Paying engineers for output exposes AI's real bottleneck: process

The model. Tenex, an applied AI consultancy, pays engineers by story points rather than hours, and reports a roughly 4x spread between its highest and lowest paid engineers. The bet is that AI widens output variance, so compensation should follow output. The two classic failure modes, sandbagging story points and shipping bad code, have not shown up. The harder problem is paying for non-coding time such as architecture and client meetings, which is forcing the model to evolve.

The evidence. The firm claims per-engineer output roughly doubled over the past year, with models like Opus 5.5 doing more long-running work. Story-point criteria are recalibrated quarterly to keep pace with model improvements. Tenex also describes a "meta harness": structured documentation, a CLI, and agent decision logs checked into the repo, so context never lives only in someone's head. These are useful operating details, but all self-reported; there is no external validation or quality measure.

The adoption gap. From client work, Tenex estimates 90% of enterprises have rolled out a general LLM, 60% have training and enablement, only 50% have coding agents in engineering, and just 5% have a non-engineering agent in production. That drop is the real story. It is not model quality that holds enterprises back; it is workflow standardization, data permissioning, and governance. CIOs and CISOs are struggling to control what agents can read across the company.

The process lesson. CJ Hess says you cannot automate a process that does not exist. Teams with 40 people doing similar work slightly differently cannot build reliable background agents. Standardization has to come first, then evals and handoff. The tradeoff is that standardizing too early can lock in weak processes. Tenex's advice is to build process before automation, which is easy to agree with and hard to do in large organizations.

The SaaS debate. The host's example of replacing a $40,000-a-year middle-tier SaaS for $1,000 in his competition shows the pull of one-of-one software. Alex pushes back: in large enterprises, the risk of ripping out core systems is too high, and the deciding factor is whether the prize is big enough to justify the cognitive load of maintaining non-core software. Tenex built its own OS and CRM because iteration speed matters there.

What's missing. There is no data on code quality, retention, or long-term outcomes under output-based pay. The 2x output claim could hide growing architecture debt. CJ hints that the model may move away from story points toward true output value as agents do more of the work. The signal to watch is whether Tenex can keep compensating fairly as models improve.