BusinessMarkets, money & strategy
Tenex says 90% of its enterprise prospects have rolled out Claude or Codex, but only about 5% run non-engineering agents in production
AI consultancy Tenex says most large clients have deployed Claude or Codex, yet few run agents in production beyond software work. It traces the gap to unstandardized processes and data permissioning, and describes its own Git-based agent workflow and output-based engineer pay.
From the Latent Space podcast episode of October 21, 2026 and cited sources: Tenex describes an adoption funnel based on about 100 enterprise conversations. [1] [5]
SOURCES[1] AI for the Fortune 2000: How Tenex Builds an AI-Native… latent.space[5] The State of AI in 2025 (McKinsey) mckinsey.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Tenex describes an adoption funnel based on about 100 enterprise conversations. [1] [5]
- Original source
- Latent Space podcast episode, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 21, 2026 podcast episode published
- Covered
- Oct 21 edition
- Evidence
- 7 sources across 6 sites; transcript sourced
- 1AI for the Fortune 2000: How Tenex Builds an AI-Native…latent.space
- 2Paying Engineers like Salespeople (AI Engineer)ai.engineer
- 310x AI Engineers with $1M Salaries (Latent Space)latent.space
- 4Building an AI-Native SDLC (Tenex)tenex.co
- 5The State of AI in 2025 (McKinsey)mckinsey.com
- 6Google Partner Tenex Raises $250 Million (Bloomberg)bloomberg.com
RisingHow it’s measured
YouTube video · 33.9K views · 120 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
Tenex describes an adoption funnel based on about 100 enterprise conversations. About 90% have given staff Claude or Codex, roughly 60% train employees to use them, about 50% have deployed coding agents, and only around 5% run an agent in production for non-engineering work. McKinsey's 2025 survey measures something different: 23% of organizations scaling an agentic system somewhere, and no more than 10% in any single function. [1] [5]
According to Tenex, the main obstacles are organizational rather than about the models. Agents and their evals are hard to build when many staff do the same task in different ways. Giving every employee assistant access has also made data permissioning a major problem for CIOs and CISOs. Tenex first built custom RBAC gateways for clients and has since turned that work into its own product. [1]
Tenex's internal engineering workflow keeps project context as versioned markdown in Git. The record covers current docs, epics, specs, a decision log and a scratchpad, and agents use a CLI and reusable skills to query project state, validate work and sync with Linear and GitHub. The team says it has enforced plan structure less strictly as models have improved. [1] [4]
Tenex pays engineers per story point completed. The gap between its highest- and lowest-paid engineers is about 4x, not the 10x suggested in earlier framing. The model has needed changes to pay for architecture work and client meetings. Tenex says average monthly output per engineer has roughly doubled since the Opus 4.5 release in December 2025. [1] [2] [3]
The evidence comes from one consultancy's client conversations and its internal metrics.
Read the full assessment
If the pattern holds more widely, the gap between rolling out AI assistants and running production agents depends mainly on standardized processes, evals and data permissions, not on model capability.
Executive brief
Tenex, a 75-person New York firm that helps large companies adopt AI, says that among roughly 100 enterprises it is talking to, 90% have rolled out Claude or Codex to staff. Only about 5% run an agent in production for work other than software engineering (transcript, 19:45–21:33). Tenex also pays its engineers by output (story points) instead of by the hour. It reports a 4x gap between its highest- and lowest-paid engineers and says average engineer output roughly doubled over 2026. All of these figures come from the company and none has been independently checked. Do not confuse Tenex (tenex.co) with TENEX.AI, a separate cybersecurity startup.
What changed and event timeline
Story-point pay pitched on stage
At AI Engineer Code 2025, co-founder Arman Hezarkhani explained the model: clients pay Tenex per story point delivered, and engineers are paid per story point they complete ().
$1M engineer forecast
On an earlier Latent Space episode, the founders said they expected several engineers to earn more than $1M in cash the next year from story points alone ().
Long-running agents arrive, per Tenex
CJ Hess calls Opus 4.5 the point when models began handling long-running tasks, and dates the doubling in output from then ().
Tenex publishes its process
The company described its "MetaHarness" in eight phases, built on a versioned markdown project record, a Tenex CLI and more than 50 reusable skills ().
Latent Space follow-up
Alex Lieberman and Hess give updated figures on headcount, pay spread, enterprise adoption and model choice ().
Capabilities and access
- Tenex sells "AI transformation as a service."
- It began by building custom data-permissioning (RBAC) gateways for clients and has since turned that into a product of its own.
- Model choice, per Hess: Opus 5.5 is the team's default.
Read the full section
- Tenex sells "AI transformation as a service." About half its customers are in the Fortune 2000; the other half are mid-market firms with $100M–$500M in revenue (transcript, 04:05).
- It began by building custom data-permissioning (RBAC) gateways for clients and has since turned that into a product of its own. Few public details exist (transcript, 22:08–22:46).
- Model choice, per Hess: Opus 5.5 is the team's default. Codex is used for review and computer-use work, and AMP's cloud "orbs" run the agents (transcript, 11:55–13:41).
Technical analysis for researchers and developers
- The project record is split into folders for current truth (
1-docs/), epics (2-epics/), specs (3-project-specs/), a decision log (4-log/) and an ungoverned scratchpad (6-scratchpad/). - Agents use a CLI to query project state, validate their work and sync with Linear and GitHub.
- Hess says the team has reduced how strictly plan structure is enforced as models have improved (transcript, 08:20).
Read the full section
- Context as code. The project record is split into folders for current truth (
1-docs/), epics (2-epics/), specs (3-project-specs/), a decision log (4-log/) and an ungoverned scratchpad (6-scratchpad/). All of it is checked into Git, so agents can see why decisions were made as well as what was built (Tenex blog). - Machine-readable interfaces. Agents use a CLI to query project state, validate their work and sync with Linear and GitHub. Project status is derived from these records instead of reported by hand (Tenex blog).
- Plan structure. Hess says the team has reduced how strictly plan structure is enforced as models have improved (transcript, 08:20).
- Reproducibility. The blog post contains no quantitative evaluation (Tenex blog).
Claims and evidence
- Output doubled during 2026. This is measured as average story points per engineer per month (transcript, 10:44).
- Token spend roughly 1–2x employee spend, stated without exact figures (transcript, 15:29).
- Enterprise adoption funnel (90/60/50/5%). This comes from Tenex's own conversations, not a survey. The two use different definitions.
Read the full section
- Output doubled during 2026. This is measured as average story points per engineer per month (transcript, 10:44). Tenex sets the point estimates itself and recalibrates them each quarter (transcript, 11:19), so the figure is self-defined and only partly comparable over time.
- Token spend roughly 1–2x employee spend, stated without exact figures (transcript, 15:29).
- Enterprise adoption funnel (90/60/50/5%). This comes from Tenex's own conversations, not a survey. For comparison, McKinsey's 2025 survey found 23% of organizations scaling an agentic system somewhere, and no more than 10% doing so in any single business function (McKinsey State of AI 2025). The two use different definitions.
- Kill My SaaS. The host says a $40,000 annual quote was replaced with about $1,000 of build cost (transcript, 30:35). The show notes put the quote at $43,000 (episode page).
Context and prior work
- The story-point pay model grew out of Hezarkhani's previous company, Parthian. There, cutting the team by 90% and rebuilding the process around AI reportedly raised output 10x (Latent Space, Nov 2025).
- Name collision: TENEX.AI, an AI security-operations company, raised $250M in 2026 (Bloomberg; PYMNTS). It is unrelated to Tenex.co.
Limitations, safety and contested findings
- The gap is 4x, not the "10x" in the earlier framing.
- Lieberman is bearish on vertical AI tools. The host disputes this, citing Harvey's and OpenEvidence's success, proprietary data moats and specialized interfaces (transcript, 25:47–27:33).
- Lieberman calls the idea that AI will only take over work people dislike "a bit of cope."
Read the full section
- Pay spread. The gap is 4x, not the "10x" in the earlier framing. Tenex has had to adjust its pay model to cover time spent on architecture and client meetings (transcript, 05:20–06:30).
- Vertical AI. Lieberman is bearish on vertical AI tools. The host disputes this, citing Harvey's and OpenEvidence's success, proprietary data moats and specialized interfaces (transcript, 25:47–27:33).
- Job displacement. Lieberman calls the idea that AI will only take over work people dislike "a bit of cope." He argues that both employers and employees share responsibility for reskilling (transcript, 37:17).
- Inflation risk. The 2x output claim depends on point estimates that Tenex both assigns and revises (transcript, 11:19).
Business and practitioner implications
- Hess's rule is "you can't automate a process that does not exist."
- Governance is the bottleneck. Giving every employee access to Claude or Codex has made data permissioning "a nightmare" for CIOs and CISOs (transcript, 22:46).
- Cost controls.
Read the full section
- Standardize before automating. Hess's rule is "you can't automate a process that does not exist." When 40 people do a task 40 different ways, it is hard to write evals for an agent (transcript, 23:22).
- Governance is the bottleneck. Giving every employee access to Claude or Codex has made data permissioning "a nightmare" for CIOs and CISOs (transcript, 22:46).
- Cost controls. Falling per-task token costs mean one heavy user can now do less financial damage than six to eight months earlier, according to Lieberman (transcript, 17:52).
- Build vs. buy. Replacing SaaS gets harder the further up-market a company is. Owning internal tools carries a maintenance cost of its own (transcript, 31:48–33:41).