Enterprise AI Adoption

Tan reports YC's media, events, and finance teams (non-programmers) are building skill files and scheduled agents. One finance person compiled 100 Excel workbooks into an agent-built app. This signals horizontal adoption of agent management across all knowledge work, not just engineering, expanding the TAM for agent orchestration platforms.

AI deployment across product (natural language search), growth (agentic traffic partnerships), and internal productivity (coding agents) showing measurable ROI; domestic travel demand strong despite price increases.

Paper enables designers without coding skills to build and ship complete websites (marketing pages, animations, interactions) by directing agents visually; Paper's brand designer built entire marketing site in one week using Paper + agents, a task that would have taken months traditionally.

AI coding assistants compress engineering squads from 7 to 4 people; automation of fraud investigations and dispute resolution; proprietary cross-product data enables SoFi Coach financial chatbot; real-world ROI visible in fintech operations.

Enterprises like PE firms can shift from intuitive top-down pruning of opportunities to exhaustive agent-driven simulations across vast search spaces, creating new work and value rather than just automating existing workflows.

Takeoff's thesis: agents must own the entire revenue-critical workflow (loan funding, patient onboarding) across all systems (CRM, dialers, notaries, underwriters) with an always-on 'heartbeat' harness; horizontal agents that only handle a slice fail because the customer must still orchestrate the rest.

Matthew Prince's strategy — hiring 1,000+ new grads to embed in existing teams — offers a template for enterprises: use human capital to bridge the AI adoption gap instead of waiting for agents to replace workers.

As foundation models commoditize, value shifts to vertical AI firms that combine deep domain expertise, legacy system integration, and workflow automation to capture the much larger labor/services TAM — not just software budgets.

Firms are moving beyond chaos to long-term planning, understanding AI's specific use cases, and investing in upskilling existing workforces rather than layoffs, with AI automating distracting tasks to unlock strategic human potential.

Meta envisions its 200M business users scaling to billions as AI tools lower entrepreneurship barriers, creating a dynamic ecosystem where business agents interact with personal super-intelligence agents to drive economic activity.

Enterprises avoid locking into single model labs because model leadership rotates rapidly; they need partners who drive organizational change (team structure, planning, specs, design) not just token throughput, creating moat for independent application-layer companies.

Large companies now perceive status quo as higher risk than adopting unproven startup solutions, creating unprecedented sales velocity for early-stage B2B companies — deals that used to take years now close in months.

Enterprise adoption is driven by CFO focus on token economics (cloud, on-prem, desktop via GB10/DGX Station) and IT governance of shadow AI; productivity gains come from sharing agent skills across teams, but require business process review to operationalize.

Enterprises want to own their intelligence stack (cost, sovereignty, self-improvement, data privacy); this creates massive market for AI modernization services (fine-tuning, integration, workflow restructuring) that open source model companies can capture via deployed-engineer model.

Consumer AI is near saturation but enterprise needs massive intelligence for complex workflows (M&A due diligence, life sciences, manufacturing); real-world diffusion limited by permits, clinical trials, human coordination — not model capability — making bureaucracy the ultimate rate limiter.

AI adoption is broadening beyond frontier labs to enterprises in every industry — financial services, healthcare, retail, media — using Amazon Bedrock to automate processes and run agentic workloads that drive efficiency and new customer experiences, indicating a maturing enterprise AI cycle.

Most companies are 'token maxing' — throwing more model calls at problems without understanding where AI actually adds value. The winners will build frameworks that match AI to the right workflows, audit results, and empower employees rather than replace them, turning AI from a cost center into a productivity multiplier.

Collison observes that businesses now perceive the risk of sticking with legacy systems as higher than the risk of buying from unproven AI startups, creating an unprecedented window for early-stage companies to sell into enterprises at meaningful scale immediately.

Levy argues the rate limiter for enterprise AI value is diffusion — humans must integrate model intelligence with proprietary data, navigate permits/regulations, and close real-world feedback loops; this makes enterprise AI a multi-decade transformation, not a quick flip.

AI adoption not concentrated in frontier labs; enterprises of all sizes integrating models via Bedrock for automation, agentic workloads, and new customer experiences across every industry vertical.

Companies are broadly adopting AI but lack understanding of how to use it efficiently, leading to 'token maxing' where more AI is assumed better. The real opportunity lies in building the right frameworks and use-case discernment, similar to how Ajax and real-time frameworks took years to mature after the internet hype.

Taxdown runs 'AI Fridays' where every employee builds AI projects with dedicated engineering support, plus a separate 5-month 'entropy' squad that shipped four regional tax products in three months; this bottom-up experimentation culture converts skeptics and surfaces production-ready tools faster than top-down mandates.

Industrial AR pilots are demonstrating ROI in warehouse logistics: workers use AR glasses to scan barcodes, receive real-time placement instructions, and confirm picks without handheld devices or SAP terminal trips, reducing errors and increasing throughput in high-SKU environments.

Deep enterprise trust and access to all documents/emails allows Legora to build proactive agents that can structure data rooms, run diligence questions, and identify missing content — moving lawyers from real-time collaboration to delegating 20-30 minute autonomous tasks.

CEO requires every employee to use AI multiple times per month; AI handles significant percentage of customer queries with faster response times critical for next-day delivery business; AI enables new greetings card category with much larger SKU count; internal evangelist program drives adoption beyond cost savings to customer proposition improvement.

Pedro identifies three adoption tiers—token maxers, average engineers, and the rest using chatbots—and argues companies must build OpenClaw-style harnesses to give non-technical teams virtual employees with Slack, email, and meeting access.

In agentic systems, not every step needs frontier intelligence; reducing a model's decision domain lets Mistral shrink size and energy use dramatically. Enterprises in air-gapped or regulated environments accept a 6-month capability lag for full control, customization, and on-prem deployment, making tailored open models the preferred enterprise architecture.

Current org charts ship siloed tools (Copilot per department) rather than reimagining end-to-end processes; real AI productivity requires assigning owners to cross-functional processes (e.g., supplier onboarding) with KPIs and AI harnesses, turning science problems into narrow engineering problems.

AI-native companies easily acquire customers due to novel value props or labor replacement, but struggle with retention as users constantly switch to newer shiny tools; solving ICP fit and retention is the core challenge for AI application companies.

Pincus cites a statistic that 90% of enterprises investing in AI have gained no benefit, attributing it to shallow adoption (few users on old models like GPT-3.5) rather than 'token maxing' by empowered individuals—suggesting the bottleneck is organizational, not technological.

AI-native vertical SaaS companies are achieving unprecedented growth curves (200%+ ARR doubling at $200M+ scale) in boring industries like property management and legal, resembling early Standard Oil market dominance in intelligence.

Kimi K3 and Inkling enable internal fine-tuning on proprietary data without vendor lock-in. Companies need crash programs with best advisors to decide build vs buy. Inference providers (Fireworks, Modal) will optimize open models cheaper than APIs. Delay means inability to catch up later.

Enterprise usage of Claude (coding and general) is in early diffusion but growing, suggesting large latent demand as AI integrates into workflows.

AI services must price on value (per unit or outcome-based) not cost-plus or seat-based, and founders should strictly limit pilot customers to avoid being overwhelmed before product scales—humans-in-the-loop must scale non-linearly with revenue.

Ping Wu argues successful CX transformation requires distinguishing three conversation buckets: (1) calls that shouldn't exist (fix root cause), (2) low-emotion transactions (full automation), (3) high-emotion complex interactions (human augmentation). Automation alone fails; a unified platform applying different AI roles per bucket is needed.

CEOs must directly build with AI tools to understand frontier capabilities viscerally, enabling them to disseminate AI-native practices throughout the organization and avoid becoming 'redundant CEOs' of legacy-style corporations.

Zaffino argues current LLM capabilities — data extraction, coding velocity, agent orchestration — are already sufficient to transform insurance underwriting, claims, and portfolio management, with 10x execution speed improvements achievable now.

Companies realize passing proprietary business logic (ontology) to frontier model APIs enables model providers to train competing products; economic incentive (frontier API costs) aligns with privacy incentive to drive specialized model adoption on neutral infrastructure.

Resolve AI's autonomous incident response and Crust AI's voice agents for Marriott demonstrate early but tangible enterprise adoption of agentic workflows, validating the infrastructure rewrite thesis.

Enterprise adoption of AI coding tools follows a product-led growth motion where individual champions (PMs, designers, ops) bring tools in via weekend experimentation, then sales-assisted hackathons and education convert them to enterprise deals, with security/compliance as the key differentiator.

Users cannot articulate needs for capabilities they don't yet understand (Henry Ford 'faster horses' problem); adoption lags because people lack mental models for what AI can do, making problem-focused development more valuable than tool-first approaches.

Individual contributors can now spend thousands on API calls via agents, forcing companies to treat token allocation like hedge fund capital allocation — evaluating ROI per IC strategy, using smaller models where possible, and building LLM-based auditing of generated token value.

AI-native companies (ChatGPT, Claude, Replit, Cursor, etc.) launch globally from day one; 57% of new Stripe companies are non-US; the 2025 cohort grows 50% faster than 2024, with 2x companies reaching $10M ARR within three months, GitHub pushes up 41%, and 20% of Atlas startups monetize within 30 days vs 8% in 2020.

Enterprises adopt coding agents organically through individual developers, then approach vendors for security compliance, signaling true product-market fit without traditional top-down sales.

Engineering teams (SRE, SWE) already use agent managers internally; the next wave requires solving identity/access controls, data permissions, and blast-radius management for AI-generated code — Google expects 2027 to be the crossover year where non-engineering functions (finance, forecasting) adopt fully agentic workflows, with startups having an AI-native organizational advantage.

Enterprises gain trust not through exhaustive upfront testing but by shipping fast with small eval sets (5-10 cases), observing real runs via tools like LangSmith, and iterating weekly—because agent architectures become obsolete every 9 months.

90% enterprise AI project failure stems from 90% of enterprises not knowing IT; this organizational drag acts as a brake on fast takeoff, giving society time to adapt to log-linear scaling improvements.

The most AI-forward companies are buying Granola not just for notes but to build a proprietary context engine from meeting transcripts that makes their internal AI agents more effective.

CEOs across all industries recognize they must become AI companies or die, driving massive demand for workforce retraining; however, organizations are stuck at basic Copilot/ChatGPT usage, face employee resistance to automation, and have no systematic infrastructure for identifying, upskilling, and redeploying AI-fluent talent.

To make the company brain legible to AI, all communications (emails, Slack, DMs, meetings) must be captured, diarized, and synthesized into structured context; unrecorded knowledge is invisible to the intelligence layer.

In a rapidly shifting AI landscape, the only durable strategy is picking a narrow ideal customer profile (in-house legal teams) and making them extremely happy; product-led referral growth compounds while competitors confuse their ICP.

Single-condition digital tools (e.g., maternity apps) hit revenue ceilings and deliver 3-5x returns over 9+ years — insufficient for VC — while integrated platforms combining telehealth, pharmacy, and marketing (Hims & Hers) build defensible flywheels worth billions.

As enterprises get better at deploying AI, they shift toward open source models for cost efficiency; broader main street adoption will drive next leg of demand growth beyond hyperscalers.

The maturation of AI agent frameworks (harnesses, Cloud Code) since summer 2024 created a step-change in reliability, allowing Factorial to go all-in on embedding agents that can proactively manage complex workflows rather than just assist users.

The remediation bottleneck is manual QA processes; autonomous QA agents that codify functional correctness can validate security fixes instantly, allowing continuous pressure testing of new code before production deployment and closing the loop between finding and fixing vulnerabilities.

Anthropic observed latent demand from non-engineers — designers, finance, data science — who were jumping through hoops to install a terminal tool. This led to Co-work, a GUI wrapper with VM isolation and guardrails, built in 10 days by Claude Code itself, signaling broad enterprise adoption beyond developers.

PhotoRoom mandates English internally even in its Paris office, using language as a filter to attract only those committed to building a global company. This cultural mechanism aligns team mindset with the scale required for category leadership in AI applications, where winner-take-most dynamics reward early global orientation.

Adam Swiecicki argues the market's split between 'AI companies' and 'software companies' is false; all future winners will embed AI, and Rippling is already benefiting from OpenAI's advances while Ramp leads in AI-forward CFO positioning.

Real-time payment data reveals far faster AI tool adoption (ChatGPT, Anthropic, Cursor, Cognition) across mainstream businesses than official surveys capture, suggesting productivity gains and revenue acceleration are already underway but underreported.

ElevenLabs sees enterprise adoption driven by developers pulling API into organizations, then expanding to agent platforms, with executive referrals creating a self-reinforcing growth loop that accelerates deal velocity.

Agents will evolve from automated ICs to organizational leaders that receive business goals, diagnose root-cause problems (e.g., onboarding localization gaps), and propose strategic fixes; board members may eventually oversee agent CEOs.

Jain argues that even if base model improvements stopped today, enterprises have barely scratched the surface of current capabilities. He predicts 5 years of massive growth as utilization moves from 1% to 10-20% across verticals, making application-layer innovation the primary value driver regardless of model progress.

Horizontal tools like Rig, Zapier, or Airtable lack a single ICP; buyers are AI-obsessed non-technical leaders (CEOs, CROs, RevOps) who want to rip out legacy workflows, build a data foundation, and automate heterogeneous processes — adoption spreads virally across functions once seeded.

Companies stuck with legacy systems and offshoring manual work represent a large addressable market for AI agents that can integrate with existing tools, as they cannot afford lengthy digital transformations.

Companies should restructure around three archetypes — builder ICs, outcome-owning DRIs, and AI-founder leaders — eliminating human middleware because queryable, artifact-rich organizations let intelligence route information directly, making velocity a function of token usage not headcount.

General-purpose AI provides baseline translation, but enterprises need specialized quality with full-stack control for compliance-critical use cases like legal contracts and FDA-regulated healthcare, creating defensible moats for vertical AI applications.

Software reliant on ease-of-use and interaction (e.g., legacy lease management) is vulnerable to AI agent replacement, while tech spend as a share of enterprise budgets will double from 5-6% today to ~10% by 2035, driving demand for IBM's software portfolio and consulting.

After 18 months of pilots, enterprises realize point solutions create unmanageable technical debt; the winning approach starts with API quality lift, centralized discovery/security, then deterministic workflows, finally agent-created workflows at scale.

Purpose-built AI agents that reason over unstructured data and compliance documents can fully automate complex KYC, KYB, and content review, eliminating rules engines, classifiers, and human reviewers while adapting rapidly to new fraud patterns.

AI gives every role code access, blurring traditional swim lanes; however, Karri argues the winning teams will still separate exploration (designers in Figma/Canvas) from production (engineers reviewing agent output) and will pause to reflect on feedback rather than ship continuously.

Enterprises in operationally heavy industries are deploying voice AI agents at scale to handle core functions like quoting, tracking, and collections, with one customer replacing a planned 500-person hiring spree with an agent making 20-40k daily calls. Customer acceptance is high when AI is disclosed upfront.

The biggest barrier to enterprise AI adoption is the need for forward deployed engineers to configure policies; building an AI forward deployed engineer that joins Slack/Meet and automates policy changes will unlock the next wave of enterprise AI deployment.

Standard RAG adds unacceptable latency for voice agents. Streaming RAG processes audio chunks in real time, triggering retrieval when early chunk document sets match final query results, or using a fine-tuned model to decide when partial queries contain sufficient information. Meta's paper shows 0.5s latency reduction on synthetic and 1.5s on human speech with unchanged accuracy — critical because audio hallucinations are harder for users to detect than text.

Banks are still grappling with basic digitization after 20 years; the opportunity for AI-native fintechs is massive because incumbents' legacy infrastructure and compliance burden prevent rapid AI adoption in core processes.

Organizations making most effective use of AI share two traits: CEO-led mandate to automate everything with hands-on leadership, and uncomfortably permissive data access allowing AI to read all meetings, code, Slack, and email.

US business formations re-accelerating driven by non-employer firms; Stripe Atlas 2026 cohort tracking 5x revenue of 2025 cohort; top 100 AI startups median selling into 55 countries in year one — AI enables lean, globally distributed companies from inception.

CEOs using LLMs to synthesize Slack, PRs, and docs directly gain unfiltered visibility into engineering velocity and quality, reducing reliance on biased middle-management reporting and enabling flatter, more efficient org structures with more junior 'agent managers'.

Companies treating AI as a transformation initiative with dedicated resources and employee training see real returns, while those running superficial POCs abandon them; the usage gap (e.g., Copilot licenses vs. actual usage) creates opportunity for platforms that drive adoption to value.

Brands need 100% accurate logo/typography translation and product photography; open-weight compact models enable on-prem licensing solving IP concerns for design, manufacturing, and defense workflows.

The limiting factor for AI adoption in SMBs is not model capability but the expertise to prompt, integrate, and operate them; vertical SaaS that packages models into outcome-oriented workflows will capture this demand.

Primary users are small/medium business owners currently running operations on email, WhatsApp, spreadsheets; they would have paid dev shops but now build themselves, creating a new category of 'personal software' and solopreneur ventures that raise funding on AI-built products.

Contrary to the stereotype of lawyers as tech-laggards, Gabe reports senior partners at top firms rapidly adopting Harvey because LLM technology is inherently suited to language-intensive legal work, creating a pull-based adoption dynamic rather than push-based sales.

Recording all meetings, transcripts, and artifacts creates a shared organizational brain. Agents that meta-prompt on this corpus (dream cycles) continuously improve skills — e.g., YC's two-sentence pitch skill now outperforms individual partners. This compounding loop is the micro-mechanism for building superintelligence inside companies.

Adjacent professional services verticals (accounting, insurance, claims processing, compliance) offer massive opportunity for domain-specific AI combining deep expertise with automation, where mundane enterprise work creates high-value wedge products.

The non-destructive testing sector faces a demographic cliff with 40% of inspectors retiring within five years, eliminating the option to maintain status quo with Excel and legacy tools and creating urgent demand for AI that captures expert knowledge and multiplies remaining workforce productivity.

Satya Nadella's vision of enterprises owning their intelligence drives demand for infrastructure that ingests decades of proprietary data (emails, docs, code) into searchable memory, enabling agents to act as knowledgeable insiders rather than generic assistants.

A fashion brand with 10-20M revenue and zero technical staff built a centralized 'company brain' by dumping all legal, financial, and operational docs into a repository, connecting data silos (Shopify, Klaviyo, Holded, banks), and using Claude Desktop as the universal execution layer — proving non-tech SMEs can deploy high-impact AI by focusing on context centralization over model sophistication.

Current phase requires harder work to stay ahead — follow AI developments, master frontier models, and seek messier roles inside organizations; the productivity dividend arrives only after this investment period.

The toolkit provides open Nemotron models (3 Ultra: 5x faster, 30% cheaper, hybrid SSM-MoE), Open Shell secure runtime (adopted by Red Hat, Canonical, Microsoft), and CUDA-X libraries as agent skills — allowing enterprises like Cadence, CrowdStrike, Palantir, SAP, ServiceNow to build super agents rather than be disrupted.

Target's integration with OpenAI and focus on app, loyalty, and delivery improvements could drive customer loyalty and sales, but execution risk remains high.

DBS drives AI adoption by requiring every employee to build their own generative AI use cases within their customer journeys, creating ownership and practical learning that scales experimentation across 40,000 staff.

Companies try to insert AI into human-shaped workflows rather than redesigning processes around AI capabilities (as Amazon did with warehouse robots). This 'human replacement' mindset prevents the efficiency gains seen when systems are architected for AI from the ground up.

Frontier intelligence is now a perishable asset (shelf life: weeks). Value shifts to 'interfaces' — orchestration layers that hot-swap models (Kimi, Inkling, Fable, proprietary) per task. Enterprises with technical maturity (JP Morgan, etc.) will build internal AI stacks using open weights as commodity inputs.

Major enterprises are now deploying agents at scale for software development, DevOps, and QA; workflow reimagining is just beginning, signaling an inflection from potential to actual enterprise AI revenue.

Enterprises gain trust through observability (LangSmith) and eval-driven development; starting with just 5-10 scenarios forces product thinking about desired agent behavior, and living eval datasets capture real-world usage to guardrail future prompt changes—shipping iteratively with limited blast radius beats months-long waterfall builds.

Investors are fleeing software names like Salesforce and Intuit because AI could negate the need for traditional SaaS platforms, fueling a flight to energy, utilities, and consumer defense that has pushed those defensive sectors to expensive valuations.

Dassault Systèmes is restructuring its 3DEXPERIENCE platform—used by 45M users across 400K customers—to deliver AI agents (Virtual Companions) as a core service, embedding reasoning, physics simulation, and regulatory compliance into engineering workflows rather than bolting AI on top.

Enterprises need models specialized to their domain, language, and private codebases — smaller tailored models run faster, cheaper, and satisfy data sovereignty; Mistral's Forge platform and on-prem deployment model directly address this, turning customization into a compounding infrastructure investment for each customer.

Zeta's Athena lets customers simply voice desired outcomes (e.g., 'add a couple million customers') while AI executes across the platform. This shifts the product from a tool users must master to an autonomous agent that delivers value directly, dramatically increasing platform utilization and stickiness across the enterprise.

Enterprises that succeed with AI adoption run organization-wide hackathons, tolerate failure, and empower ICs/tech leads (not boards) to drive change; EY exemplifies this by embracing experimentation despite not being an AI-native firm.