AI for the Fortune 2000: How Tenex Builds an AI-Native SDLC — Alex Lieberman & CJ Hess

Alex Lieberman reintroduces Tenex as an “applied AI partner” for some of the world’s largest companies. The company’s premise is that frontier models provide raw intelligence, but the difficult “final mile” is diffusing that intelligence through an enterprise’s actual systems, processes, and workforce. Tenex began about 18 months earlier as an AI-powered development shop, initially helping companies whose engineering teams had not adopted coding agents. As Tenex moved from startups and mid-market companies toward larger customers, the work expanded from staffing and AI engineering into “AI transformation as a service.” Alex says 50% of current customers are Fortune 2000 companies and the other 50% are mid-market companies with roughly $100 million to $500 million in revenue. The company now has 75 employees, 90% of whom are engineers or AI strategists—technical consultants who lead transformation work. Alex frames Tenex’s role as helping enterprises figure out how to use AI throughout the organization rather than merely adding an AI feature to an existing application.

CJ Hess says he was Tenex’s first hire, after trying several startups and eventually running an engineering consulting company. As a software engineer, he found that AI was making his work “better and faster,” while his hourly billing model meant he was effectively getting paid less for producing more. Tenex appealed to him because its founding philosophy was to reward engineers for output rather than time. Tenex uses story points as the principal unit for compensation. Engineers agree with the company on the size of the work, and compensation is tied to the points delivered; as a result, some engineers can earn millions of dollars per year while others earn substantially less. Alex says the highest- and lowest-paid engineers currently differ by about 4X, not 10X. He describes the model as a leveraged bet on an individual’s intellect and work ethic: AI increases the possible output, but also increases the variance in earning capacity. The model is intended for engineers with a strong software foundation who actively push themselves to use frontier tools while maintaining quality.

Alex says the expected objections to output-based pay were that story points are subjective, engineers would inflate estimates, or they would ship low-quality code merely to check boxes. He says those problems have not been the main issue, partly because Tenex’s AI strategists are incentivized around customer retention and growth, creating a counterweight to engineers who might otherwise optimize only for points. The harder compensation question is how to pay for work that is not directly coding: architecture, client meetings, discovery, and other activities that enable the eventual output. CJ says models can now get to roughly 90% of a solution when given the right context, making the human value increasingly concentrated in sitting with a customer, absorbing a detailed brain dump, and translating that context into a useful implementation. Tenex still anchors compensation to story points, but it periodically recalibrates what a point means. Alex estimates that overall engineering output has roughly doubled during the year, especially since models became better at long-running tasks. The company reviews its estimation criteria at least quarterly to balance rewards for output with the changing value of AI-assisted work.

CJ explains that Tenex began standardizing engineering skills and processes into a structured repository, supported by a command-line interface. The system guides agents through planning, visualization, client readouts, implementation, and logging decisions. A canonical plan document gives projects a common shape, although Tenex has reduced the amount of enforced structure as models have improved. Once a plan exists, agents build it out while recording what they are doing and why. The plans, decisions, and other artifacts are checked into the repository, increasingly as Markdown or HTML, so the code is not the only source of memory. CJ argues that teams need durable records of why a decision was made and what context surrounded it; otherwise that knowledge remains in conversations or in someone’s head. Keeping it in the repo gives agents and internal tools access to context that many engineering organizations lack. Tenex measures whether the process works partly by watching average story points per engineer trend upward. CJ describes his current special-project work as pushing this “MetaHarness” toward the bleeding edge: giving models the tools and workflow structure needed to perform longer, more complex tasks.

CJ says Opus 5.5 is currently the preferred model across much of Tenex because it is fast, relatively inexpensive, and capable enough to cover tasks that previously required two or three different models. Engineers still reach for Codex for some computer-use tasks, and CJ says the strongest workflow often has Opus do the main work while Codex reviews it or works back and forth with it on the machine. CJ is also “amp-maxing,” using Opus through Amp. He says Amp’s cloud-based setup keeps work off his computer, and the ability to connect a Claude Code subscription substantially changed the economics for him—he estimates it cut costs by about 80%. He prefers Amp over Claude Code and Claude’s remote environment because Amp is more opinionated and supports cross-model workflows without trying to be a general-purpose solution. Alex asks about token spending relative to employee spending. CJ estimates Tenex is somewhere between roughly 1:1 and 2:1 in raw token spend to employee spend. He says the company worried about token-maxing as headcount grew, but falling model prices and subscription arrangements have made runaway costs less threatening than they were six to eight months earlier.

Alex describes enterprise adoption as a sequence of milestones. Of roughly 100 enterprises Tenex is speaking with, he estimates 90% have rolled out a general-purpose LLM such as Claude or Codex to employees. About 60% have also invested in training and enablement, although he argues that teaching people how to use the tools may be as important as deploying the tools themselves. The percentage that has deployed coding agents to engineering teams is closer to 50%, despite outsiders often guessing 70% or 80%. The largest gap appears beyond engineering: Alex estimates only about 5% of enterprises have agents in production doing non-engineering work. Many companies ask Tenex to build an agent for a back-office task, but CIOs and CISOs face a major permissions problem when every employee has access to AI tools that need different slices of company data. Tenex initially built bespoke solutions and is now productizing its own gateway to control what agents and employees can read. CJ adds that “you can’t automate a process that does not exist.” If 40 people perform the same function in 40 slightly different ways, an autonomous agent cannot reliably know what correct execution means. Tenex therefore often standardizes the process first, creates evaluations around it, and only then hands the work to an agent.

Alex is generally bearish on vertical-specific AI products whose main proposition is an agent performing work inside a particular profession. He points to the operational difficulty of serving enterprises: every customer has nuanced requirements, legacy systems, internal politics, and a growing expectation for a one-of-one implementation. That convergence of software and services can make it difficult to preserve software margins. He also expects frontier model companies to absorb many narrowly defined tasks unless the vertical vendor has a genuine advantage such as strategically important data or compliance access. The host pushes back with examples such as Harvey for lawyers and healthcare-focused products, arguing that specialized data, interfaces, workflows, and user habits can remain valuable even when general-purpose models improve. He compares the opportunity to Bloomberg or Superhuman: users may already have general access to information or email, but will still pay for a specialized interface, integrations, and fewer mistakes. The participants agree that the market for software will grow and that reasonable people can disagree about each company’s share; their sharper distinction is between a specialized app with useful workflow and data integrations, which can succeed, and a thin vertical agent that merely asks customers to buy another $20-per-month subscription. They then discuss replacing a mid-tier SaaS product. The host says his company paid about $40,000 per year for a mediocre product and spent roughly $1,000 building a replacement, which is entering production for its next conference. Alex says replacing deeply entrenched enterprise systems such as Salesforce is much harder because the risk can involve hundreds of millions or billions of dollars. For mid-tier software, the decision depends on whether the savings or improved capability justify assigning engineers to maintain a non-core system. The host argues that the benefit is not only saving $39,000 but creating a 10X-better experience. Alex agrees that startups and high-growth companies can do this more readily than large enterprises.

On whether AI will eliminate jobs, Alex says he believes AI will produce benefits for humanity and augment many workers, while also changing industries and eliminating some roles. He points to call centers as an area already declining as AI handles more of the work. Tenex’s approach with customers is to replace a particular function and then identify the next important problem for the affected employees to solve. He argues that large companies generally have an endless supply of unresolved work, so subject-matter experts can be moved from repetitive spreadsheet-level tasks toward cross-team and strategic work. Alex also rejects describing job displacement only as something that happens to passive employees. Employers should train and reskill people because doing so helps the business, but employees also have responsibility to assess the stability and replaceability of their own jobs and decide what skills to develop. He calls the idea that AI will only remove work nobody enjoys “cope,” since many people perform jobs for income regardless of whether they like the tasks. The conversation closes with New York’s AI culture. Alex says New York has a smaller AI bubble than San Francisco but a larger commercial environment, making it strategically important for AI diffusion. Tenex intends to expand that bubble, bringing more people into an AI-native way of working while using New York’s business concentration to connect technology with customers.