Most large companies have handed their employees an AI assistant. Very few have put an AI agent to work on anything other than software, according to Alex Lieberman, a co-founder of the AI consultancy Tenex. Speaking on the Latent Space podcast, in an episode recorded at the AI Engineer conference in New York and published October 21, 2026, Lieberman put rough numbers on the gap.
The figures are his estimates from Tenex's own client conversations, not a survey. He said the firm is talking to "let's call it 100 enterprises" at the moment.
What Tenex does
Lieberman described Tenex as "an applied AI partner for some of the biggest companies in the world to help them go from AI absent to AI native." In his view AI is "raw intelligence," and the hardest problem is "the final mile" of getting it to spread through the economy. The company started about 18 months ago as what he called an AI-powered dev shop, meant to speed up the engineering teams of companies that were not yet using coding agents. As it moved toward larger clients, they began asking for "AI transformation as a service." Today, he said, about half of Tenex's customers are Fortune 2000 companies and the other half are mid-market firms with $100 million to $500 million in revenue. The firm has 75 people, 90% of them engineers or AI strategists. Tenex's website describes its work as identifying AI opportunities, changing workflows and delivering software through engineering teams.
Four milestones in a race
Asked what customers most often want, Lieberman described enterprise AI adoption as "milestones in a race." He said people at the conference were surprised by how few companies had reached the later ones.
The first milestone is rolling out a general-purpose AI model to staff, a tool any employee can use. By his count, 90% of the enterprises Tenex talks to have given employees either Anthropic's Claude or OpenAI's Codex.
The second is teaching people to use it. Training and enablement "never gets the love that it should get," Lieberman said, though he called it "arguably more important or as important as the technology itself." He put this at about 60%. Some companies, he said, assume "our people will just figure it out," while a good share decide they need to teach staff what skills are and how the tools' features work.
The third is coding agents: AI systems that write and change software with some independence, given to a company's engineering and product teams. When Lieberman asked people at the conference to guess the share of enterprises that had done this, he said most answered "it has to be 70 or 80%." His estimate is closer to 50%. Every other enterprise Tenex speaks to, he said, either has no coding agent or uses something "closer to not having a coding agent than having a coding agent."
The fourth is the one he called the biggest "diffusion gap": an agent in production doing non-engineering work. An agent here means software that carries out a task on its own, using tools and data, rather than only answering questions in a chat window. Lieberman put this at about 5%. Most enterprises Tenex talks to with more than a thousand employees, he said, "do not have an agent in production that is doing work that is not building software."
What clients ask for
Lieberman said the current wave is driven by a small group of people who understand what the technology can do and "a way larger group of people driven by FOMO." He did not count the fear of missing out as a bad thing: "It creates some forward momentum, but it also means that people will latch onto the things that are most in the zeitgeist." People have been saying it is the year of the agent for about a year and a half, he said. "Every year is the year of the agent."
One of the most common requests follows from that. Enterprises ask Tenex to build an agent for some back-office task, the internal administrative work that keeps a company running. Lieberman said much of Tenex's work now involves deploying agents outside the engineering organization for repeatable work.
The second big request is about permissions. Once every employee has Claude or Codex and needs access to certain company data, Lieberman said, managing role-based access control (RBAC, the rules that decide which roles can see which data) has "turned into an absolute nightmare" for chief information officers and chief information security officers. They lack a good gateway, he said, something "almost like a control panel for data permissioning." The conversation boiled the problem down to which information each person or tool can and cannot read. "And that is a huge issue," Lieberman said.
Tenex first built custom permission setups for individual companies, he said, and has since turned that work into its own gateway product. Later in the episode, the firm's internal special projects were described as including that platform, with the goal of getting background agents, which work on their own without someone prompting them step by step, into enterprises and solving the messy IT problems discussed earlier.
AI vendors are working on the same problem from their side. Anthropic's Claude Tag, a shared agent for Slack announced in June 2026, lets administrators create separate agent identities for different functions, each with its own memory and tools. Anthropic's example is a company's sales and engineering teams: the engineering agent should not inherit sales memories or give engineers access to sales tools and data.
A process has to exist first
The discussion then turned to a problem Tenex keeps running into when it deploys agents: a process that does not exist cannot be automated. The example given was a team of 40 people doing roughly the same job, each in a slightly different way. Tenex can sit down with those people, give them tools and show them how to improve their own workflow. Building an agent that does the job proactively in the background is much harder.
The reason is measurement. Without a standard way of doing the work, the agent has no way to know whether it is doing the job right, and Tenex cannot build evals, the tests that check an agent's output against an expected result. The bigger payoff, as described in the episode, comes when agents bring work to people, who review it and send it back if needed, turning employees into what has been called "agent managers."
So Tenex's recommendation to clients is often to build a defined process for the team first, then lay out the steps for handing that work to a background agent. Its transformation methodology follows a similar order: employee surveys, stakeholder interviews and baseline measurements come first, with automation introduced alongside training and change management.