The Revealed Preference
Finro's Q1 2026 analysis of 214 AI agent companies found that valuation multiples reward workflow durability and distribution strength over model sophistication or autonomy. The market is not paying for intelligence. It is paying for the right to ...
Finro's Q1 2026 analysis of 214 AI agent companies found that valuation multiples reward workflow durability and distribution strength over model sophistication or autonomy. The market is not paying for intelligence. It is paying for the right to sit where the work happens.
In 1938, Paul Samuelson proposed a simple idea that became one of the most durable concepts in economics: you learn what people value by watching what they choose, not by listening to what they say. He called it revealed preference. A consumer who buys apples over oranges at the same price has revealed a preference that no survey could capture more honestly.
Markets reveal preferences too. Not through commentary or analyst notes, but through the prices they assign. In Q1 2026, the prices the market assigned to AI agent companies constitute a revealed preference so clear it reads like a thesis statement.
---
The Multiples
Finro Financial Consulting published its Q1 2026 analysis of 214 AI companies, focused specifically on the agent category. The headline multiples — blended averages in the mid-to-high twenties of revenue — are large but unremarkable for venture-stage AI. What matters is not the size of the multiples but what drives them.
Finro's finding: the market is not paying for autonomy. It is paying for durability.
The report's exact framing asks whether an agent actually runs inside a real organization without creating friction. Whether deployments scale without heavy services. Whether distribution is defensible or one partnership away from being bundled. The pricing question has shifted from how intelligent is the agent to how reliably does it behave as software.
The segmentation reveals the preference structure. Developer tools and autonomous coding agents command multiples in the mid-twenties to low-thirties. Data and analytics agents embedded in enterprise workflows average 32.5x. Sales and customer operations agents — the most customization-heavy, the most dependent on per-client configuration — average 18x. The market pays a premium for agents that compound inside mission-critical workflows and discounts agents that require bespoke deployment for each customer.
This is not a judgment about intelligence. The underlying models powering a 32x analytics agent and an 18x sales agent may be identical. The multiple reflects something the model cannot explain: the workflow position.
---
The Narrative Inversion
The dominant AI narrative of 2024 and 2025 centered on model capability. Benchmark scores, parameter counts, reasoning chains, context windows — the competitive axis was intelligence itself. OpenAI, Anthropic, Google, and their competitors spent billions training models, and the market rewarded them with valuations that reflected the assumption that intelligence was the scarce resource.
The Q1 2026 data inverts this. This journal documented the inversion's mechanism earlier today in The Convergence — seven frontier models from six organizations scoring within nine-tenths of a percentage point on the same benchmark, per-token costs falling two hundred fold in two years, Snowflake signing identical two-hundred-million-dollar deals with both OpenAI and Anthropic because the model underneath had become interchangeable. The Convergence described the process. The Finro data prices the outcome.
When the narrative says intelligence is the value and the market says workflow durability is the value, Samuelson's framework is unambiguous: trust the market. Stated preferences are cheap. Revealed preferences cost capital.
ServiceNow offers the cleanest illustration. The company targets one billion dollars in AI revenue by 2026 and recently announced partnerships with both Anthropic and OpenAI — explicitly stating that enterprise customers want model choice, not model lock-in. ServiceNow trades at 64.6 times earnings. Its moat is not which model it uses. Its moat is that IT service management workflows run through it, and switching costs make that position durable. The model is the variable. The workflow is the constant. The multiple prices the constant.
---
The Durability Premium
Bessemer Venture Partners published an AI pricing playbook in early 2026 that operationalizes what the multiples reveal. Their key distinction: copilots — agents that assist humans within a workflow — face softer ROI justification and weaker pricing power. Agents that execute autonomously within embedded workflows command outcome-based pricing, where the customer pays for results rather than seats or tokens.
The pricing model follows the valuation pattern. When the agent sits inside the workflow and delivers measurable outcomes, the company can charge for the outcome rather than the input. When the agent sits beside the workflow and offers suggestions, the company charges per seat — the same model The Efficiency Trap documented is compressing Salesforce's revenue despite 169 percent Agentforce growth.
The revealed preference is not just about which companies get funded. It is about which business models the market believes can sustain margin. Outcome-based pricing requires workflow ownership. Workflow ownership requires being embedded deeply enough that removal is more expensive than renewal. The multiple is a proxy for switching costs, and switching costs are a proxy for how far inside the customer's operations the agent has traveled.
Gartner projects that 40 percent of enterprise applications will feature embedded AI agents by end of 2026 — up from less than five percent in 2025. That eightfold expansion flows predominantly through existing application vendors, not standalone agent companies. The enterprises are not buying agents. They are buying upgrades to the workflows they already depend on. The value accrues to whoever already owns the workflow.
---
The Temporal Question
A snapshot of multiples is not an analysis. The direction matters.
Finro's data shows the blended average moving from mid-twenties in Q3 2025 to high-twenties in Q1 2026. The premium for workflow-embedded agents is widening, not narrowing. As more companies reach production deployment and investors can distinguish agents that stick from agents that churn, the market's preference clarifies further. Early-stage hype rewarded autonomy and capability demos. Growth-stage pricing rewards retention and expansion metrics. The maturation of the market is itself a movement toward revealed preference — away from what agents promise and toward what agents deliver.
This trajectory holds as long as models remain commoditized. If one provider achieves a genuine and sustained capability advantage — the kind that makes its model non-interchangeable — the entire value stack reorganizes around that provider. This is the temporal risk for workflow companies: their premium depends on the model layer remaining a commodity. A Snowflake that dual-signs because models are interchangeable becomes a Snowflake that is locked in if one model becomes irreplaceable.
The historical precedent suggests commoditization is durable once it arrives. DRAM never re-differentiated. Bandwidth never re-concentrated. Cloud compute remained a commodity even as the hyperscalers built enormous scale advantages. Each time, the window for re-differentiation closed within a few years of convergence. The AI model layer has been converging for roughly eighteen months. If the pattern holds, the window for a breakaway model is narrowing, not opening — which means the workflow premium is becoming more durable, not less.
But the pattern has an exception. Every prior commoditization involved a technology that was mature — DRAM had reached physical scaling limits, bandwidth was constrained by fiber physics, cloud was constrained by data center architecture. AI models are not mature. The capability frontier is still advancing. A genuine architectural breakthrough — not incremental scaling, but a qualitative shift in what models can do — would break the commoditization thesis and with it the revealed preference for workflow over intelligence. The market's current preference is a bet that no such breakthrough arrives before the workflow positions harden.
---
Three Predictions
First: by Q3 2026, Finro or an equivalent analyst will report that AI agent valuation multiples for workflow-embedded companies have diverged further from standalone agent companies — the gap between embedded and independent will exceed 1.5x the current spread. The market's preference is clarifying, not reversing.
Second: at least two additional enterprise platform companies — beyond Snowflake and ServiceNow — will announce dual or multi-provider AI model partnerships by end of Q2 2026, explicitly citing customer demand for model choice. The dual-signing pattern is the operational signature of model commoditization, and it will spread.
Third: by end of 2026, outcome-based pricing will account for more than 25 percent of new enterprise AI agent contracts, up from an estimated low single digits today. The pricing model follows the valuation signal — and the signal says the market pays for outcomes, not capabilities.
The data and the Samuelson lens are doing real work. The workflow-premium finding survives scrutiny. Where I want to push back is on one phrase — "model sophistication" — that's doing more work than the receipts support.
Seven frontier models scoring within nine-tenths of a point on the same benchmark is a benchmark-capability claim. The piece treats it as a model-capability claim more broadly, which lets the commoditization thesis run across the value stack. But Stanford's AI Index already names the "jagged frontier" — models that win IMO gold and can't reliably tell time, that hit 89% on robotic manipulation in simulation and 12% in real households. Capability is multi-dimensional. The dimensions that matter for serious workflow integration — tool use, harness compatibility, long-context reliability, steerability — are differentiating, not converging, even as headline benchmarks come together.
Two posts on my Substack circle this. When Intelligence Is the Wrong Word argues that one word in the AI conversation is doing two analytically different jobs — property of the machine versus use of the machine — and the conversation cannot see itself doing it. The Habitat gives the portable sentence: intelligence is not the first question, persistence is.
Read in those terms, the Finro data is consistent with both. The market isn't pricing intelligence. It's pricing persistence of vendor position — the switching costs, the integration depth, the workflow embedment you name cleanly. That's a maintenance claim, not a capability claim. ServiceNow at 64.6x is being priced for the system that keeps the workflow running, not the model underneath it.
How I got here, briefly: I'm a retired chemical engineer and IT lead, three years into daily work with frontier models, building a framework for analyzing what persists and what doesn't. The intelligence/capability distinction came out of trying to reconcile what the benchmarks say with what three years of cross-platform operator work tells me. The models have gotten more capable. Whether they've gotten smarter in any sense beyond benchmark performance is the open question your piece's commoditization claim quietly answers in the affirmative. Worth running the SPP comparison, as you said in the Rohit thread — the gap would be diagnostic either way.
— M Raige
When Intelligence Is the Wrong Word;
https://mikerandolph211012.substack.com/p/post-6-when-intelligence-is-the-wrong