AI Applications

UK accounting rule change moved driver payments from cost of revenue to contra revenue, creating an 8 percentage point revenue growth headwind with zero impact on gross profit, operating profit, or free cash flow. Investors focusing on GAAP revenue miss the underlying 22% bookings growth and consistent margin improvement.

The hosts discuss Notebook LM's ability to generate educational podcasts from source material and envision the next step: live voice mode with memory acting as an interactive tutor ('Tutor LM') where users guide the conversation depth in real time. They contrast this with static podcast generation, noting the 'choose your own adventure' dynamic better matches learning needs, and observe the divergence between stated anti-AI sentiment and massive revealed usage.

Founders who grew up in vertical industries (HVAC, pharmacy, music) have irreducible edge — they see workflows AI cannot yet replicate; these 'vertical nepo babies' (Rebar, PillPack, Suno) build sticky AI applications that generic horizontal AI cannot displace.

Services industries representing ~70% of European GDP are ripe for AI-native transformation via roll-up strategies that acquire local operators and migrate them onto shared AI infrastructure, automating compliance and operations while enabling customized service at scale.

Accounting shifts (e.g., UK driver payments moving to contra revenue) can depress reported revenue growth by 8 percentage points while bookings, margins, and free cash flow accelerate, creating valuation disconnects for investors who focus on GAAP revenue.

AI-enabled roll-ups across property management, MSPs, software, and business services are emerging as a primary use case for venture debt, with debt accounting for at least 50% of capital structure as acquisitions become the go-to-market strategy.

AI tools are dramatically accelerating the pace at which founders can build and commercialize complex technologies, raising competitive standards and compressing time-to-market.

Boris introduces 'product overhang' concept — each model generation has capabilities no product yet elicits (whole-file writing, language rewriting, drawing with OpenCV) — creating massive startup opportunity to 'unhobble' models by removing scaffolding and giving them harder tasks with verification.

Legal AI platforms like Lorra build defensibility through complex multi-party workflows and organizational context that generic frontier models cannot easily replicate, creating durable application-layer value.

AI coding agents and regulatory analysis agents turn months-long compliance work into rapid processes, making 'AI-native' infrastructure a multiplier for scientific teams rather than a replacement.

Simile's models predict subjective human preferences at scale using representative population data. Enterprises filter populations and test products/messages via interactive agents traversing Figma/web demos. This democratizes access to human insight — previously only affordable for large companies — and brings human voices into high-stakes decisions. As agent production costs plummet, the alpha shifts to knowing what humans actually want.

Dean advises founders to target problems where general models succeed 0-1% of the time, not 20%, citing AlphaFold for protein folding and suggesting material science, chip design, and personal data organization as ripe domains for specialized models with proprietary data or evaluators.

Models improve exponentially but need application layer to connect to real workflows; this 'bridge layer' (agent-first companies, infrastructure, post-training, data) is where massive value accrues, analogous to Palantir's early role bridging tech and operations.

AI-native interfaces (NL search, design agents, answer engines) driving engagement and new revenue models; conversion lags initially but intent data enables better matching; agent partnerships (Claude connector) creating incremental booking channels.

AI-generated designs converge on identifiable 'vibe coded' patterns (excessive bold, purple gradients, card overload), making exceptional human-directed design a critical differentiator for startups to build trust and stand out; every great company of the last 20 years had exceptional design.

Voice AI leader Eleven Labs is vertically integrating from model provider to enterprise application layer (customer support agents), capturing more value as AI shifts from infrastructure to deployed workflows.

Leopold's short on software (Microsoft, cybersecurity) bet on AI models absorbing applications, but record software earnings suggest the absorption timeline is longer or the thesis is wrong.

Park predicts synthetic panels will surpass the human panel market because they unlock the 95% of hypotheses currently untested due to cost, time, and scale constraints, enabling society to simulate every decision before deployment.

Taxdown's moat comes from 200+ tax advisors continuously labeling conversations, creating a proprietary dataset that generalist LLMs cannot replicate; guardrails ensure zero false negatives, enabling automation of 80%+ of tickets while maintaining the 'declaration well done' brand promise that a 1% error rate would destroy.

Legora is moving beyond task-level augmentation to proactive agents that can execute multi-step legal workflows (e.g., M&A due diligence) autonomously, accessing full document repositories and email contexts to deliver completed work products.

Successful horizontal AI applications must serve both novices and experts simultaneously, similar to how iPhone serves both casual users and photography nerds, enabling a journey from first-time builder to sophisticated creator.

Instead of automating existing KYC, Brex redesigned onboarding end-to-end, enabling KYC at lead qualification which changed targeting and credit decisions—illustrating how AI-native processes unlock new business logic.

Internet shifts power from supply/distribution to demand ownership; companies that become the default consumer interface (Hims for prescriptions, Uber for rides) gain pricing power over suppliers and expand into adjacent high-margin services.

AI is causing existing acquisition channels (SEO, written outbound, cold calling) to collapse, with CAC payback increasing 5x to 57 months for public SaaS companies, while product building barriers plummet; this makes go-to-market the primary bottleneck and creates demand for referral-based growth infrastructure.

Checkout.com's long-standing AI use in intelligent acceptance (optimizing issuer-specific transaction routing) and new AI customer support bot (80% CSAT, exceeding human agents) demonstrate practical AI deployment that directly improves merchant revenue and operational efficiency.

Airwallex is layering AI agents onto its infrastructure to handle financial operations, instruct payments execution, and auto-reconcile cash — creating a 'giant finance' layer where software autonomously drives financial outcomes.

The moat is not whether an app can be copied — AI makes building easy — but the velocity at which a company's proprietary ontology improves. If the flywheel spins faster than competitors', the leader captures the market because customers default to the platform that iterates fastest and solves more problems, as seen with Spotify vs. Apple/Amazon.

Duolingo and Zeta Global demonstrate how AI dramatically accelerates content creation and campaign building in education and marketing software, driving 35-50% revenue growth at 16-20x FCF multiples, representing a value AI segment overlooked by the market's infrastructure obsession.

The next paradigm in enterprise software is 'organic software' — systems that store knowledge at individual, team, and company levels (memory), automatically infer preferences from behavior (adaptability), and self-monitor with guardrails (self-referencability). This replaces static SaaS playbooks with dynamic, AI-driven experiences that learn and reshape themselves continuously.

OpenAI and generalist LLMs are expanding into vertical applications; specialists survive only with unique data, distribution, or metaknowledge in a defensible niche. Otherwise the 'big player takes all' at scale.

Positioning in ChatGPT/Perplexity answers now mirrors early SEO: high-quality content, structured data, and authority signals determine inclusion; companies (Skinvity, Metricool, Go Trendier) are rebuilding content engines to capture AI-mediated discovery.

Model Context Protocol lets non-technical VCs deploy AI agents for sourcing and evaluation in natural language; shifts work from data gathering to conviction-building on proprietary deal memos.

Acquiring traditional services firms and layering AI on top fails because legacy operations have incompatible metrics, hiring, and performance expectations; only regulatory moats (e.g., insurance licenses) justify buy-over-build.

AI-native services companies in regulated industries (tax, audit, insurance, law, healthcare) can achieve 50%+ gross margins by combining frontier models with human-in-the-loop operations, capturing markets 2-3x larger than pure software TAMs.

AI automation of integration and operations enables thousands of acquisitions per year versus handfuls in traditional PE roll-ups, creating a new model for consolidating labor-intensive, communication-centric service industries like lettings, accounting, and legal.

LLMs eliminate manual rostering (2-4 hours/week saved), auto-digitize menus from photos, optimize recipe costing against inflation, and manage last-minute shift coverage via automated backup pools — high-value workflow automation in a labor-intensive vertical.

Traditional air defense fails against mass cheap drones because of cost asymmetry and human cognitive overload; embedding recurrent neural networks for radar fusion and computer-vision terminal guidance on interceptors raises kill probability and lowers cost-per-kill, creating a defensible AI application layer in defense.

Agentic AI workflows that search literature, generate novel candidates, predict properties, simulate dynamics, and plan synthesis can systematically explore chemical space far beyond human trial-and-error, unlocking solutions for PFAS removal, semiconductor materials, and energy storage.

AI coding platforms enable domain experts (physical therapists, pool maintenance, sports clubs) to build sophisticated vertical SaaS without engineers, disrupting traditional agencies and addressing millions of underserved niches still running on legacy software like MS-DOS.

The rapid release of developer-facing tools like Cloud Code by Anthropic indicates a step-change in AI application usability, making them a bellwether for the broader AI application sector and a magnet for ecosystem events.

Hims & Hers is deploying capital to scale AI capabilities across its closed-loop data ecosystem, aiming to improve consumer experience and drive cost efficiencies that could justify current leverage.

LLMs have crossed the threshold for compliance automation — GitHub now auto-fills 92% of security questionnaires via Vanta; the roadmap includes hundreds of agentic workflows for policy mapping, risk assessments, evidence collection, and auditor communication, collapsing GRC team headcount while elevating strategic work.

The vast majority of valuable data has never been on the internet or seen by models; the real AI winners will be those who can access and train on this offline proprietary data, not those with temporary speed or idea advantages.

Legal tech proves vertical AI success requires owning the ground-truth data layer (Vlex's 25-year case law corpus) and constraining generation via RAG — generic LLMs hallucinate legal citations because they lack verified retrieval; the defensible moat is proprietary data, not model architecture.

Anthropic's enterprise-first strategy (40% share vs OpenAI 27%) and coding-specific product (Claude Code) drove revenue from $87M to $45B ARR in 2 years. Constitutional AI and safety positioning build trust for enterprise adoption.

Granola is transitioning from a personal meeting notes tool to a team workspace that aggregates meeting transcripts to serve as the context layer powering internal AI agents, which will handle half of company work in the future.

The half-life of AI skills is compressing rapidly, making traditional long-form courses obsolete; the highest-value applications are micro-upskilling (continuous 10-15 minute learning), curated tool discovery to cut through 'AI tool overwhelm,' and moving workforces from passive LLM prompting to building autonomous agent workflows — a transition most organizations have barely begun.

Manus targets underserved non-technical users who cannot access engineering talent, enabling them to build functional web applications and businesses without coding skills — evidenced by users in countries with scarce developer ecosystems paying $1,000+ to build sites on Manus.

Generative AI enables unbundling of professional legal services to serve the 12.5M annual UK labor violations where only 115K reach pre-action stage, creating a new mass-market category.

Holiday home management's unique intensity — 6-7 guest turnovers per month per property vs years for residential — creates a coordination problem (pricing, check-ins, cleaning, maintenance, payments) that agentic AI solves better than human teams, making it a prime vertical for vertical AI automation.

The AI rollup thesis — buying fragmented, labor-intensive SMBs (holiday rental managers, home services) and deploying agentic AI for pricing, customer communication, and operations orchestration — has been validated at scale by Arbio, Buena, Dwelle, and Ilya, turning 15% margin businesses into 70% EBITDA software-like platforms.

ML models now forecast millions of units daily with no humans in loop, making supply chain more agile; GenAI-powered search advertising tools drive higher ROAS for brands, growing ad revenue to hundreds of millions ARR; internal automation eliminates nearly all third-party SaaS spend.

Vast untapped personal datasets — photos, geo-location, health records, communication history — become platforms for AI-generated consumer apps when layered with LLMs and multimodal models, enabling hyper-personalized discovery, social, and utility experiences.

Meta's first major AI release under new leadership integrates image generation and editing into consumer apps, signaling a push to monetize AI through creative tools and compete in the consumer AI application layer.

AI tools like Cloud Code allow a single person to handle design, product management, coding, and operations, drastically lowering the barrier to launching consumer AI products.

AI video generation has moved from experimental to standard workflow in media production, driving enterprise partnerships and equity deals with studios like Lionsgate.

The most ambitious AI startups should target regulated sectors (legal, healthcare, financial services) as full-stack providers owning the entire value chain, not as software vendors, because this creates moats through regulatory permission and outcome accountability.

Founders report that AI tooling now allows building minimum viable versions (V0) in 3 days versus 3 months previously, fundamentally increasing the number of experiments a team can run and the speed of learning. This structural shift lowers the cost of failure and favors teams that aggressively reduce cycle time.

PhotoRoom's trajectory demonstrates that depth — sequentially dominating photo, then e-commerce photo — compounds more reliably than breadth. Each 10x focus step expanded the addressable market rather than limiting it, suggesting AI application founders should prioritize vertical depth before horizontal expansion.

Finance leaders at Rippling and Glean are using AI agents to automate monthly close commentary, flux analysis, and Salesforce deal verification, freeing accountants for anomaly investigation and shifting finance from reporting to value-add activities.

Liaw states most IVP AI investments are application-layer and expresses satisfaction; he sees Europe producing a scaling cohort of applied AI companies (e.g., Cradle in drug discovery) that leverage local domain density.

No software vendor builds for 5-person firms (95% of market), leaving legacy systems that can't handle scale. At €80M revenue, Afianza can finally fund proprietary client portal and OCR/automation to productize service delivery — a threshold where AI-driven efficiency becomes economically viable in a historically low-tech sector.

Call centers and web analytics are first verticals where AI agents replace humans but create new supervision burden. Humans cannot monitor 100% of AI interactions (sampling 2-3%). AI supervisors (Infinite Watch) analyze every voice call and session replay for compliance, quality, hallucination, latency, UX friction. Insurance/regulated verticals adopt first despite slow procurement. Pattern generalizes to any domain deploying autonomous agents.

Sora-class models now produce 30s+ coherent video. Market will segment: Freepik/Creative (ad/ecom images), HeyGen/Synthesia (short clips), InVideo (education), Mito (collaborative multiplayer for films/music videos/docs). Differentiation via product experience (multiplayer, timeline, asset management) not raw model. Lightspeed backing signals conviction. Parallels 2006-2010 social media segmentation (FB/LinkedIn/Twitter).

The market needs AI writing tools that assist the thinking and editing process while keeping the human in control of every word, rather than generating text that anchors thinking and restricts creativity.

General-purpose LLMs outperform custom recommenders by leveraging massive context windows and world knowledge, but the optimal consumer interface for AI-driven commerce (chat vs visual vs hybrid) remains undetermined; end-to-end job completion matters more than low-friction entry.

Intercom's Apex model (post-trained open weights) beats Sonnet/GPT on resolution rate, latency (0.6s faster), and cost by focusing narrowly on customer service; argues all vertical AI companies must eventually own model layer to sustain competitive moat.

Vertical AI products initially added scaffolding around weak models, but as models get better that scaffolding value disappears; defensibility shifts to network effects within the vertical, reverting to classic 2010 SaaS playbooks.

Frontier labs cannot be model-agnostic because they're incentivized to sell their own tokens; application-layer platforms that let users switch models dynamically will win, analogous to energy-agnostic factories.

Back-office automation is highly transferable across verticals because processes share structural similarities despite different tools; a horizontal platform with vertical-specific go-to-market can scale across insurance, healthcare, and legal.

Companies like Starbucks using AI to build custom inventory, sales, and CRM applications ('vibe coding') demonstrates structural shift: enterprises replace purchased SaaS with internal AI-generated tools. This validates bear case for legacy software vendors facing growth erosion and pricing pressure from AI-native competition.

DeepL is applying its full-stack AI approach beyond translation to build a horizontal agent that automates administrative workflows in finance, legal ops, and people ops, democratizing access to all employees to drive bottom-up adoption.

Leading artists are integrating generative AI for image/sound creation and workflow automation, but the broader creative community's livelihood anxiety — most artists live hand-to-mouth — creates a bifurcated market where tools that respect provenance and pay royalties will win over pure scraping models.

When enterprise data is trapped behind human-built UIs without APIs, AI agents can operate browsers to scrape and reason over that data, unlocking automation for previously inaccessible systems.

Investors agree AI is not a standalone category but an enabling layer for vertical B2B solutions (finance, sales, support, healthcare); pure model/infra plays absent in Spain, while application-layer defensibility is questioned due to low barriers to entry.

AI agents for customer support can reach 60-70% deflection rates immediately and target 90-95%, dramatically outperforming traditional IVR/chatbot systems that only achieve 10-15% deflection, creating massive efficiency gains for enterprises.

AI is demonstrably increasing engineering output 3-4x; same productivity leap is coming to sales, legal, and support. CIOs must choose: double down on current processes, wait for AI-native tools, or rip and replace — but organizational capacity for simultaneous change is impossible.

AI application companies are achieving hypergrowth revenue trajectories — hundreds of millions ARR in months — justifying large venture rounds despite high valuations.

A stealth portfolio company built a full integrated banking product in two weeks versus two years for Penta and eight weeks for her second startup, demonstrating how AI tooling and open banking APIs are collapsing fintech development cycles.

Apps that merely manage data (fitness, notes, reminders) will be displaced by local AI agents that have full computer access and can store memories as user-owned markdown files, offering a more natural, proactive interface.

Foundation models commoditize reasoning, but enterprise vertical applications (CPQ, enterprise search) retain defensibility through deep workflow integration, multi-stakeholder coordination, regulatory compliance, data ontology, and liability assumption — creating moats that pure model access cannot replicate.

James Hawkins became convinced AI progress will continue indefinitely after studying OpenAI founders' reasoning — human intelligence is physical and replicable — leading PostHog to pivot strategy toward building an AI product manager that autonomously ships features via PRs based on unified customer data.

There are not enough applied AI companies working on high-value business processes (legal, finance, back office, marketing) versus tooling around AI itself; these vertical agent markets will go through competitive waves like coding and customer service before consolidating.

Brands must build 'marketing for machines' — distributing structured, factual content aimed at retrieval agents (bots) rather than narrative-driven human content, because AI models ingest raw information and generate the narrative post-ingestion, fundamentally changing SEO and content strategy.

Specialized data relationships (tax, creative, medical) beat generalist AI. Intuit can do AI tax prep for $100; Hims owns patient relationship; Adobe owns creative workflow. Generalist AI (Perplexity) lacks vertical data moat. Incumbents win by embedding AI into existing workflows.

Gabe argues that successful vertical AI companies like Harvey will differentiate by building deep enterprise infrastructure — client matter management, collaboration platforms, workflow orchestration — around foundation models, because the model layer is commoditizing and the value accrues to the application layer that solves complex regulated business problems.

The winning AI-native architecture is recursive loops where AI ingests multimodal data (support tickets, session recordings, logs, Slack), identifies problems, ships fixes via PRs, measures impact, and iterates — moving beyond copilots to autonomous product improvement.

ChatGPT-style interfaces will replace broker-mediated distribution for standardized insurance (auto, home, pet) because they offer instant quoting, transparent coverage explanation, and frictionless purchase — Goldman Sachs research validated this by predicting broker disintermediation after TUIO's ChatGPT app launch wiped $39B off broker market caps.

ChatGPT, Gemini, and Claude are becoming the first stop for AI image generation, bypassing Adobe Firefly; freemium users may never discover Firefly, eroding Adobe's top-of-funnel acquisition.

Garmin is integrating AI nutrition tracking (photo-based food logging) into its Connect subscription, illustrating how consumer hardware companies can layer AI services to increase lifetime value and justify premium pricing.

DLSS 5's one-step pixel-space diffusion transformer runs frame-by-frame causally within the 16ms game frame budget, making generative enhancement a standard real-time graphics primitive rather than an offline batch process.

GitHub commits tripled in early 2026 vs 2025, turning $3T developer salaries into ~$9T productivity; Cadence chip design agents compress RTL verification from weeks to hours (40x) — coding and chip design are the first killer apps for agentic AI, with NVIDIA providing the underlying acceleration.

Adobe's industry-standard position in professional video editing and graphic design creates a durable moat that consumer AI tools like 'nano banana' cannot disrupt; professionals require precision, workflow integration, and file compatibility that current AI image generators lack.

DBS's journey-based management framework with live data dashboards and AB testing captures incremental revenue from AI/ML nudges across customer journeys, delivering $750M last year and targeting $1B+ this year, proving AI can directly drive top-line growth in banking.

Duolingo uses AI to rapidly add math/science/history courses; Zeta launches Athena conversational marketing agent. Both demonstrate AI as force multiplier for existing subscription/software models.

As AI generates infinite content, trusted human perspective becomes scarce and valuable; transparency about AI use lets readers choose, reinforcing Substack's subscription model.

The company plans to use proceeds to enhance AI-driven consumer experience and leverage proprietary data, representing a telehealth rollup strategy betting on AI-powered personalization to improve margins and retention.

Companies with proprietary data and workflow moats that embed AI into core products (not infrastructure) gain margin tailwinds from falling model costs while expanding TAM; memory and commodity hardware remain poor investments regardless of AI demand.

Intuit's tax/accounting moat persists because AI will be integrated into trusted platforms rather than replace them; users trust established software with sensitive financial data and AI features like taxes, making incumbents the natural AI interface.

Zeta Global's Athena platform shows marketing tech companies can expand from ad placement optimization into broader business intelligence (ROI optimization across sales, marketing, R&D), creating larger TAM and higher-value AI subscription revenue.