submissions, and medical affairs, personalized per division. Starter data packs of 13,000-plus scored use cases tell teams which modules to enable first, removing the guesswork from sequencing.
Compliance teams review components that inherit the same audit grammar across every initiative. Organizations stop treating each AI project as a from-scratch validation exercise when shared evidence and rule versions already exist. The governing principle holds throughout. AI proposes, and humans dispose. Submission to a health authority, approval of a claim, and release of a batch record remain human-gated with a recorded rationale. That is the design regulators expect, not autonomy with a disclaimer buried in the user manual.
Auto audit replay, version control of rules, correlated cross-functional cases, and third-party chain governance together convert compliance from a high-stakes scramble into a predictable, repeatable export. Programs built with strong model accuracy should not be shelved because the evidence story was never told. TFives is built to make sure they are not.
Follow Tarini Mohapatra on LinkedIn or visit TFives for more insights on AI regulatory compliance, governance infrastructure, and building the evidence architecture that gets life sciences AI to production.
