Interview with Shweta Doshi from Mindtickle

1. While building agentic systems, there are two non-negotiables that often crop up-reliability and error handling. As the Head of Product, what have you found to be critical in ‘designing for trust,’ especially when it’s an agent making customer-facing decisions down the line?

Trust is not a feature you add to an agent. It is the architecture you build it on.

When an agent makes a customer-facing decision, coaching a rep on a live deal or surfacing a competitive play, reliability is not a feature. It is the architecture.

Three non-negotiables for us:

Governed autonomy – Agents observe, reason, and act, but inside explicit boundaries. High-stakes actions like publishing a playbook to thousands of sellers require human approval. Low-stakes actions run on their own. The boundary is configurable, not hard-coded. That is the difference between an agent a customer trusts and one they switch off.

Grounding over generic intelligence – Every AI judgment anchors to the organization’s own definition of excellence, their methodology, their winning behaviors, their deal patterns. An ungrounded model gives generic answers.

Traceability at every step – Every decision carries a visible reasoning trace. When a deal agent flags risk, it says why: 30 percent higher risk, recent inactivity. Trust is earned through transparency, not black-box confidence. We design for uncertainty to be visible.

2. There’s a substantial tension between long-term product vision and the need for agility in a highly volatile tech landscape. When businesses must modify their tech stacks to keep pace with innovation, how do you iterate your multi-year tech roadmap for immediate agility?

In today’s AI landscape, we don’t build multi-year roadmaps in the traditional sense. We have a clear long-term vision of where we’re headed, but we architect for evolvability and treat the roadmap as a living document.

When we transformed Mindtickle into an agentic AI platform, the ground was constantly shifting beneath us.New models every quarter. New interoperability protocols emerging. A rigid three-year plan would have been dead on arrival.

Here’s how we manage that tension:

Platform-first, feature-second – The layered architecture evolves independently. Swap a model at the gateway without touching product code. Add a protocol at the tool layer without rewriting agents.

Ship in concentric circles – Based on customer jobs to be done – ship the highest value first , then adjacencies. Each new agent reuses the same building blocks

Standardize the integration surface – We adopted MCP to avoid building one-off connectors for every system. When a customer’s stack changes, we adapt in days, not quarters.

The roadmap is a portfolio of bets across time horizons. Place architectural bets for where the market is going, not where it is.

3. As Mindtickle’s Head of Product, you lead Product, Design, and Product Analytics- where alignment and accountability are of the essence in innovating proactively. Do you have a practicing philosophy while helping your product and engineering teams avoid burnout through intensive, ambiguous workloads?

Intensive work does not break people. Fragmented work does.

The teams that burn out are not the ones working hard on one big problem. They are the ones context-switching across five, each half-defined, none finished. The cognitive load lives in the switching, not the intensity. So the discipline is not about working less. It is about protecting focus.

4. Can you walk our readers through the North Star metrics that business leaders must prioritize in this AI age- ones that truly reflect business health and not merely feature usage?

Feature-usage metrics in AI products are dangerously misleading. A customer can show high adoption on the dashboard and zero change in the outcome the product was bought to move.

The metrics business leaders must prioritize instead:

Behavior change, not activity – Not whether a user engaged with the product, but whether their behavior in the real workflow actually shifted. The gap between product interactions and real-world behavior change is the only honest measure of effectiveness.

Outcome velocity – Time to the result that matters. AI should shorten the time from first use to consistently better outcomes. If your AI features are not measurably accelerating the outcomes, they are just sophisticated toys.

Consumption of value as a leading indicator – As the market shifts toward usage-based models, value consumption becomes a shared North Star between Product and Customer Success. The question isn’t whether customers logged in, but whether they consumed enough AI-driven value to justify and expand the investment. Declining consumption predicts churn earlier and more accurately than NPS.

5. With the enterprise tech domain extensively saturated with similar products, pricing remains a stark buying metric. Do you think there’s a future where pricing points will lean more towards value-based models defined by the level of autonomy an agent exercises?

Yes, and it is already happening. The market is moving to usage and outcome-based models, and we are leading that shift with an AI Credits model: a common currency across every AI offering, where customers pay for value consumed rather than seats purchased.

The deeper point is your framing. Pricing will increasingly reflect the level of autonomy an agent exercises. An assistant that suggests a next step is worth less than an agent that does the work. As autonomy rises, so does the value created per interaction, and pricing should track that, not seat count.

6. You’ve spoken about designing and building habit-forming products. With buyers increasingly aware of behavioral engineering behind such models, how do you cultivate credibility around cognitively demanding solutions?

The credibility shift is from engagement to earned habit.

An earned habit forms because the product makes the user better at their job, not because it exploits a dopamine loop. A rep returns to an AI roleplay because the last one exposed a real gap before a real call and acting on it produced a measurable improvement. A manager trusts the deal agent because its risk signals have consistently proven accurate. The habit is a consequence of value, not a substitute for it.

Three things keep this honest. First, tie every habit to a measurable business outcome, not a vanity metric. If daily usage rises but quota attainment does not, the habit is hollow, and the buyer will see through it. Second, make the value legible in the moment, so the user understands why the recommendation is worth acting on. Third, reduce cognitive effort without removing human judgment. AI should make complex work feel lighter, not replace critical thinking.

The goal is not more time spent in the product. It is less time spent reaching a better decision.

7. The increase in AI tax is inevitable, especially with the cost of computing and inference hitting new highs every single week. Can businesses continue to maintain attractive price points for enterprise adoption while sustainably absorbing these costs? Or is a market-wide price correction inevitable?

I see this differently. Inference cost per unit of intelligence is not rising. It is collapsing. The cost to run a capable model has fallen dramatically year over year. What continues to rise is total AI spend because we’re asking models to do more. . That is a very different problem from broken unit economics.So the question is not whether businesses can absorb an ever-growing AI tax. It is whether they architect to ride a falling cost curve.

The first move is to decouple price from raw cost. We price on value delivered, not on tokens consumed, so when inference gets cheaper, the customer experiences more value rather than feeling a tax. The second is to route intelligently, because not every task needs the most expensive model.A well-architected platform sends simple judgments to the cheap model and reserves the frontier model for high-stakes reasoning, and because the gateway layer makes this swappable. The third is to make intelligence compound, so when the system learns from every call, it reaches a better answer with less compute over time.

The platforms that win are the ones that turned falling inference cost into rising customer value rather than protected margin.