The Intelligence Dividend: How Enterprise Tech Leaders are Earning Returns From AI

Every enterprise is paying an AI bill. Far fewer can show the dividend. The latest episode of the iOPEX leadership series asks what separates the two: where AI actually earns, how to choose the right workflows, and what it takes to move from pilot to production.

The conversation features Aparana Gupta, Director, Software Engineering at Microsoft, who leads employee experience solutions built on behavior analytics and agentic AI running at enterprise scale. Ravi Kusuma, SVP at iOPEX Technologies, hosts the discussion, and Ashish Sopori, Senior Director, Marketing at iOPEX Technologies, opens the episode and introduces the theme.

Right Information, Right Place: Where AI Earns

The conversation opens on a simple premise: when the right information reaches the right place, AI can act on it. But every organization is now invested in AI, and the value gap remains. How are we defining ROI, and are we getting enough value out of it? Aparana’s answer is three horizons: efficiency, effectiveness, and optionality. AI ROI is a combination of all three, weighed against computation and engineering cost, governance, and operational risk, and the value the workflow itself generates. It is also shaped by competing priorities, which is why the decision framework matters.

Choosing the Workflow: The VFIT Framework

What thought process decides which workflows benefit from AI first, and what separates a high-yield candidate from a science project with executive sponsorship? Aparana described what we’ll call VFIT: Value, Friction, Intelligence advantage, and Trust. The strongest candidates are workflows where intelligence is slow, expensive, and inaccessible today.

She also placed humans at three points in the system:

“Human-in-the-loop governs action, human-on-the-loop governs the system, and human-in-the-design governs the intent: shaping objectives, failures, constraints, evaluations, and escalations. That is the most important one, the intent at play.”

Pilots Test Capability. Production Tests Reality.

What does running enterprise AI in production teach you that no amount of piloting could? A pilot tests capability, and production tests reality. The mindset shifts from model-centric engineering to system-centric engineering, where the goal is predictable behavior within understood boundaries and constraints.

That raises the question of what breaks first when a workflow graduates: the data architecture, the governance model, or the security and privacy posture. Aparana’s view is that governance, including guardrails such as PII protection, should not become a bottleneck. At scale, good governance creates certainty. Ravi captured it this way:

“Within a governance model, the actualization of AI is easier.”

Building Trust Before Taking It to Customers

Employee experience is where these ideas are being tested in production. Making employee-facing AI real beyond the pilot, and then taking those changes to customers, starts with building three types of trust: capability, reliability, and intent, the last of which is organizational. Enterprise leaders can borrow this sequence: earn trust internally first, then extend the experience outward.

Feedback Loops: Art and System Design

How do you engineer a feedback loop so that what you learn in production dictates the next wave of AI investment? On the surface it looks like a thumbs-up or a one-line comment, but underneath there are mechanisms and trade-offs. Aparana outlined four types of feedback: explicit, behavioral, outcome, and system signals such as latency, overrides, and hallucinations. Designing a better feedback system is a combination of art and system design, and it shows up in both the architecture and the org chart.

Memory Architecture: Compounding, Not Isolated, Intelligence

What has to be true of an enterprise’s data and memory architecture for intelligence to compound rather than reset with every new project? That, Aparana argued, is the real AI dividend: compounding intelligence rather than isolated intelligence. Data connects to context, memory, semantics, metadata, ontology, and policy, which leads to action and feedback. Enterprises need a reusable semantic foundation, retrievable identity and lineage, governed memory, and interoperable knowledge layers. Once the first results are net positive, the advice on scaling is clear: don’t stop at one successful output, extend it to 50 more use cases.

Looking Ahead to 2027

Imagine it is early 2027 and someone replays this episode. What will have separated the enterprises that earned an intelligence dividend from those that simply paid an AI bill? Aparana pointed to differentiators built around intelligence: feedback loops will matter most, followed by governance, as enterprises ask how to operate around AI across systems, people, and processes, and how to measure the value it creates.