Applied AI firm Tenex, led by Alex Lieberman with engineering leader CJ Hess, outlined how they scale an unusual incentive model—paying engineers for delivered outcomes rather than hours—and how they make project context durable for both humans and agents.
Paying for output, not hours
Tenex long ago adopted a compensation model centered on story points instead of hourly billing. The rationale is simple: if AI lets engineers finish work faster, an hourly rate can reduce their earnings for the same job. By tying pay to output, Tenex aims to reward productivity.
In practice, Tenex still uses story points as a core component of pay. Classic failure modes—sandbagging estimates or shipping low-quality code to collect points—haven’t materialized as the main problems because Tenex’s AI strategists are measured on customer retention and growth, which aligns incentives toward legitimate value delivery. The harder challenge is compensating non-ticketed work: understanding customer needs, designing architectures, and deciding what should be built. As models improve, engineers spend relatively less time directly coding and more time converting customer context into model-ready inputs, so any output-based comp system must account for that work.
Company size and customer mix
According to Alex Lieberman, Tenex has grown to 75 people. Their customer base is split roughly 50/50 between Fortune 2000 companies and mid-market businesses (typically $100–500 million in revenue). About 90% of the team are engineers or AI strategists (technical consultants). Tenex began as an AI-powered dev shop and now increasingly delivers enterprise-wide AI transformation services.
Compensation lessons and productivity gains
Tenex reports roughly a 4x pay dispersion between its highest- and lowest-paid engineers—less than the oft-cited "10x" but still a wide spread. The firm also observes productivity gains: over the past year engineers’ average monthly story points roughly doubled, driven by model improvements (the team cites Opus 5.5 among their primary models).
Because model capabilities evolve, Tenex recalibrates story-point valuations at least quarterly. The firm also tracks metrics like average story points per engineer and token-to-employee spend to understand ROI and cost dynamics.
MetaHarness: putting plans and decisions in the repo
CJ Hess described Tenex’s MetaHarness: a shared set of skills, canonical project documents and a CLI that surround the work. Plans, decisions, acceptance criteria and other context are checked into Git alongside code so that humans and agents returning to a task can read not only what was built but why the team made those choices.
Tenex’s description of an AI-native SDLC treats the project record as a single source of truth for agents and humans, with integrations and views into tools like Linear and GitHub. The approach increases legibility of surrounding work (plans, evidence, decisions) as distinct from merely making agents write code faster. The company cautions, however, that this level of structure can be overkill for small teams or one-off projects.
Kill My SaaS: the cost of owning what you build
In a practical experiment called Kill My SaaS, the team tested replacing a $43,000-a-year speaker-management SaaS. The AI-driven replacement cost approximately $1,000. The winning submission by human judges ranked 19th on the AI judge leaderboard, underscoring that verifying scenario coverage is different from proving users will want and use a product.
Alex’s counterpoint is that as you move upmarket the complexity, cost and risk of ripping out established systems grow—there is ongoing maintenance and integration burden to consider. The experiment’s defender argued that owning the workflow lets a team iterate faster and improve experiences for speakers and operators, and that potential must be weighed against the maintenance Alex described.
The debate on vertical AI
Tenex’s leaders are broadly skeptical about purely vertical AI products if buyers expect highly customized, one-off solutions: such expectations make it hard to deliver at software margins, and as frontier models improve, many domain tasks may be absorbed by general-purpose tools.
Their interlocutors pushed back with examples (Harvey, OpenEvidence) where domain data, specialized interfaces and integrations create real moats. CJ suggested a useful distinction: a vertical product whose main feature is an agent performing a task versus a product that reshapes the workflow around the task. The more interesting question, he argued, is which products become durable places where people do their work and which become features of the agents they already use.
Models, token costs and operations
Tenex’s teams use models such as Opus 5.5 and occasionally other specialist models (e.g., for code-heavy tasks). The company estimates its raw token-to-employee spend ratio at roughly 1:1 to 2:1. Competition and model improvements have driven per-task costs down, and Tenex has shifted to cheaper models for many tasks without sacrificing ROI. They also noted the operational need to control token spending and tune when to call the most expensive models.
Jobs, reallocation and reskilling
Tenex frames AI-driven change as a shift that will displace some roles but also create opportunities for redeployment. They emphasize reallocation—moving people away from repetitive work to higher-value problems—and the joint responsibility of employers and employees for upskilling.
Takeaway
Tenex’s approach combines an output-based pay philosophy with disciplined project recording (MetaHarness) to make AI-assisted work auditable and repeatable. The company’s growth to 75 people and its split customer base show demand for both engineering augmentation and broader AI transformation, while their experiments (like Kill My SaaS) and debates on vertical AI highlight the trade-offs between buying, building and owning software in an AI-native world.
