Yogesh Joshi, SVP of Global AI Platforms at TransUnion, explains why individual productivity gains from AI often stall before they reach the enterprise—and why the fix is redesigning workflows around business outcomes, not just adding AI to the tasks already in place.
Ask any engineering leader whether AI has made their development team more productive, and the answer is almost always yes. Employees can draft content, analyze data, write code, and automate routine work in a fraction of the time it used to take. Ask whether the organization is moving faster as a result, though, and the answer gets murkier. The output at the keyboard has changed, but the business throughput often hasn't.
That disconnect is the real AI productivity paradox: individuals are moving faster, but organizations aren't capturing the gains at scale. The reason isn't the technology. It's the operating model. Sustainable AI impact doesn't come from making tasks faster. It comes from improving how work moves across functions and through the enterprise—connecting the people, processes, data, and decisions that make work possible.
Nowhere have I seen this more clearly than in application development, where I spend a good deal of my time as SVP of Global AI Platforms at TransUnion, a global information and insights company spanning credit reporting, fraud prevention, and marketing analytics. The pattern isn't unique to writing code—any workflow with its own approvals, handoffs, and dependencies runs into the same wall. What follows is what we've learned: why individual speed doesn't automatically translate into organizational speed, how embedding AI into a workflow rather than bolting it on changes the math, and what to measure so leaders can see what’s changed.
Faster Work, Same Bottlenecks
Individual productivity gains from AI run into organizational friction almost immediately.
Organizations still operate through layers of approvals, disconnected systems, and functional silos. AI speeds up the work within the workflow. It doesn't automatically fix the workflow itself.
We saw this at TransUnion with an internal tool that helps engineers generate code and create tests. It delivered a 25 to 30 percent improvement in developer productivity. But code isn't the outcome of development. A production release is. And releasing software into production depends on sign-offs across multiple dependent steps: confirming existing features still work as expected (regression testing), verifying the software meets user needs (user acceptance testing), a security and vulnerability review, product approval, operational readiness, and coordination across engineering, quality assurance, product, information security, risk, operations, and release management. Each of those groups can approve or hold a release; the decision about whether and when to ship rarely moves as fast as the code that made it possible.
We've seen the same pattern elsewhere: a recent platform release finished full regression testing weeks before it shipped and waited on enterprise risk review and cross-functional release governance. The engineering work was done; the critical path had moved beyond software development. AI can accelerate creation. It can't remove those dependencies.
From Add-On to Architecture
The next wave of productivity will come from AI becoming part of how work gets done, not from occasional copilot use.
That's the thinking behind our AI analytics orchestrator, an agentic AI system that unifies the end-to-end analytics workflow into a single, coordinated process. A lender might use it to predict the likelihood a borrower defaults, helping balance access to credit with portfolio risk—or apply the same workflow to fraud-risk scoring or marketing audience selection. Previously, data scientists moved through separate steps to interpret a business question, identify data and attributes, assemble analytical code, execute the model, document the work, and prepare it for review. Faster code generation sped up one step, but the handoffs and repeat work remained.
We redesigned the workflow around reusable and auditable analytical templates. The orchestrator converts a natural-language request into a governed code-template workflow. It maintains human review and creates a repeatable path from question to analytical output. We didn't just add an AI assistant to the analyst's desktop. We embedded AI, domain context, controls, and documentation into the workflow itself.
For selected model-development workflows, what once took 45 to 65 hours now takes approximately 1 to 1.25 hours, delivering a 45x to 52x faster development time with performance parity to benchmark models. These figures come from observations across multiple customer engagements and measure the analytical workflow segment, not the full customer lifecycle. But they capture the difference between accelerating a task and redesigning a repeatable workflow.
Measuring What Matters
Many organizations focus on adoption metrics: AI prompts sent, AI tool licenses deployed, code produced, or theoretical hours saved. Those numbers don't prove the enterprise is moving faster. The better question is whether AI improves customer outcomes, accelerates decisions, reduces risk, increases resilience, or drives growth.
The orchestrator results above are a good example: the gains matter because they help the organization make and act on decisions more quickly. Ultimately, the goal isn't to deploy more AI. The goal is to create more value.
The Operating Model Is the Product
AI can't fix a broken operating model. The organizations seeing the greatest productivity gains are rethinking their structures. They're building cross-functional teams around business capabilities, reducing dependencies, embedding governance earlier, and establishing common measures of success. AI is an accelerant, but transformation comes from aligning people, process, data, and technology.
The practical leadership question should be "What constraint governs the end-to-end outcome, and how do we redesign the workflow around it?" If AI makes one contributor faster while work still waits for data, approval, validation, or release, the gain stays local. If the operating model changes so that controls, decisions, and evidence are built into the flow of work, the gain becomes organizational.
The real issue isn't that AI doesn't work; it's that AI is working faster than many organizations can adapt. The next wave of value will come from an organization’s ability to reimagine how work gets done.
Written by Yogesh Joshi
Yogesh Joshi is senior vice president of global AI platforms at TransUnion, where he leads OneTru, the company's core AI and data platform. Before TransUnion, he held data, analytics, and identity leadership roles at Neustar, AIG, and Experian. He holds a master's degree in management from UCLA Anderson School of Management.