My work sits at the intersection of AI product management, AI strategy, and decision intelligence.

Over the last 10+ years, I’ve built AI products across startups and global technology companies, from early natural language processing work in Germany to scaling AI products at HubSpot and Intuit. I focus on turning messy, unstructured information into products that help people make better decisions and take action, where trust, reliability, and outcome quality drive adoption.

At HubSpot, I evolved the sales CRM experience from system of record to AI-guided selling, driving 20+ points of adoption improvement. At Intuit, I incubated and shipped an AI Customer Agent from 0→1 as part of its AI-native CRM initiative, scaling it to 1,000+ users and improving measured accuracy from roughly 60% to 90%, while supporting workflows tied to $400M+ in SMB business opportunities.

At Intuit, I led a P0 AI-native CRM initiative, taking an AI customer agent from zero to one.

The problem was simple: SMB operators were losing leads in their inboxes because they didn’t have the time to manage the customer lifecycle. We hypothesized that an AI agent could identify and qualify those leads, and ultimately engage customers on the seller’s behalf.

I deliberately started narrow. We built a single prompt-based experience to validate the core use case before investing in a broader agent architecture. Once we validated it, we expanded into orchestration across lead identification, prioritization, qualification, and engagement.

I partnered closely with engineering and our AI teams to build the system and the evaluation infrastructure around it. We started with human evaluation, built a golden dataset of roughly 1,000 lead emails, and introduced LLM-based evaluators. That took accuracy from roughly 60% to 90%.

Within three months, we scaled to 1,000+ users, with customers telling us the agent surfaced leads they otherwise would have missed.

The broader outcome was establishing the architecture and evaluation practices for building AI agents at speed—not just shipping one AI feature, but creating a foundation for scaling agentic products.

I start with the outcome, not the model.

I think about an AI product as a system that transforms a goal and a representation of reality into better decisions, actions, and ultimately outcomes. So before I think about agents, prompts, or models, I work through six questions:

Goal → Governance → Capability → State → Observability → Feedback

1. Goal — What outcome are we trying to improve?

First, I identify the user’s goals and choose the primary goal. That goal becomes the most important constraint on the system. It determines which workflow matters, which decisions are valuable, and what success looks like.

For example, in revenue, the goal might not be “summarize calls.” It might be improving win rate, reducing sales cycle, or helping a seller decide where to focus.

2. Governance — What should the system be allowed to do?

Then I define the system’s boundaries: permissions, decision rights, policies, escalation, and where humans need to remain in control.

This is particularly important with AI because capability can exceed what we should actually delegate. I want to distinguish between what the system can do and what it should do.

That leads naturally to progressive autonomy: the system can start by informing or recommending, then earn the ability to take more action as reliability and trust increase.

3. Capability — What intelligence is actually required?

Only then do I determine what the system needs to be capable of.

That might include reasoning, prediction, synthesis, prioritization, planning, recommendation, tool use, or execution. I don’t start with “Where can I put an LLM?” I start with the decision or workflow problem and determine what intelligence is actually necessary.

4. State — What does the system need to know about the world right now?

This is where I define the system’s world model. I think about it as:

Context + Memory + State

• Context is the relatively stable background — who or what something is.
• Memory is history — what happened previously.
• State is current reality — what is true right now.

That distinction matters because the same context and history can produce completely different decisions depending on current state. A good AI system therefore isn’t just retrieving information; it’s maintaining a representation of the world that is relevant to the decision.

5. Observability — How do we know what the system is doing and whether we can trust it?

For AI systems, observability has to go beyond traditional system health. I want to know:

• What did the system see?
• What did it infer?
• What evidence supported the inference?
• What decision did it make?
• What action did it take?
• How confident was it?
• What happened afterward?

That creates the foundation for explainability, evaluation, and trust. Critical calculations can remain deterministic, while inferential components are explicitly evaluated and appropriately gated.

6. Feedback — How does the system get better?

Finally, I design the learning loop.

The action produces an outcome. The outcome generates information. That information updates the system’s understanding and improves future decisions.

I think about it as a spectrum of intelligence and autonomy, based on what the workflow actually requires.

Deterministic systems work best when the logic is well understood and precision matters. At HubSpot, for example, Guided Selling used deterministic rules to translate known signals into specific recommended actions—for example, when a deal stage changes to “Negotiation,” recommend “Draft Contract.”

AI-assisted systems are useful when the problem requires intelligence and/or action, but the system needs defined boundaries. With Topic Explorer, I built a conversation-intelligence capability that helped teams understand which topics and conversation patterns were associated with won versus lost deals. We defined the analytical questions and structure upfront, then used an LLM to synthesize and reason over the conversation data within those boundaries.

Agentic systems become useful when intelligence needs to be connected to action. For example, if a deal enters negotiation, an agent could understand the deal context, determine what contract is appropriate, draft it, route it for approval, and coordinate the next steps. The system isn’t just recommending an action—it is acting toward a goal.

This is the right treatment when the workflow requires dynamic reasoning across multiple steps and systems, rather than simply applying a rule or generating an insight. The agent can own the bounded responsibility of moving the contract process forward while governance defines what it is allowed to do and when a human needs to intervene.

I think there are four conditions that determine whether people will adopt AI: relevance, reliability, observability, and control.

Relevance — “Do I want this?” Is this a problem worth delegating—frequent, painful, costly, or error-prone—and is the user comfortable handing it off?

Reliability — “Can I trust it?” Can I depend on the system to perform consistently and correctly, including when it encounters unexpected situations?

Observability — “Can I see and verify it?” Can I understand what the system is doing, inspect the evidence behind its actions, and know why it reached a conclusion?

Control — “Can I intervene?” Can I set boundaries, review consequential actions, override the system, or take back control when necessary?

Reliability + observability + control create trust. Relevance + trust create adoption. And adoption only matters if it improves outcomes.

My background across economics, product management, and applied AI has shaped how I think about technology: I’m less interested in technical capabilities than in how systems improve institutions, expand opportunity, and drive better outcomes.

Economics taught me to think in terms of incentives, institutions, and systems. Product management taught me to turn that thinking into products. And 10+ years building AI products has taught me that AI’s greatest promise is expanding human capability—not simply automating work.

That conviction shapes my product philosophy: AI should strengthen human judgment, not replace it. The best systems help people navigate complexity, make better decisions, and act with greater leverage.

And the strongest systems are built with users, not simply delivered to them. That combination of augmentation, human agency, and real-world feedback guides how I design AI products today.