The Two-Stream Retention Problem: Why Your SaaS Metrics Can't See Half Your User Base

The Accounting Change Nobody Wrote About

In January 2026, JPMorgan Chase disclosed a decision that received less attention than it deserved. The bank reclassified its artificial intelligence spending — approximately $2 billion of its $19.8 billion annual technology budget — from discretionary innovation to core infrastructure.

On the face of it, this sounds like a finance department reorganization. It is much more significant than that.

When JPMorgan's CFO moved AI alongside data centers, payment systems, and risk controls, it changed what auditors examine, what regulators scrutinize, what capital ratios reflect, and how difficult it becomes to reverse the decision. Infrastructure budget is not flexible in the way that innovation budget is. You cannot cut a data center budget by 40% in Q3 without consequences; the same constraint now applies to JPMorgan's AI spending. The bank's LLM Suite — a proprietary generative AI platform now accessible to more than 60,000 employees — is core infrastructure in the same way that Bloomberg terminals are core infrastructure: not optional, not experimental, not subject to next quarter's budget review.

Three months later, in April 2026, Novo Nordisk announced a strategic partnership with OpenAI to integrate advanced AI across its entire business — from drug discovery and clinical trials to manufacturing, supply chain, and commercial operations — with full company-wide deployment planned by end of 2026. This was not a pilot. It was an infrastructure commitment, designed by a company with $36 billion in annual revenue, running across every function in a regulated global industry.

Two announcements. Two of the most conservative enterprise buyers on Earth. Both treating AI not as an experiment but as infrastructure.

For the B2B AI vendors still running the pilot-first playbook, this is the signal they missed.

What Crossing the Chasm Actually Means

Geoffrey Moore's technology adoption lifecycle has a structure that B2B software markets tend to follow with uncomfortable regularity. Early adopters — visionaries who buy based on potential and tolerate significant ambiguity — precede the early majority, whose behavior is fundamentally different. Early majority buyers are pragmatists. They do not buy until the category has proven itself among peers they trust. They do not run experiments; they deploy infrastructure.

The chasm between these two groups is where most enterprise technology companies stall. The challenge is that the marketing that works for early adopters — capability-forward, use-case expansive, innovation-language heavy — actively repels pragmatic buyers. Pragmatists are not evaluating your feature roadmap. They are evaluating whether your product can reliably perform a defined function at scale, with a compliant data architecture, for five years.

By the standard chasm definition, enterprise AI crossed in 2026. The evidence is not subtle. Gartner projects that 80% of enterprises will have deployed GenAI-enabled applications by end of 2026, up from less than 5% just a few years ago. Agentic AI spending is projected to reach $201.9 billion in 2026, 141% growth year-over-year. Companies deploying AI at JPMorgan's scale and Novo Nordisk's risk profile are, definitionally, pragmatist buyers who have seen enough proof from peers to commit.

For AI vendors, this means the market has bifurcated. Some buyers remain in the early adopter phase — startups, digital-native companies, innovation-forward practitioners — and they continue to respond to traditional AI GTM. A growing set of buyers, concentrated in financial services, healthcare, pharma, manufacturing, and regulated industries, have moved to infrastructure status. The GTM playbook for these two groups is not merely different; it is structurally incompatible.

How the Enterprise Buyer Changed

The buyer of enterprise AI in 2023 was typically the head of an innovation lab, a chief digital officer, or a particularly enthusiastic VP of engineering. They had a budget in the range of $50,000 to $200,000. They could approve a vendor engagement with minimal procurement involvement. They evaluated based on capability: what can this technology do? How does it perform on our test cases? Their internal credibility came from being ahead of the trend, not from operational track record.

The buyer of enterprise AI in 2026 is the CIO, the CFO, or a procurement committee with both of them on it. They have a budget in the range of $1 million to $20 million. They cannot approve a new infrastructure vendor without a formal RFP process that typically takes 6 to 18 months. They evaluate based on risk: what can go wrong? How do we recover? Who is liable? Their internal credibility comes from operational reliability, not from innovation optics.

The implications run through every stage of the sales process.

Early-adopter buyers gave vendors access to a champion who could push a deal through on enthusiasm and technical merit. Post-chasm buyers have champions too, but those champions are now subject to overrule by committees that include legal, IT security, compliance, finance, and business continuity. A champion who loves your product can lose an RFP to a less capable competitor that scored better on the vendor stability assessment.

Early-adopter buyers ran pilots. They were comfortable discovering unknown capabilities and tolerating rough edges during a defined evaluation period. Post-chasm buyers run procurement processes. They want the pilot equivalent — a proof of technical fit — but they want it embedded in a structured evaluation that produces a formal vendor recommendation, not an informal champion endorsement.

Early-adopter buyers tolerated pricing ambiguity. They understood that AI pricing models were evolving and accepted a "we'll figure it out" conversation. Post-chasm buyers have been burned by the pilot-to-production cost gap and now require production-realistic pricing transparency before a contract is signed.

The gap between enterprise AI pilot costs and production costs has become a specific source of enterprise distrust. A simple agentic workflow that costs $0.04 per interaction in a constrained pilot environment can cost $1.20 per interaction in a production deployment with full orchestration, tool access, reasoning loops, and output validation. Enterprises who discovered this gap at invoice time are now building cost scrutiny into RFP requirements. For AI vendors, pilot-to-production pricing transparency is no longer optional.

JPMorgan's $2 Billion Infrastructure Signal

The significance of the JPMorgan reclassification goes beyond the number. The specific act of moving AI from innovation budget to infrastructure budget inside a systemically important financial institution has legal, regulatory, and operational consequences that make the commitment effectively irreversible in the short term.

Infrastructure spending at a bank of JPMorgan's scale is subject to a different category of regulatory oversight than discretionary technology spending. The OCC and Federal Reserve examine infrastructure investments as part of operational risk assessments. The board has fiduciary responsibility for infrastructure reliability in a way it does not for innovation experiments. Capital ratio calculations may reflect the infrastructure classification differently than innovation spending. Auditors now scrutinize AI as a core operational dependency rather than a side project.

This means that JPMorgan's decision was not just strategic — it was structurally entrenching. The AI infrastructure that bank has built is now defended by the same organizational gravity that defends its core payment infrastructure. A future CFO cannot zero out the AI line item the way they could zero out an innovation lab.

The LLM Suite deployment to 60,000 employees confirms the operational scale. This is not a pilot. It is a product that tens of thousands of professionals use daily to summarize regulatory documents, draft client communications, generate analytical frameworks, and manage routine information workflows. It runs on private infrastructure with enterprise security controls, model governance policies, and data residency restrictions that would not exist in a pilot deployment. The vendor relationship that enabled this is an infrastructure relationship, not an innovation partnership.

For AI vendors, the question is not whether to try to sell into JPMorgan. The question is whether their product, their compliance posture, their pricing model, and their enterprise sales motion are designed for buyers of this type — or whether they are still designed for the innovation labs that no longer make the purchasing decision.

Novo Nordisk: The Cross-Industry Confirmation

The JPMorgan reclassification might be dismissed as financial services-specific. Banks have always treated technology as infrastructure. Their regulatory environment demands it. But Novo Nordisk's full-company OpenAI integration announcement removes that out.

Novo Nordisk is a pharmaceutical company. Its core business is drug discovery, clinical development, and manufacturing — domains where the regulatory environment is intense (FDA, EMA, ICH guidelines), the timelines are long (10+ years from discovery to approval), and the consequences of system failure are measured in patient outcomes, not quarterly earnings. It is not an organization that deploys infrastructure impulsively.

The scope of the partnership is also notable. This is not "we are using AI in our data science team." The partnership covers drug discovery, clinical development, manufacturing, supply chain, and corporate functions — the full value chain of a $36 billion company operating in regulated markets on multiple continents. The timeline is end of 2026. The scale is company-wide.

What makes this important as a cross-industry signal is the buyer profile. The AI decision-maker at Novo Nordisk is not an innovation lead. It is the C-suite, deploying at infrastructure scale, with full regulatory accountability for every output the AI system produces. If Novo Nordisk's AI systems make an error in a regulatory submission, the consequences are measurable in years of lost approval time and billions in delayed revenue. The vendor they chose had to meet a standard that most AI vendors have not built toward.

The combination of JPMorgan and Novo Nordisk represents two of the most conservative, most regulated, and most complex enterprise buyers in the global economy. Both have committed to AI at infrastructure scale. That is the chasm crossed.

The GTM Playbook That No Longer Works

The standard enterprise AI sales motion of 2024 was: find a champion inside the target organization, run a 90-day pilot with promotional pricing, demonstrate a compelling result, use that result to pitch executive expansion. This motion worked when the buyers were innovation teams with discretionary budgets and short approval chains.

It has four structural failures in the post-chasm market.

The champion problem. Post-chasm enterprise buyers have champions too, but those champions no longer control the budget or the vendor approval process. A data science lead who loved the pilot will lose at the procurement committee if the vendor does not have SOC 2 Type II, does not have an approved data processing agreement, does not have peer references from companies of comparable scale, and does not have a production pricing model the CFO can underwrite. Champions generate meetings. They do not close infrastructure contracts.

The pilot pricing problem. Running a pilot at promotional pricing or at pilot-scale resource allocation creates a trust problem at contract time. Enterprises who have experienced the $0.04-to-$1.20 per-interaction jump between pilot and production environments are now structurally suspicious of vendors who cannot show production-scale pricing before the pilot begins. Vendors who cannot price honestly in the pre-sales phase are disqualifying themselves with informed buyers.

The compliance discovery problem. In the pilot-first motion, compliance requirements were discovered during or after the pilot. Post-chasm buyers with formal procurement processes now complete compliance assessments before the first technical evaluation. A vendor who cannot provide a completed security questionnaire, evidence of audit logging, and data residency documentation at the RFP stage does not advance to the technical evaluation.

The timeline problem. A 90-day pilot cadence made sense when the buyer could make a decision within a quarter. Post-chasm procurement processes at major enterprises run 6 to 18 months. The vendors who lose in this timeline are the ones whose sales motion was designed for the former pace — they run out of stakeholder momentum, lose their champion to a reorganization, or get displaced by a competitor who submitted a more complete compliance package three months earlier.

The enterprise AI adoption failure pattern documented this year is not simply a post-deployment problem. It begins in the GTM motion. The vendors who land large enterprise contracts and then struggle with activation are often selling to the wrong buyer persona — the innovation team that greenlit the vendor but does not have the authority to mandate adoption across the organization.

The New Enterprise AI GTM Framework

Rebuilding enterprise AI go-to-market for the post-chasm market requires resequencing the entire sales motion.

1. Lead with compliance posture, not model benchmarks. The first conversation with a post-chasm enterprise buyer should be about your security architecture, your data governance model, your audit logging capabilities, and your regulatory compliance evidence — not your model performance on benchmarks. Compliance posture is the prerequisite that determines whether you proceed; model performance is the differentiator within the set of compliant options. Vendors who open with demos and close with security reviews have the sequence backwards.

2. Design for procurement, not for champions. Map the procurement process before the first meeting. Identify who has budget authority, who has compliance review authority, who has final vendor approval authority, and who has informal veto power in IT security or legal. Build a parallel engagement plan that gives each stakeholder what they need to say yes. A champion-only sales motion produces champions who cannot get the contract signed.

3. Price for production from day one. Pilot pricing and production pricing should be the same pricing model, even if the initial contract volume is small. Show the production unit economics in the pre-sales conversation. This signals vendor maturity and eliminates the budget surprise that destroys trust at renewal. Enterprises who understand the full production cost model before signing are far more likely to expand than enterprises who discover the real cost six months into production.

4. Build audit trails and observability as first-class features. Every AI output, every data access, every model invocation should be logged and accessible to the customer's audit team. This is not a compliance feature — it is a product feature that enterprise buyers increasingly require as a baseline. Vendors who treat observability as an enterprise add-on are signaling that they were not designed for infrastructure-grade deployment.

5. Target infrastructure stakeholders, not innovation stakeholders. Identify the CIO, CISO, or VP of Enterprise Architecture as the primary relationship target, not just the data science lead or digital transformation director. Infrastructure stakeholders have the budget authority, the vendor approval authority, and the organizational longevity to turn a proof of concept into a multi-year contract. They also have longer sales cycles, but those cycles produce contracts with stronger expansion economics than champion-driven POC expansions.

The Buying Criteria Have Shifted

The table below summarizes the evaluation criteria that separate pre-chasm enterprise AI buyers from post-chasm infrastructure buyers.

The agentic AI production failure data shows that 88% of enterprise AI agents never reach production. The failure mode is not technical — it is organizational. The vendors who succeed at getting to production are the ones who understood the organizational requirements before the deployment, not after. In the post-chasm market, the organizational requirements are visible in the procurement process itself.

Who Wins the Post-Chasm Enterprise Market

The structural winners in post-chasm enterprise AI are not the vendors with the best models. They are the vendors who built for infrastructure buyers before infrastructure buyers became the majority.

Security-native vendors — those who built compliance architecture from day one rather than bolting it on — have a compounding advantage. Every enterprise compliance assessment they pass generates a reference. Every reference reduces the time-to-close for the next deal. The vendors who were selling to innovation labs in 2023 with lightweight security postures are now facing a retrofit problem: adding enterprise-grade security to a product not designed for it is expensive, slow, and visible to buyers who ask detailed questions.

Observability-first vendors — those who built audit logging, output traceability, and usage analytics as core product features — are in a similar position. Infrastructure buyers require observability not as an add-on but as a prerequisite. Vendors who treat it as optional have a product debt problem that shows up in every enterprise security questionnaire.

Vendors with production-validated pricing models — those who have been transparent about the pilot-to-production cost economics with early customers — have a trust advantage that is difficult to replicate. Post-chasm buyers share notes with their peers. Vendors known for honest pricing conversations get introduced by reference; vendors known for pricing surprises get screened out before the first call.

The enterprise GTM playbook transition from PLG to enterprise sales is playing out in real time in enterprise AI. The companies that built product-led growth motions for innovation buyers are not automatically equipped to serve infrastructure buyers. The motion, the team structure, the compensation model, and the product architecture all require adaptation.

The Window Is Still Open — Briefly

The JPMorgan and Novo Nordisk announcements mark the crossing of a threshold, not the closing of a market. The early majority is large — it contains many organizations that have not yet committed to AI at infrastructure scale and are still in the evaluation process. For AI vendors who adapt their GTM now, there is still time to build the compliance posture, the reference base, and the enterprise sales motion that the next wave of infrastructure buyers will require.

The vendors who wait for the market to come to them — hoping that the product-led, demo-first motion that worked in 2023 will continue to work — will discover the chasm the hard way: watching enterprise deals stall at procurement, getting displaced by vendors with better compliance documentation and longer enterprise track records, and losing renewal conversations with customers whose requirements have shifted faster than the sales motion.

Takeaway: JPMorgan's $19.8 billion technology budget now includes AI alongside data centers and payment systems. Novo Nordisk is deploying AI across its entire global value chain. These are not outliers — they are the leading indicators of a structural shift in how the enterprise majority buys AI. The go-to-market strategies built for 2023's innovation buyers are not equipped for 2026's infrastructure buyers. The vendors who recognize this distinction now, and rebuild their sales motion, compliance posture, and pricing model accordingly, will win the post-chasm enterprise market. The ones who do not will spend 2027 wondering why their pipeline stalled.