We’re now accepting applications for PearX, our exclusive small-batch program for pre-seed companies. Our S26 cohort kicks off in July 2026 and runs for 12 weeks. We welcome founders at the idea stage as well as teams already gaining early traction to apply. Below are some of the ideas and opportunities we’re most excited to see founders tackle next.
Plan mode for every knowledge worker
One of the most underappreciated design breakthroughs in AI tooling is a workflow change. Claude Code introduced “plan mode,” where the agent and developer iterate on an approach before any code gets written. This simple forcing function — make a plan, pressure-test it together, then execute — dramatically improves output quality because it surfaces ambiguities, bad assumptions, and missing context before they compound into wasted work. Meanwhile, the chatbots most knowledge workers use today are optimized for agreeability. Ask Claude Chat or ChatGPT to write a sales deck, build a financial model, or draft a marketing brief, and it will happily produce something that looks polished but is built on a stack of unchallenged assumptions about your audience, your constraints, and your goals. The result is output that’s 60-70% of the way there: impressive enough to feel useful, but not good enough to actually ship without heavy rework.
We believe there’s a large opportunity to build vertical AI tools that bring the plan-then-execute paradigm to domains like sales, marketing, and finance. In practice, this means an agent that behaves less like an eager intern and more like a sharp collaborator: one that asks clarifying questions, flags where it’s uncertain, proposes a structured plan of attack, and only moves to execution once the human has weighed in on the approach. For a sales team, this might look like the agent and rep co-developing an account strategy before drafting outreach. For a finance team, it could mean aligning on assumptions and edge cases before building a forecast.
This is a wedge that could evolve into something much bigger. A tool that earns trust in the planning phase naturally becomes the system of record for how decisions get made, and eventually the agent that executes on them autonomously. If you’re building the agentic workflow layer for a specific vertical, we want to talk.
Financial-grade agent infrastructure
By Ryan Sells
AI agents are already useful. But they’re useful in the way a smart advisor with no signing authority is useful. They can research, summarize, and recommend, but the moment you need an agent to actually move money, execute a contract, or make a credit decision, the entire stack breaks down. Today’s agent infrastructure was built for a world of reversible, low-stakes actions: drafting an email, writing code, searching the web. But financial systems demand auditability, policy enforcement, liability attribution, and the ability to handle irreversible actions where mistakes aren’t fixable with an undo button. The gap between “ChatGPT-style agent that suggests things” and “autonomous agent that can execute in production financial systems” is enormous.
We’re looking for founders building the trust, policy, and execution layer that makes AI agents safe to deploy in financial contexts. This could span identity and authorization frameworks that let agents act on behalf of humans with scoped, revocable permissions; policy engines that enforce compliance constraints and spending limits before an action is taken, not after; audit and explainability infrastructure that satisfies regulators and counterparties; or composable transaction primitives that let agents interact with banks, payment rails, and financial APIs with the same reliability we expect from production software. The design space is rich because every financial vertical — payments, lending, treasury, insurance, procurement — has its own set of regulatory and operational requirements, but the underlying need is shared: a common infrastructure layer that makes agents both capable and trustworthy.
The companies that build the trust layer for agentic finance will sit at the center of an enormous amount of economic activity as agents move from advisors to actors. The teams that solve this will define how AI participates in the real economy. If you’re building here, we’d love to hear from you.
Verified consumer health marketplaces
Consumer health is flooded with products that promise results but rarely prove them. Supplements, protocols, devices, and programs are marketed with glossy claims, thin evidence, and little accountability. Meanwhile, real health outcomes are deeply personal, heterogeneous, and dynamic — what works for one person may do nothing for another. The core failure is the absence of a feedback loop that connects interventions to measurable outcomes in the real world.
We now have the technical ingredients to close that loop. Continuous biomarker measurement, at-home diagnostics, longitudinal health data, and AI-driven analysis make it possible to evaluate efficacy at both the individual and cohort level. But today, this data is fragmented, underutilized, and rarely integrated into the actual marketplace where health decisions are made. The result is a system optimized for distribution and marketing.
We’re looking for founders building verified consumer health marketplaces where products, services, or protocols are evaluated based on real biological impact. This could include platforms where vendors are rewarded for measurable outcomes, where consumers make decisions based on verified response data, and where consented health data becomes a moat rather than a liability. The companies that introduce verification and accountability into consumer health will reshape how trust is earned and how health markets function at scale.
Ambient AI for vertical workflows
Jump AI went from zero to $10M ARR in 18 months by building what amounts to Granola for wealth managers. They charge $100/month per user (vs. Granola’s $10) because they have deep context about conversations as they happen and are fully integrated into the wealth manager’s actual workflow. Abridge did something similar in healthcare: transcription plus ambient intelligence plus workflow integration. The pattern is clear, and it works because these tools don’t just record meetings — they understand the professional context around them and can act on it.
We want to fund founders applying this ambient AI pattern to other verticals. The playbook is: listen to the actual conversations a professional has, connect that to their existing data and tools, and surface useful context in real time. The value is in domain-specific integration, not generic transcription. Which industries have expensive professionals spending time in meetings where context from other systems would change the quality of their decisions? We think there are many, and the 10x pricing premium Jump commands over Granola shows how much value vertical specificity creates.
Selfishly, we’d love one for venture. An ambient agent plugged into our Airtable, Crunchbase, and past meeting notes that can whisper “you saw a similar company six months ago” or “that data point doesn’t match what Crunchbase shows” during a live pitch. If you’re building the ambient workflow layer for any professional vertical, we want to hear from you.
Foundational datasets for biological intelligence
The defining bottleneck in AI-driven biology is data. In language and vision, large curated datasets like ImageNet and Common Crawl were sufficient to train foundation models because the core tasks are largely single-modality. Biology is harder: the problems that matter (drug discovery, disease understanding, clinical prediction) are inherently multimodal, requiring connections across sequence, structure, function, and clinical outcome. Biology has deep datasets in individual modalities, but nothing that connects them.
We have one proof point for what happens when a single modality gets a good enough dataset: the PDB yielded AlphaFold for protein structure prediction, and the next generation of tools (like Isomorphic Labs’ IsoDDE) is extending that into binding affinity prediction, pocket identification, and molecular generation. But even the most advanced of these systems stops at molecular interaction. Downstream questions (stability, toxicity, pharmacokinetics, patient selection, clinical response) remain largely disconnected from the structural breakthroughs upstream. The existing landscape of large biological repositories (e.g., UniProt, GEO, Human Protein Atlas, etc.) reflects the same limitation.
We’re looking for founders building the canonical training datasets for biological foundation models: companies generating proprietary human biological data at scale (multiplexed perturbation screens, longitudinal multi-omic cohorts, paired genotype-phenotype datasets), or infrastructure that integrates and harmonizes existing fragmented human data into model-ready representations. We’re especially excited about teams that already have a proprietary dataset in hand, have rigorously benchmarked it against meaningful predictive tasks, and can credibly demonstrate that performance improves with scale. If you’re building the foundational data layer for biology, we want to hear from you.
Systems of record → Systems of coordinated action in healthcare
Clinical care today is dominated by systems of record. EHRs, billing platforms, claims systems, and registries store enormous amounts of data, yet care itself still depends on people stitching systems together. Clinicians and operators act as the connective tissue, moving information across platforms, reconciling inconsistencies, and deciding what happens next. The result is delays, preventable escalations, administrative burden, and billions in avoidable cost.
The first generation of healthcare software digitized documentation. The next generation made that data searchable and analyzable. What has not changed is execution. Work still fragments when it crosses teams, organizations, and payment models. Healthcare is also inherently multi-party. Providers, nurses, therapists, administrators, payers, revenue teams, vendors, and patients each operate within distinct incentives, permissions, and regulatory constraints. Today, AI tools are deployed within these silos. Insights surface, yet responsibility remains distributed. When action requires coordination across entities, momentum slows and accountability diffuses.
The next shift in healthcare software is coordinated execution. Systems that move work forward across stakeholders, route tasks across care teams, reconcile documentation before claims are denied, escalate clinical risk within guardrails, and synchronize updates between providers and payers in real time. Execution becomes part of the system itself rather than dependent on individuals chasing context.
In a reimbursement-shaped and highly regulated environment, alignment matters as much as intelligence. The companies that win will build inspectable, compliant infrastructure that integrates deeply into clinical and financial operations and ensures that intent becomes aligned action across the system.
If you are building infrastructure that turns fragmented healthcare processes into coordinated healthcare action, we want to talk.
IRL trust networks
Digital systems were never designed for a world where seeing is no longer believing. As generative AI advances, trust online is actively decaying: deepfakes, synthetic media, and agent-driven interaction make it increasingly difficult to know whether a person, action, or relationship is real. Identity can be forged. Content can be fabricated. Reputation can be manufactured at scale. The cost of deception is falling faster than the cost of verification.
Meanwhile, real trust is still built the old-fashioned way — by showing up in person, spending time together, following through, and being accountable over repeated interactions. But almost none of this real-world trust is legible to our digital systems. Today’s platforms rely on brittle proxies: profiles, ratings, social graphs, and credentials that are easy to fake and hard to port. As a result, the signals that matter most in an AI-saturated world — reliability, presence, and shared experience — are invisible where decisions are actually made.
We’re interested in founders building IRL trust networks: infrastructure that captures trust earned through real-world interaction and makes it portable, composable, and economically meaningful. Not identity verification, and not social media — but systems that translate lived experience into durable trust signals others can rely on. This could underpin access to jobs, housing, healthcare, lending, marketplaces, and communities where stakes are high and mistakes are costly. As digital trust erodes, the value of verifiable, IRL-earned trust only increases. The teams that build this layer won’t just create products; they’ll help restore trust in a world where authenticity is no longer guaranteed.
Intent-native commerce infrastructure
By Ryan Sells
Every layer of modern commerce – catalogs, pricing, merchandising, checkout – was designed around a human navigating a screen. AI agents don’t browse. They take intent (“reorder packaging supplies from the cheapest reliable vendor”) and go directly to sourcing, evaluating, and transacting. This breaks the existing stack. Catalogs are structured for browsing, not programmatic reasoning. Pricing assumes static SKUs, not dynamic negotiation against a buyer’s constraints. Checkout assumes a human with a credit card, not an autonomous actor with scoped authorization. The entire infrastructure needs to be rebuilt around intent, not clicks.
We want to fund founders building commerce infrastructure where agent-mediated transactions are first-class. Storefronts that function as intent-aware APIs rather than visual catalogs. Pricing that can be dynamically constructed based on context and constraints rather than fixed price tags. Discovery layers where relevance is determined by fit-to-intent rather than SEO or ad spend. As agents become the default purchasing interface, merchants who can’t sell to agents will lose distribution the same way businesses that couldn’t sell online did two decades ago. The infrastructure powering this transition is a generational platform opportunity, and it’s early.
Personalized people search for the agent era
The explosion of AI agents is creating a massive, underserved need: better people data infrastructure. Today’s people search and enrichment platforms — companies like People Data Labs, Clearbit, and ZoomInfo — were architected for a world where a human salesperson or recruiter ran a query, eyeballed the results, and applied their own judgment to fill in gaps. That workflow tolerates stale records, generic filters, and flat profiles. But when an AI agent needs to decide who to contact, how to personalize a message, or which expert to loop in, the bar changes completely. Agents need real-time, contextual, structured people data they can reason over. Static CSV exports designed for a human in the loop don’t cut it.
We want to fund the team building a people search and context layer for agentic workflows. Think: an API that doesn’t just return a name and LinkedIn URL, but synthesizes someone’s recent work, public writing, professional network, and inferred intent into a representation an agent can actually use. This could be a new data platform with novel collection methods, a middleware layer between agents and existing sources, or a protocol that lets agents query and exchange people context with each other. The consumer of this data is now a model making decisions at scale, and the data infrastructure should reflect that.
We’d especially love to back a team that understands both the data engineering challenge and the emerging agent ecosystem. There are hard problems on both sides: getting fresh, structured data at scale is an engineering grind, and building representations that agents can reason over well requires deep familiarity with how these systems actually work. If you’re building in this space, reach out.
Agents to manage business expenses
SMBs leak money constantly across insurance, rent, lines of credit, taxes, utilities, and dozens of other recurring costs. In most cases, nobody is actively managing these expenses because the savings on any single line item don’t justify the time. But in aggregate, the waste is real. The tools that exist today require a human to initiate each comparison or negotiation, which means it mostly doesn’t happen. A ten-person company doesn’t have a procurement team, so bills go unexamined and contracts auto-renew at whatever rate the vendor set.
We want to fund founders building agents that run in the background and continuously look for savings across every category of business spend. Not a dashboard that shows you where you’re overpaying — an agent that actually renegotiates your electric bill, finds a cheaper insurance policy, or flags when your credit line terms are no longer competitive. The wedge could be any single expense category, but the long-term play is a general-purpose cost optimization agent for SMBs that quietly saves them money while they sleep.
From paying for tools to paying for validated outcomes
Healthcare has historically paid for activity. Time spent in visits. RVUs generated. Minutes logged under asynchronous codes. Licenses purchased per clinician. Even digital health has largely followed this pattern, charging per seat, per message, or per device. That model is increasingly misaligned with how value is created.
As reimbursement tightens and health systems face sustained cost pressure, activity-based metrics are becoming harder to defend. As technology enhances productivity and extends clinical capacity, engagement metrics become less meaningful. When workflows become more efficient, time-based reimbursement can penalize adoption. When work shifts outside the traditional encounter, billing frameworks struggle to recognize its value. Measuring screen time or user activity tells us little about whether care improved, risk declined, or costs were contained.
We believe the next generation of healthcare companies will anchor around validated outcomes delivered across real workflows. Payment models will increasingly reward measurable clinical and operational impact rather than software usage or time logged. The companies that win will quantify their impact on cost, quality, and workflow performance in ways that are legible to providers and payers. They will build traceability into their systems from day one and align their revenue with measurable value created across stakeholders.
Healthcare is entering a period where software must prove its impact beyond activity metrics. If you are building with this mentality, we would love to chat.