AI Engineer Code 2025

Paying Engineers like Salespeople

Loading slides…

Read the talk

Paying Engineers Like Salespeople

Tenex ties engineer compensation to accepted story points, pairing a direct reward for faster delivery with architecture work, client approval and incentives for customer retention.

From a talk by Arman Hezarkhani

What happens when story points determine pay?

Tell a room of experienced engineers that you pay people by completed story points, and the first reaction is skepticism. That was Arman Hezarkhani’s experience in the green room: at Tenex, the AI transformation company he co-founded, clients pay for delivered story points, and engineers earn compensation based on the points they complete. The same unit connects what the company sells with what it rewards internally.

Slide lists paid on output, uncapped upside, and incentives to work smarter, faster, harder, beside a statue using a phone.Open full source frame
Tenex’s premise: pay software engineers like salespeople.

Hezarkhani credits this arrangement with helping Tenex recruit engineers who have started and exited companies, machine learning and AI researchers, and former NASA rocket scientists. He also reports fast delivery of maintainable, high-quality code. These are practitioner claims, not a controlled comparison of compensation systems. The question is what compensation could look like as AI changes engineering productivity—not whether every team should adopt Tenex’s model.

0:340:45
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:34 · section reference included

Two engineers, different reasons to experiment

Hezarkhani came to this question through his own engineering career: studying and teaching computer science at Carnegie Mellon, helping Google scale its AI, cloud and mobile practices internationally, and then founding venture-backed startups. While working on his last startup at the WeWork at 33 Irving in New York, he noticed a neighbor with Visual Studio Code open. There was no chat window. The neighbor was typing directly into the editor—character by character, beginning with for.

Illustration of a crouching caveman holding a club beside a glowing laptop displaying AI, with the presenter at left.Open full source frame
A caveman encounters a laptop marked “AI.”

He contrasts that with a comic picture of his own machine: dozens of agents, including several ordering lunch, others writing code and another doing research. Beneath the joke is a real question: why would two engineers sitting beside each other use such different workflows? Some people try AI and find it worse than their own work; others do not want to try it. Hezarkhani’s explanation centers on incentives. As a founder, he personally benefited from every incremental improvement in efficiency, so searching for better tools was part of the job.

That incentive sent him to Twitter, LinkedIn and blog posts to find ways to produce better code faster. He imagined that his neighbor probably received a base salary, an annual bonus and equity. The neighbor’s actual compensation was unknown; this was a hypothesis about why conventional rewards might not encourage the same investment in experimentation. The person who captures an efficiency gain has a stronger reason to seek it out.

2:192:30
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

2:19 · section reference included

Who keeps the benefit of finishing faster?

An intentionally invented history of compensation makes the tradeoffs visible. Start with hourly labor: the buyer pays for time, and the engineer receives a predictable rate. When Hezarkhani asks the audience what is broken about that arrangement, the answers converge on speed, output and the absence of upside. Finishing sooner can actually reduce the engineer’s earnings.

Now turn the estimate into a fixed price. Five hours at $100 an hour becomes a $500 project. Finish faster and the engineer keeps the $500; take longer and the engineer absorbs the overrun. That creates an incentive to improve efficiency, but it also makes estimation risk consequential. An engineer unsure of a five-hour estimate may quote fifteen or twenty hours to protect against the downside. The buyer then has a new reason to dislike the arrangement.

Salary and bonus remove the need to price every project, but in Hezarkhani’s illustration they can leave little marginal reward for extra effort. His fictional Larry Page vignette introduces the next step: share ownership so employees benefit when the company succeeds.

ArrangementWhat the engineer is rewarded forIncentive tension
HourlyTime spentFaster completion can reduce pay
Fixed projectCompleting an agreed scopeOverrun risk encourages padded estimates
Salary and bonusEmployment and bonus criteriaExtra effort may have weak marginal upside
Salary, bonus and equityWork plus company successOwnership connects effort to an eventual exit

Equity is the strongest alignment mechanism in this progression: employees participate in the upside that motivates the founder. Hezarkhani recognizes it as a foundation of the startup model, rather than dismissing it as ineffective.

The speaker stands beside a timeline titled “A Brief History of Compensation,” showing hourly labor, project-based work, salary, salary plus bonus, and salary plus bonus plus equity.Open full source frame
A brief history of compensation, from hourly labor to salary, bonus, and equity.

The difficulty is that most startups do not become Google. Engineers know that equity may never produce a payout. Some people willing to take substantial risk start companies themselves; others negotiate firmly for cash and treat equity as an optional upside. Hezarkhani’s proposed next step is a more direct reward for using AI tools well, with high code quality remaining a requirement.

4:535:08
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:53 · section reference included

From a client request to an accepted ticket

Tenex separates its work into roadmapping and execution. A client may arrive with a broad request for AI or a specific goal, such as increasing customer-service output. Tenex studies the business and produces a roadmap that the client can implement independently or ask Tenex to execute. Execution can involve off-the-shelf tools or custom builds; the story-point compensation model enters with the custom work. Clients can also submit new requests after seeing the roadmap.

Two client-facing roles turn those requests into delivered software:

  1. The strategist distills the requirements. Mostly technical people, including former product managers and engineers, work with the client to define the product requirements.
  2. The AI engineer designs the implementation. The engineer creates an architecture design document containing tickets, each assigned story points. Hezarkhani says this design work consumes most of Tenex’s engineering time.
  3. The engineer implements the tickets. Code follows the architecture and the agreed decomposition of the work.

The underlying delivery process is familiar. The unusual part is how completed work becomes compensation.

Ticket acceptance is the trigger for earning a fee per completed story point. Engineers receive a flat base, and Tenex performs a quarterly true-up based on the story points completed. The model therefore retains a base payment while connecting variable compensation to accepted output. Hezarkhani presents the following projects as examples of work delivered through this process.

Three stages connected by arrows: product requirements, architecture design, and writing code. Strategist is labeled beneath requirements; AI engineer spans design and code.Open full source frame
Product requirements lead to architecture design and writing code.
7:357:44
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:35 · section reference included

Billboard moderation has two costs

One client sells billboard inventory, including placements in Times Square. Customers can buy through a traditional sales process or use a self-service product: upload a PNG, select a location and choose how long the advertisement should run. The experience resembles buying a Facebook or Google ad, but the uploaded image must pass two rounds of moderation—one inside the company and another with the billboard owner.

The internal review creates both an operating cost and a revenue delay. People must be paid to moderate the images, and Hezarkhani describes the process as slow and error-prone. While an advertisement waits for approval, it is not running and generating revenue. That makes moderation a concrete AI opportunity: reduce the work and waiting time attached to an existing transaction.

Tenex scoped the moderation system, wrote the architecture design document, divided the work into tickets and built it. Hezarkhani reports a two-week build and 96% accuracy compared with the human moderator. The talk does not supply the sample size or evaluation protocol, so the result should be read as a reported human-comparison result, not a general production accuracy guarantee or a measured revenue improvement.

9:319:34
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:31 · section reference included

Expanding what a retail camera can do

A second client supplies devices to retailers around the world. Its low-power devices could run one AI model on the device, producing traffic heat maps from camera footage. A camera looking down into a space could show where people moved throughout the day—useful information for a retailer, but only one application of the installed camera and compute.

Tenex explored what more capable chips and model quantization could enable. The proposed combination was additional processing capacity plus models made small enough to support parallel execution. After producing an opportunity report, Tenex built an expanded model suite. Hezarkhani reports building five models that can run in parallel, including heat mapping, queue detection and theft detection. He supplies no device specifications, quantization settings or runtime benchmarks.

The workflow remained the same: product requirements, architecture, implementation and payment based on output. The deliverable was an expanded set of capabilities for the device; the talk does not describe rollout status.

10:5211:05
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

10:52 · section reference included

What keeps the incentive from being gamed?

Paying by story points creates obvious failure modes:

  • Point inflation: An engineer could assign forty-five story points to adding a button.
  • Rushed delivery: Finishing quickly might mean producing something that fails to work or is difficult to maintain.
  • Unhealthy competition: Individual output incentives could encourage engineers to compete for credit rather than collaborate.

These risks are part of the model’s design problem. A short build time alone does not answer whether the result is good.

Tenex’s first counterweight is the strategist. Strategists are compensated on net revenue retention, or NRR. Hezarkhani calls this customer happiness, but NRR measures retained and expanded revenue from existing customers, including revenue losses; it is not a direct satisfaction score. The intended balance is between engineers rewarded for accepted output and strategists rewarded for a continuing customer relationship. Each ticket must pass multiple internal QA rounds involving the strategist and receive client approval.

Point inflation has a related control: strategists scope the work, and the scope is reviewed. Engineers still author the architecture and its tickets, but they do not have unchecked authority to determine the scope on which their compensation depends. Acceptance gates address quality; strategist scoping and review address inflated estimates.

12:0012:14
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:00 · section reference included

AI amplifies the people the system rewards

For the collaboration problem, Hezarkhani’s answer begins with hiring. Tenex deliberately makes hiring difficult so subsequent work is easier. In this model, selecting people who can exercise judgment and work well together is an essential condition, not something the compensation formula can replace.

His co-founder Alex describes AI as a distorting mirror that enlarges existing attributes. A strong engineer becomes more capable; a sloppy engineer can produce more sloppiness. This is an amplification metaphor, not a measured productivity multiplier. It explains why rewarding output must coexist with judgment about who produces it and how they work.

Hezarkhani’s closing challenge extends beyond engineering: if AI makes people faster and more capable, does their compensation give them a reason to develop and use that capability? His belief is that current arrangements can hold people back. The invitation is to reconsider how the people doing the work share in the gains they create.

13:2413:34
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

13:24 · section reference included

Resources

From the talk

  • TenexDocumentation0:34

    Tenex's overview of its AI strategy, transformation and outcome-based engineering services.

Transcript download
Full transcript

Loading transcript…