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

What it claims

Tenex, a 75-person applied AI firm in New York, says paying engineers by story points works, and that engineer output roughly doubled over the year after Opus 4.5 (December 2025). The interview also gives Tenex's figures on enterprise AI adoption. Of about 100 enterprises it talks to, 90% have rolled out Claude or Codex to employees and about 60% run training and enablement. Roughly 50% have deployed coding agents to engineering and product teams. About 5% have an agent in production doing non-engineering work. Customers are split 50/50 between Fortune 2000 companies and mid-market firms with $100M to $500M in revenue. Of the staff, 90% are engineers or "AI strategists," who act as technical consultants.

How it works

Compensation. Story points drive pay, and the gap between the highest- and lowest-paid engineer is about 4x, not 10x. AI strategists are paid on customer retention and growth. Tenex says this cross-incentive keeps engineers from inflating estimates or shipping weak code. The unsolved problem is paying for time not spent coding, such as architecture and client meetings. Point criteria are recalibrated at least quarterly because better models should deflate what a point is worth. The measure of whether a process works is the trend in average story points per engineer per month.

MetaHarness. It started as a shared set of agent skills. It grew into structured documentation inside each repo plus a CLI that agents use to produce plans, visualizations and client readouts. Each project has a canonical plan document. Tenex has loosened the enforced structure of that document as models improved. Agents follow a set convention for logging decisions and progress. Everything is checked into Git, increasingly as HTML "picture books" instead of Markdown. The stated rationale is that code alone does not record why a decision was made. Tenex's own SDLC article notes this much structure can be overkill for a small team on one project.

Tooling. Most of the team drives with Opus 5.5 and uses Codex for review and computer-use tasks, with the two working back and forth. CJ Hess runs Opus through AMP's cloud "orbs" using a linked Claude Code subscription. The host estimates that setup cuts cost by about 80%.

Evidence

All figures are self-reported in conversation, with no method given. Tenex puts its raw token spend at between 1:1 and 2:1 against employee spend. The host gives a frame for that ratio: most companies spend $10 to $200 per employee on tokens, and AI startups spend 1x to 10x employee cost. Tenex says moving off Astra and Fable to Opus 5.5 cut costs sharply.

In Latent Space's Kill My SaaS experiment, the host was quoted about $40,000 to $43,000 a year for a speaker-management tool and spent $1,000 replacing it. The contest drew 72 submissions. The winner chosen by human judges ranked 19th on the AI judge's leaderboard.

Limits

The output and pay figures have no baseline beyond "this year." The 2x is CJ Hess's estimate, and he says it varies widely by person. The transcript is auto-generated, contains errors and does not label speakers.

Tenex's customer findings are qualitative. Back-office agents are the most common request, and data permissioning (RBAC) for company-wide Claude or Codex access is a major CIO and CISO pain. Tenex now sells its own gateway product for that. CJ Hess warns that agents fail where 40 people each do the same job differently: "you can't automate a process that does not exist." Tenex's advice is to standardize the process first so evals are possible.

Alex Lieberman argues that replacing SaaS gets riskier further upmarket and that maintenance cost often outweighs savings. Tenex built its own CRM and internal OS. Alex Lieberman is bearish on vertical AI such as Harvey and OpenEvidence. The host disputes that view.