How Tenex Runs an AI-Native Engineering Team: Outcome Pay, Shared Context and Where Enterprises Are Stuck

Lieberman introduces Tenex as an applied AI partner for some of the biggest companies in the world, helping them go from AI absent to AI native. He says the company's belief is that AI is incredible but is raw intelligence, and that the hardest problem is the final mile of actually diffusing it through the economy. Tenex's view is that it should act as the player that helps enterprises work this out.

CJ Hess, who describes himself as a special projects lead at Tenex, says he was the first hire and came from a software engineering background. He says his output was improving and speeding up while his hourly billing was pulling his pay down, and that he was drawn to the company's idea of rewarding engineers for output rather than hours.

Lieberman explains that Tenex started about 18 months ago as an AI-powered development shop, with engineering as the canary in the coal mine for AI. The idea grew out of co-founder Armand, who was using coding agents heavily while some people he knew were still typing into their IDEs without adopting the technology. The company has since moved up-market: about 50% of its customers are Fortune 2000 firms and about 50% are mid-market companies with $100 million to $500 million in revenue. Customers increasingly ask for AI transformation as a service rather than only staff augmentation and AI engineering. Today the company has 75 people, about 90% of whom are engineers or AI strategists, which Lieberman describes as technical consultants.

Under the model, engineers are paid against story points, so some engineers can earn millions of dollars a year while others earn less because they are less productive. Lieberman argues that AI is a leveraged bet on a person's intellect and work ethic, so earning capacity will show more variance than before. He says the model is not for everyone, but suits people who want to go long on themselves, have a strong software foundation, keep pushing to use the frontier, and stretch their time while keeping quality high.

Asked for an update on how many people earn over $1 million a year, Lieberman says earlier there were two or three such people, and that the gap between the highest- and lowest-paid engineers at Tenex is now roughly 4x rather than the 10x that the model might suggest.

Hess adds that a feared failure mode, subjective story points, sandbagging or shipping bad code, has not materialized. He attributes this partly to AI strategists being incentivized on retention and growth. Hess notes that the hardest part is compensating time spent on non-coding work such as architecture and client meetings, which the compensation model handles through nuance.

Hess says his coding has been dropping over time. With the right context, models can now get to about 90% of a solution. He argues that the real value now sits in sitting down with a customer, hearing everything they can brain-dump, and then feeding that to the model to get closer to the solution.

Hess says Tenex may in future look more at the actual output, how it works and how it looks, rather than the traditional story point model, while still anchoring to story points as the unit of measurement.

Lieberman says everything is checked into version control, including the agents' plans and documentation, which he says are now generated more in HTML. Hess says the team has standardized a set of skills, grown into a structured set of documentation within a repo plus a CLI that the models use to drive work such as creating plans, visualizations and client readouts. As models improved, the team reduced the structural enforcement it imposes on the canonical plan document.

Lieberman raises the debate over whether docs are needed at all, since the code itself can act as the memory of a codebase. He says documentation is still needed to record why a decision was made and the context of the conversation behind it. The aim is for that context to live in one durable place so that agents can take on more of the work without being given everything up front.

Lieberman says Tenex measures whether these processes are working by tracking the trend in average story points per month for every engineer. Hess estimates the increase overall has been about 2x this year, while noting that results differ across people.

Hess says the team lets people choose tools but that Opus 5.5 is the current vibe and is the main driver for most of the team. He says Opus is fast, cheap and covers most of what used to require two or three models depending on speed, intelligence or the ability to solve specific problems. He says some of the team still reaches for Codex for work that is more computer-use oriented, and that using both models, with Codex reviewing and the two working back and forth on the machine, still produces the best results.

Hess says he runs AMP at maximum. AMP recently added the ability to connect a Claude Code subscription, which Hess says can cut costs by about 80%.

Hess says he prefers AMP because it takes everything off his computer and offers cross-model functionality: Opus can write code and another model can review it from a different environment. He says he could not replicate his local setup with Claude Code in the cloud or Claude Tag, which he sees as better suited to teams collaborating on non-technical tasks.

Asked about token spend relative to employee spend, Lieberman notes that many companies spend roughly $10 to $200 per employee on tokens, and that more AI-native companies express this as a percentage. Hess says Tenex's raw token spend is somewhere between 1:1 and 2:1 relative to employee spend, and that the team thinks about when to reach for the largest model. Lieberman says costs have recently come down substantially since the team moved to Opus 5.5, after an earlier period in which many teams used more expensive models that gave the best results.

Asked about the long-term outlook for AI businesses, where most revenue currently comes from the labs, the questioner notes that some founders believe enterprises will drive revenue in the next few years. Lieberman describes what that would require: enterprises would need to do substantial fine-tuning of their own models. He says the case for that depends on either better quality, which is a hard bet because the frontier keeps moving, or on price.

Lieberman says the price argument looks difficult because token costs keep falling across models. He suggests costs per task may have reached a kind of Pareto point, and that it is now much harder for a company to token-max and seriously harm its budget than it was six to eight months ago. Hess agrees that costs are trending down, but notes that a future with very expensive frontier models that deliver the best results could reverse that trend. He cites the recent launch of a model that is at the frontier yet cheap, and says competition in both open and closed source will keep costs falling. He adds that many economically valuable tasks can already be done with Opus 5.5.

Lieberman describes enterprise AI adoption as a series of milestones. The first is rolling out a general-purpose LLM to employees, which he says about 90% of the roughly 100 enterprises Tenex talks to have done with Claude or Codex. The next is training and enablement, which he argues is as important as the technology itself and is at around 60%. Some companies assume employees will figure it out, while others need to teach people what skills are and how the features of these tools work.

Next comes coding agents. Lieberman says that when he asks people what share of enterprises have rolled them out to the engineering product organization, most guess 70 to 80%, but the figure is closer to 50%. The steepest drop comes after that: Hess says agents in production for non-engineering work fall to about 5%, even at companies with more than a thousand employees. Lieberman says the most common request is for a back-office agent, driven by FOMO and what is trending. He says FOMO creates momentum but can also lead people to latch onto the wrong ideas.

Hess says that as every employee has Claude or Codex and needs access to different data, the question of permissioning has become a serious problem for CIOs and CISOs. Companies lack a control panel that answers simply what can be read and what cannot. Tenex started by building bespoke solutions for companies and has since productized its own gateway.

Lieberman says a recurring problem is that you cannot automate a process that does not exist. When 40 people do roughly the same job in slightly different ways, it is much harder to build an agent that works proactively in the background. Tenex's recommendation is first to build a process around the work inside the organization, then to hand the work to an agent. Without some standardization, it is impossible for the agent to know it is doing the job right, and it is hard to build evaluations around it.

Lieberman and Hess say they are generally bearish on vertical-specific AI solutions across functions and industries. Lieberman says that spending all day inside enterprises shows how much nuance, inertia, systems and people have to be managed. He adds that buyers of vertical solutions increasingly expect a one-of-one experience, which is why many software companies are building teams of forward deployed engineers and why services and software are converging.

Lieberman raises two questions: how vertical solutions can achieve software margins if every customer expects a bespoke result, and why frontier models should not simply absorb them. He names Harvey for lawyers and OpenEvidence for doctors as classic examples of companies winning their domains, and says he is generally bearish on them. He acknowledges that so far this view has been wrong. His answer on the moat is access to data that ChatGPT will not have, although he notes that Anthropic and OpenAI are also acquiring similar data sources.

The speakers agree that the total market for work software will grow, and that for many players the question becomes percentage market share under pressure. Lieberman says he loves Superhuman, which he pays for to avoid worrying about email, and that anything with good hotkeys appeals to him. Hess says products where an agent is the product doing specific vertical work will struggle unless they have strategic compliance data or another moat, because people can build their own tools quickly or the models will absorb the tasks. Apps that build a tool around a process and place an agent on top of it can still succeed.

Lieberman raises a related view: people may not want to use the agent a vendor built, and would prefer to use the agent they already have and plug in the vendor's data. In that world, a small set of companies may own the interface and everyone else plugs into it. He notes that Salesforce, a system-of-record company, is embracing that path by providing APIs and MCPs rather than insisting on its own agents.

Lieberman describes his Kill My SaaS competition, in which the team replaced a mid-tier SaaS product. He says they were quoted $40,000 a year for the tool and spent about $1,000 to replace it, and that the replacement is going into production for the next conference. Lieberman says Tenex is itself the customer that wants a one-to-one tool, and that it is building a domain- or vertical-specific product for its own business. He says so far it is going reasonably well.

Lieberman argues that the further up-market a business is, the harder the case for ripping out and rebuilding from scratch becomes. He says he does not agree that Salesforce or SAP are seriously at risk, because they accrue value in this moment, and there is so much money at stake that a company getting a rip-out wrong could mean hundreds of millions or billions of dollars.

Hess and Lieberman note that Tenex has built its own CRM and operating system from scratch. The team found what every company eventually realizes: much of the value of commercial software lies not only in the utility it provides but in the maintenance and improvement headaches the buyer no longer has to think about when the software is not core to the business.

Lieberman frames the decision as whether the prize is big enough to justify the cognitive load of nurturing something that is not core to the business. He says the $39,000 difference in annual cost implicitly shows that Tenex judges the effort worthwhile, and that the case is not only about cost but about the better experience it could deliver. Hess says the question is whether a product has been mediocre for a long time and what a 10x better version would look like. Lieberman says startups and nimble mid-sized companies can move that way more easily than enterprises, which are entrenched in their capabilities and rarely build a better version of a SaaS product they use.

Lieberman says he does not like treating AI's effects as something that just happens to people, because it gives too little credit to people's agency in their own lives. He says reskilling is both the employer's job, because it wants to drive growth, and the employee's job. He argues that saying AI will replace only the work people do not enjoy is a form of cope for the reality that people will be put out of jobs, because many people work to make money rather than because they like the job.

Lieberman encourages people to ask themselves how stable or replaceable their job is over the next few years, and what skilling up to be better protected would involve.

The speakers say the conversation in San Francisco is dominated by agents, with discussion centered on which model is best rather than whether people like AI. Lieberman says New York has a smaller AI bubble than San Francisco. He describes Tenex as something like a microcosm of San Francisco because of the people who self-select to work there, and the companies and people it engages in New York.

Lieberman says San Francisco feels all-encompassing, comparing it to AI Engineer on steroids, and says many attendees overlap with people who have lived in San Francisco at some point. He notes that AI Engineer represents roughly 1% of the city during World's Fair week. He argues that a group of people who live in the future can also be found in New York, and that Tenex's goal is to make New York a hub for AI in the US, because there is so much commerce there that diffusing the technology could meaningfully affect the country's GDP. He says the team takes it as a charge to grow the bubble.

Tenex is hiring forward-deployed engineers and AI strategists, who are technical consultants leading client engagements. The New York team is 75 strong, and Lieberman says the company intends to keep New York as its home. He invites enterprises that have tried AI and found it is not working to get in touch. Hess says he is particularly looking for engineers. He describes his special projects as work outside client engagements, including building the internal meta-harness system that finds the frontier of what models can do and gives them tools to do it better, and an internal platform to deploy background agents to enterprises in ways that are genuinely helpful.