Key takeaways

• AI should earn the right to scale. Move beyond pilots only when accuracy meets a predefined business threshold, leadership ownership is clear and incentive compensation (IC) experts trust the output enough to reduce manual review.
• Fix the IC foundation before adding more AI. Standardized processes, reliable data and connected workflows allow AI to create value across the IC cycle. Without them, AI can scale inconsistency.
• Keep humans accountable for fairness-sensitive decisions. AI can inform goal design, but credible goals require human judgment, review and accountability.
• Measure AI by better IC outcomes, not more AI activity. Faster answers, cleaner payouts and more credible goals matter more than the number of pilots or tools deployed.

AI experimentation is moving quickly in IC. The harder question isn’t whether AI can help. It’s when a use case has earned the right to scale without undermining accuracy, governance or field trust.

That requires more than a successful pilot. In the discussion, the conditions for scale included agreed-upon accuracy thresholds, clear leadership ownership, standardized processes and enough confidence from IC experts to reduce manual review. The goal isn’t more AI activity. It’s better IC outcomes.

In ZS’s webinar, “Inside the AI shift in incentive compensation,” Prateek Sinha, director of Future Readiness, Launch and Field Operations at Novartis, and Anshul Khariwal, principal of customer models and motivation at ZS, discussed what’s working today, what it takes to move from experimentation to adoption and how to build trust as AI takes on a larger role in IC. Watch the webinar to get the full insights.

Anshul Khariwal (AK): Where are most organizations today in their AI journey within IC?

Prateek Sinha (PS): I believe most organizations are still in the experimentation phase. We should be comfortable saying that out loud because pretending otherwise sets the wrong expectation. The conversations have matured. The intent is real, but when you look beneath the surface at whether AI is embedded in our day-to-day IC workflow, whether teams trust the outputs and whether governance is in place, most of us are still building the foundation.

Novartis is no different. We’re in a structured experimentation phase right now where we’re training our workforce across multiple programs, actively encouraging adoption and running use cases in IC. But I wouldn’t call that scaled adoption yet. We’re being deliberate about it, which I think is the right position to be in.

The organizations that will win aren’t the ones who move fastest. They’re the ones who build trust in the AI outputs before they scale them.

AK: What does it take to move an AI use case from experimentation to adoption?

PS: For us, it comes down to three things converging. You need all three. Any two out of three wouldn’t be enough.

First would be accuracy at the threshold that the business has agreed to—not the accuracy that looks good in the demo. It has to hold up across the full IC cycle, multiple therapy areas and edge cases that the pilot probably didn’t surface. We set the bar before the pilot starts, not after.

The second would be formal leadership sign-off after the pilot. It needs to be a documented decision with a named owner.

The third—and this is probably what I’ve seen most organizations underweight—is that IC experts have to trust it enough to reduce their manual review. Not eliminate it, reduce it. If your most experienced IC person is still checking every single output line by line after six months of piloting, AI hasn’t really been adopted.

As for when to move on from a use case, if after a defined pilot period you’re explaining away why the accuracy isn’t where you need it, we should move on. Sunk cost is the enemy of a good AI adoption portfolio.

AK: Where can AI improve IC outcomes most meaningfully today?

PS: Value looks different depending on where you’re in your journey, and conflating phase one wins with phase three ambition is how organizations set themselves up for disappointment.

At Novartis, we’ve organized our AI roadmap across three domains: IC operations, IC analytics and self-serve, each with a phase progression from assisted execution toward genuine intelligence.

Where we’re seeing real value today in IC operations is parallel validation of IC calculations, AI-assisted eligibility, report creation and early-phase data anomaly detection. Payout errors erode field trust instantly. An AI layer that catches inconsistencies before they reach the field is worth more than any other headline use case.

In IC analytics, we’re using AI to generate health-check insights, pulling together performance signals, exceptions and trends that previously required a significant amount of time. We’re building it into quota-setting simulations, running modeling across therapy areas simultaneously in ways that were simply not feasible manually without AI.

On the self-serve side, we’ve started with internal standard operating procedures and process support. The next phase typically would be a field-facing chatbot that can answer questions on historical reports, calculations and data, with appropriate governance on what it can and can’t resolve autonomously.

The through line across all three is a guiding principle: the human should be in the loop for the decision. It should be explainable and compliant, and business value should take priority over experimentation.

AK: As organizations move from experimentation to scaled impact, what pitfalls should they look out for?

PS: Standardization has to come before scale.

If our processes are inconsistent, AI will scale inconsistently. If our data are unreliable, AI will automate unreliability. So the organizations that skip the standardization step and jump straight into AI pilots could be the ones who end up with expensive proofs of concept that never graduate to production.

Do the unglamorous work first, fix the foundation and then AI has something real to build on.

AK: How do you decide where to apply AI in IC—and where not to?

PS: The where-not-to question is the one I find more interesting and far more important to answer. My framework has been that AI earns the right to operate in a domain. You earn that right by providing accuracy, building the governance and getting field trust phase by phase. All of this doesn’t happen at once.

This is how we’ve structured our roadmap as well. Phase one is assisted execution: AI helps teams do faster, better work. Phase two is guided decision-making. And phase three, for select domains, is where AI starts to act with greater autonomy.

But there’s one domain that I’ll probably never fully move to phase three: AI-generated goal design for field reps. AI can synthesize quantitative signals. But it can’t weigh a territory manager’s judgment about why a geography underperformed, account for a rep’s first year on a complex brand or carry the accountability when a rep looks at the goals and says, “This is not fair.”

Goal design sits at the intersection of business strategy, market dynamics and human fairness. A rep’s livelihood depends on whether their goals are credible. That requires human expertise and human accountability. AI can inform those things, but it can never really own them.

The moment you say AI set goals without a rigorous human review layer is when we’re giving away or outsourcing our accountability to an algorithm. That erodes field trust faster than anything else I can think of.

AK: Where have you seen the biggest friction when introducing AI into IC, and what has helped overcome it?

PS: All three of the classic friction points. First is leadership friction. It’s about ROI visibility. Leaders want to know what they’re getting before they commit, but you can’t always show an outcome before you’ve done the work. So the way through this is ruthless prioritization. You pick a use case where the value is tangible quickly, and build credibility there before asking for broader sponsorship.

Next is governance friction. It’s probably subtler, but more dangerous. IC sits at the intersection of compensation, compliance and field trust. For AI outputs that feed into decisions, you need an audit trail, a human sign-off layer and a clear escalation path. Organizations that treat governance as an afterthought in an AI rollout end up rebuilding it under pressure, which, in my view, is far more costly.

The third is field readiness. It’s the one that keeps me up at night. Field reps are skeptical of anything that touches IC, and quite rightly so. The way to overcome it isn’t to explain AI. We learned that the hard way. It’s to show them a better outcome.

When a rep gets a faster, clearer answer to an IC query or sees that their goals look credible because the model was validated, trust follows the result, not the technology.

AK: How should organizations measure AI success in IC?

PS: The AI capability is an input. The IC outcome is the only measure that actually matters.

When we defined objectives in our teams, we were deliberate about what we were trying to achieve: shift IC from pure execution to a data-driven strategic partner, reduce manual effort as much as possible, reduce cycle times and enable a self-serve model so field reps can get answers without going through a queue and spending weeks getting them. Those are the measures we hold ourselves to—not the sophistication of the models or the number of use cases deployed.

The trap I see teams and organizations fall into is measuring AI investment by AI activity: how many pilots we ran, how many tools we deployed. This is measuring motion, not progress.

The question I ask when speaking to my leadership team is simpler: What did a field rep experience differently this cycle because of what we’ve built? Did they get their IC answer faster? Did the payout come through cleaner? Did the goal feel more credible?

If the answer to these questions is yes, I believe AI is working. If not, we probably have a capability without an outcome. That’s the most expensive kind of investment to defend.

What should leaders ask before scaling AI in IC?

Across the discussion, the same principle holds. AI shouldn’t scale in IC simply because the technology is available. It should scale when the organization can show that the output is accurate, governed and trusted—and when the field experiences a better outcome.

As leaders decide what to scale next, three questions are worth carrying forward:

• Did this use case meet the accuracy threshold we set before the pilot?
• Are IC experts confident enough to reduce manual review?
• What will field reps experience differently if we scale it?

Watch the webinar to get the full discussion on how AI is changing IC and what organizations need to move responsibly from experimentation to scale.