The A.I. Vendors Startups Are Really Paying

Startups tell everyone they “use AI,” but who are they actually writing checks to? Now a new report on AI spending from Andreessen Horowitz, based on transaction data pulled together by fintech startup Mercury, gives one of the clearest looks yet at where real dollars are going at this early stage for companies. The topline: spending is focused among models and productivity “copilots” with a long tail of fast-rising apps churning beneath.

Who’s Really Getting Paid by AI-Driven Startups Today

Major labs dominate paid adoption. Top of the list is OpenAI, followed closely by Anthropic—hardly shocking now that founders commonly peg their stacks to GPT and Claude for coding, content and customer workflows. On the application side, developer software is pulling above its weight. Replit among the top, Cursor high on the page too, Lovable—another “vibe coding” tool listed but lower spend. Enterprise-focused coding platforms from Cognition (such as Devin and Windsurf) and newer entrants like Emergent roll in lower down, indicating an appetite for more than one flavor of AI-assisted development.

Significant as well is this divergence between traffic and transactions. Replit was outpaced in raw usage by Lovable in a previous consumer perspective rating from the same firm. Here, the money flow favors those products with enterprise-grade controls and admin tooling and more clear procurement paths. For startups, selecting something for the source code and data you put in is as much of a decision about SSO, audit logs and SOC 2 compliance as it is model quality.

Why Copilots Beat Autonomous Agents for Startup Spend

a16z partners Olivia Moore and Seema Amble call today’s spending focused “human augmentors”—AI that makes people more productive, rather than replacing them. This aligns with what the data says: Note-taking, code assist, sales outreach and drafting tools have real budgets because they fit into established workflows and have a lower operational risk. Fully independent, end-to-end “agentic” systems are still more experimental within startups, though the report anticipates that mix to change as reliability, monitoring and guardrails get better.

How Consumer AI Tools Quietly Sneak Into the Workplace

One of the surprising themes is how fast consumer AI lands in enterprise through bottoms-up adoption. In company spend, you’ll find currencies and country-level tax systems as well as CapCut and Midjourney—popular products in use at some of the largest tech companies today, but that likely have limited crossover to others outside of certain teams for marketing assets, pitch visuals or product mocks. Canva provides the playbook: start with a happy consumer experience and add in enterprise features once usage has penetrated. The report indicates that numerous “personal” AI apps are finding their way to go-to-market, sales and support earlier in order to tap into team budget.

Horizontal Dominance and Rising Heat in Vertical AI

Horizontal, general-purpose tools that work across a team tend to account for roughly 60% of the top paid apps, and vertical, industry-specific tools for approximately 40%. The vertical leaders congregate to three predictable budget centers: sales, recruiting and customer service. Legal tech is also on the rise. Software services like Crosby Legal compress hours of contract review into minutes, transforming a service agreement into software with human-in-the-loop verification.

Why No Single Vendor Is Winning Notes and Meetings

Meeting and transcription software is still fragmented, with Otter.ai, Retell AI and HappyScribe all grabbing spend. That diffusion mirrors the way teams trial and standardize: in many cases, the winner is the tool that plays best with current calendars, CRMs and security policies—not necessarily even the one with the flashiest demo. Anticipate a lot of turnover as new entrants proffer more in-depth CRM enrichment, action item automation and multi-channel follow-up.

What the Spending Data Does and Doesn’t Actually Show

As the analysis is based on Mercury transaction data, it reflects paid usage as opposed to buzz or traffic. That’s a strength—money talks around priorities—but it also means free tiers, open-source models and pilots that haven’t yet converted might be undercounted. Some model usage is occurring from cloud marketplaces (for example, through managed offerings like Azure OpenAI or foundation models on a cloud provider) which can obfuscate spend attribution. And because Mercury’s customer base leans toward startups, later-stage enterprise trends could be different.

A Practical Playbook for AI Builders and Startup Buyers

For founders, the message is unambiguous: enterprise readiness closes deals. Get a head start with role-based access controls, data geolocation preferences, auditability and predictable pricing. Developer-facing tools should be optimized for:

• Latency
• Eval against domain data
• Tight IDE/CLI workflow

For buyers, keep an eye on total cost of ownership beyond API rates—context management, retrieval infrastructure and human QA still generate real expense. Vendor durability, model portability and such matters now make a difference as pricing and performance trends shift quarter-on-quarter.

The New Normal in AI Procurement Is Relentless Volatility

The most sobering finding in the report is how quickly things can change. New apps peak on social adoption waves, while by AI standards “legacy” could mean a product launched last year. As labs discount and launch more powerful models, as incumbents ship native AI features and governance demands harden, procurement stacks will reshuffle. Today’s copilot is tomorrow’s secret agent, and today’s category may coalesce into a few platforms—or disintegrate further, as niche workflows turn out to be profitable.

The throughline: Startups are willing to pay for AI that eliminates toil, plugs into the tools they already love and incites confidence with security and support. Everything else is a shooting gallery.