Chapter 2.2 — Applications & Monetization

This is the layer where the entire build must eventually pay for itself, and it holds the biggest open question in the atlas. The bearish headline is real: by one widely-cited study, about 95% of enterprise AI pilots produce no measurable profit. The bullish reality sits underneath it: a handful of application companies are compounding toward billion-dollar revenue, enterprise AI spending tripled in a year to $37B, and the money concentrates precisely where the workflow is deep, coding above all. Monetization is not failing; it is bifurcating, and telling the two halves apart is the whole job here.

Everything below this chapter is cost. Applications are where AI becomes a product someone pays for, either a consumer subscription or an enterprise contract, and the durability of that revenue is what ultimately justifies the hundreds of billions being spent on chips and power. The question that hangs over the whole industry is whether enterprises are getting a return, because if they are not, the capex of Chapter 2.11 has no foundation. The honest answer is that it depends entirely on what the AI is doing.

The 95% headline, and what it hides

The bearish case has a number attached. MIT's Project NANDA, in an August 2025 study, found that roughly 95% of organizations saw no measurable profit-and-loss return on their generative-AI pilots despite $30–40B of spending, with only about 5% of integrated projects extracting real value. The study's own explanation matters: the failures came from tools that did not fit workflows or retain context, not from weak models. That distinction is the key to the whole layer.

Set against that failure rate is an equally real explosion in spending. Enterprise AI spend reached $37B in 2025, up 3.2 times in a single year from $11.5B, the fastest growth in the history of enterprise software.

The two facts reconcile once you look at where the money goes. Of that $37B, applications took $19B, slightly more than half, and within applications the spend is heavily concentrated: coding alone accounts for about $4B of the departmental total, healthcare $1.5B, legal $650M. AI deals now convert at roughly 47% versus 25% for traditional software. So the picture is not "AI does not pay." It is "AI pays handsomely in a few deep workflows and disappoints in shallow, horizontal ones," and the market is already sorting winners from the rest. A further tell: startups now capture about 63% of application-layer revenue, up from 36% a year earlier, earning roughly two dollars for every one the incumbents earn.

Application-Layer Market Structure

The winners cluster by vertical, and the roster is worth recognizing because these are the names that dominate AI business headlines.

Who is actually making money

The application layer has produced a cohort of companies with real, fast-growing revenue, most of them still private, and the pattern is that the winners own a specific, high-value workflow end to end.

In coding, the standouts are Anthropic's Claude Code, at roughly a $2.5B annual run-rate by early 2026, and Cursor (Anysphere), which crossed $1B in annualized revenue and raised at a $29.3B valuation, with GitHub Copilot's 20M-plus users and Cognition's Devin (which absorbed Windsurf) rounding out the field. In customer support, Sierra reached about $200M in revenue at a valuation above $15B, and Decagon tripled to a $4.5B valuation, both on outcome-based pricing. In legal, Harvey grew to about $300M of run-rate at an $11B valuation, with Legora close behind. In healthcare, Abridge's ambient clinical documentation reached $117M of contracted revenue across Kaiser, Mayo, and Johns Hopkins, and OpenEvidence, used by roughly 40% of US physicians, doubled to a $12B valuation. In enterprise search and sales, Glean tripled to over $300M and Clay reached around $100M with net revenue retention above 200%. The common thread is defensibility through workflow depth and proprietary data, not model access.¹²³⁴

The pricing revolution: from seats to outcomes

A quiet but important shift underpins the whole layer: when an agent does the work a person used to do, charging per seat stops making sense. Seat-based pricing fell from 21% to 15% of vendors in a year, while hybrid and usage-based models rose to 41%. The vanguard is outcome-based pricing, where the vendor is paid only when the AI delivers a result: Intercom's Fin support agent charges $0.99 per resolved ticket, and Decagon prices per resolution. This transition reprices the entire software industry. It is a tailwind for the AI-native winners, who can capture a share of the labor budget rather than the software budget, and a headwind for any incumbent whose value was the number of seats it sold, which is why the enterprise-software incumbents are racing to bundle agents into every tier.

Agents in production, and the reliability gate

The next phase of this layer is agents, software that takes actions rather than just answering, and the early revenue is genuine. Salesforce's Agentforce reached a $1.2B run-rate, up more than 200% year on year, and ServiceNow's Now Assist passed $750M in annual contract value and restructured its entire product line around AI tiers. Real deployments show the potential: Klarna's OpenAI-powered assistant handled 2.3M conversations, the equivalent of 700 human agents, and cut resolution time from eleven minutes to under two, an estimated $40M profit improvement, though Klarna later rebalanced toward a human-plus-AI hybrid after over-automating.⁵

The constraint, again, is reliability rather than capability, and it is the same gate described in Chapter 2.1 from the model side. Gartner projects that more than 40% of agentic-AI projects will be canceled by the end of 2027 on cost, unclear value, and inadequate controls, even as it forecasts that a third of enterprise applications will embed agents and 15% of day-to-day work decisions will be autonomous by 2028. Only about 16% of enterprises run "true" plan-and-adapt agents today; the rest are fixed workflows. The signal to watch is the conversion rate from pilot to production, which remains in the single digits for genuine autonomous agents.

America sells subscriptions, China gives it away

The national contrast in this layer is about business model as much as capability. The United States monetizes applications directly, through enterprise API contracts and consumer subscriptions; ChatGPT alone has around 50M paying subscribers, and the $19B of enterprise application spend flows to a dense field of $100M-to-$1B revenue startups. China largely does not charge. Its leading assistant, ByteDance's Doubao, reached 382M monthly users, and DeepSeek's app around 130M, but the dominant model is free access funded by advertising and the broader super-app ecosystem. Baidu made its Ernie assistant fully free in 2025 after subscriptions failed, and ByteDance only began piloting paid Doubao tiers in mid-2026, priced from around $10 a month.

The consequence is that China's application layer generates enormous usage and little direct revenue, while its models spread through free distribution, the open-weight strategy of Chapter 2.1 applied to consumer products. For an investor this means the Chinese app opportunity is monetized indirectly, through the platform companies' ecosystems and advertising, not through the standalone application revenue that defines the US winners.

The thin-wrapper question, and where value accrues

A recurring skepticism is that most AI applications are "thin wrappers" around someone else's model, with no defensible moat. The concern has teeth: application gross margins are estimated at 50–60% against 70–90% for traditional software, compressed every time the underlying model gets cheaper, and the time for a capability to commoditize has fallen from about eighteen months to twelve. The counter-evidence is the $19B enterprises actually spent, the startups' 63% share of app-layer revenue, and the ten-plus applications already above $1B in ARR. The resolution is that value accrues not to model access, which is universal, but to distribution, proprietary data, and workflow lock-in. As one framing put it, "every company is now an AI wrapper, so go-to-market is the new moat." The durable winners own a specific workflow so deeply that switching is painful; the vulnerable ones sell a feature a model provider can absorb.

The adoption tests that matter through 2028

The load-bearing question for the whole atlas is whether enterprise ROI broadens from the deep workflows to the shallow ones. Watch the pilot-to-production conversion rate, and whether a follow-up to the MIT study shows the 5% success figure rising. Watch whether the coding-and-vertical winners sustain their growth and margins as the model layer beneath them commoditizes, which would confirm that workflow depth is a real moat. And watch the agent-reliability threshold, because agents crossing into dependable production is what would turn the demand-durability question from open to settled, and pull forward the revenue that underwrites everything downstream.

The trade: own the workflow, not the wrapper

The uncomfortable truth is that most of the clearest winners in this layer are private, reachable only through their venture backers or eventual IPOs. The public expressions are the incumbents embedding AI into existing distribution: Salesforce (CRM) and ServiceNow (NOW) in enterprise agents, Microsoft (MSFT) through Copilot, and the platform companies of other chapters, with Palantir (PLTR) as the high-multiple, high-volatility applied-AI pure-play. Anthropic's coding-driven profitability, reachable through Amazon and Google, is the single best public-market evidence that the application layer monetizes durably where the workflow is deep. The discipline is to favor owners of a specific high-value workflow with proprietary data over thin wrappers whose only asset is model access, because the latter's margins compress every time the underlying model gets cheaper.

What would settle the ROI question

The bullish read on this layer, and by extension on the whole capex thesis, breaks if enterprise ROI stays stuck, if the MIT 95% figure fails to improve and the winners stay confined to coding and a couple of verticals rather than broadening. It also breaks if the agent-reliability gate does not open, leaving Gartner's 40% cancellation forecast to play out and stranding the spend. In the other direction, the thesis strengthens materially if agents cross into broad production and outcome-based pricing lets AI capture a share of labor budgets rather than software budgets, which would expand the addressable market by an order of magnitude and validate the build faster than the market expects.

Sources

Linked evidence for this chapter's figures and load-bearing claims: ^{1 2 3 4 5}