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

Full analysis

Two co-founders of a 75-person AI engineering shop just handed us a field report from inside 100 large companies, and the number that jumps out is 5%. That's how many have an AI agent running in production for anything other than coding. Everyone has a chatbot. Almost nobody has automated real work. Tenex's Alex Lieberman and CJ Hess frame that gap as their whole business. For a reader who buys and builds with AI, it's a reality check on how far "enterprise AI" actually is from the demos.

Here's what that gap means, who it helps, and what to check before you believe the productivity numbers.

The Skeptic

Outcome-based pay for engineers surfaces real output, so the 2x-productivity claim has better plumbing than most vendor stats. But look at how it's built. CJ Hess says tasks now need less effort, so Tenex calibrates story points down quarterly. If you shrink the yardstick and the count doubles, how much is real throughput versus the ruler moving? And this is a consultancy whose pitch IS that AI doubles output. Their incentive and their headline point the same way. The 5% production-agent number is more trustworthy precisely because it's embarrassing for the whole industry, not flattering to Tenex. Trust the ugly number. Hold the pretty one at arm's length.

The Enterprise Buyer

The adoption ladder is the useful part. 90% have an LLM rolled out, 50% have coding agents, under 5% have a non-engineering agent in production. The two blockers Lieberman names explain the whole drop-off. First, you can't automate a process that doesn't exist, and most back-office work lives in people's heads, not in documented steps. Second, nobody has clean control over which employee can see which data through an LLM. That second one is a procurement wall. No CISO signs off on an agent that can read across permission boundaries it doesn't understand. Tenex productized a gateway for exactly this because it kept hitting the same wall. That tells you where the real enterprise spend goes next: permissioning, not models.

The Compute Pragmatist

The money signal here is token cost collapse. Tenex runs 1-to-1 to 2-to-1 token spend against employee spend, high for anyone, and they're relaxed about it. Lieberman says it's much harder today to accidentally run up a ruinous bill than it was six to eight months ago. The Amp plus Claude subscription trick cut their token costs roughly 80% versus paying list API prices. Read that carefully. That 80% is a billing arbitrage, connecting a flat-rate subscription to an agent environment, not a drop in the underlying cost of compute. It works until Anthropic decides subscription-fed agent loops are eating their API margin and closes the door. Bank the savings, don't build your model on them.

The Builder

MetaHarness is the part you could copy on Tuesday. Agents keep canonical plan documents, log every decision in Markdown, check it all into git. The point is to get institutional knowledge out of human heads and into text an agent can load as context every time. This is a working pattern, not a Tenex invention: the teams getting real agent output are the ones who wrote their processes down first. Their own ladder proves it. The companies stuck at "we have a chatbot" are the ones whose work was never documented. If you want agents doing real work, the task isn't buying a better model. It's writing down how your team actually operates.

The Skeptic's second pass on vertical AI

Lieberman is bearish on vertical AI tools like Harvey for legal and OpenEvidence for medicine unless they own real proprietary data or a compliance moat. He may be right directionally and early on timing. A frontier model plus a thrown-together in-house agent rarely beats a tuned vertical product on day one inside a regulated workflow. The threat is real over three years. It is not a this-quarter rip-and-replace, and he knows it, which is why he's equally firm that nobody's ripping out Salesforce or SAP.

Where they genuinely disagree

The tension sits between the Compute Pragmatist and the Enterprise Buyer. If token costs keep falling and one model covers most tasks, the cheap answer says build everything in-house and skip the vendors. But the Buyer's two blockers, undocumented processes and data-permissioning chaos, don't get cheaper when tokens do. They get solved by consulting hours and governance software, which is exactly what Tenex sells. Cheap tokens make the model a commodity and make the plumbing around it the expensive part.

The second disagreement is the 2x itself. The Builder takes it as proof the pattern works. The Skeptic notes the yardstick shrank underneath the number. Both can be true: output rose and the measure drifted, which means the figure is somewhere below 2x and nobody outside Tenex can pin it.

What this actually hinges on

One belief: is the bottleneck on enterprise AI the model or the plumbing? Everything in this conversation says plumbing. Models are good enough that one of them, Claude Opus, covers most of the work. What stops agents reaching production is undocumented process and ungoverned data access. If that's right, the money over the next year flows to the permissioning and governance layer, and the multi-model routing startups that bet on orchestrating a hierarchy of models lose relevance, because operators are collapsing onto a single frontier model instead.

Before you act on any of it, check one thing in your own shop: how many of your core workflows are written down as steps an agent could follow? That number, not your model choice, predicts whether you'll ever get past the chatbot stage.

Prediction: Anthropic will change the terms of connecting a personal Claude subscription to third-party agent environments like Amp, or restrict subscription use in those tools, by the time it ships its next flagship Claude model after Opus 4.5, closing the roughly 80% cost gap against API pricing that Tenex relies on.

Confidence: Medium — the arbitrage directly undercuts Anthropic's own API margin.

Why: Tenex gets its ~80% saving by feeding agent loops through a flat-rate Claude Code subscription instead of paying metered API prices, and Lieberman says this is now cheap enough that running up a huge bill is hard. That same cheapness is the problem for Anthropic: a subscription priced for interactive coding is being used to power heavy automated agent runs that would cost far more on the API, so every team doing this is revenue Anthropic is leaving on the table. Companies close pricing gaps that eat their highest-margin product once the behavior scales, which is why the subscription-to-agent bridge is the less likely thing to survive untouched than to be capped, re-priced, or restricted. The opposite outcome, Anthropic leaving an open discount on its own API, only makes sense while the volume is small enough to ignore.

Revisit by 2027-04-25: We're right if Anthropic adds usage caps, rate limits, pricing changes, or terms-of-service restrictions on using a consumer Claude subscription inside third-party agent tools like Amp. We're wrong if the same subscription-to-Amp setup still delivers roughly the same discount versus API pricing with no new limits.