Paying for Outcomes, Building with Agents, and the Last Mile of Enterprise AI Adoption
Last year, Alex Lieberman and his co-founders came on our show to explain Tenex’s unconventional bet: paying AI engineers for the outcomes they deliver rather than the hours they spend at the keyboard. If AI makes engineers more productive, billing by the hour actually penalizes them for doing the same job faster. Tenex wanted the incentives to work in reverse.
At AIE New York, Alex returned alongside CJ Hess to discuss what happens when that philosophy scales across an entire engineering organization. We also brought one of our own experiments to the table: Kill My SaaS, where we challenged developers to replace a conference speaker management tool that was once quoted at $43,000 per year.
Both stories ask the same question: when the cost of building plummets, how do you measure the value of a great outcome?
How to Pay for Output When the Job Keeps Shifting
Tenex still uses story points as a major component of engineer compensation. Alex noted that the failure modes people worry about—artificially inflating estimates or shipping shoddy code to game points—aren’t the primary problem. Tenex’s AI strategists are held accountable for client retention and business growth, giving them an incentive to push back against work that doesn’t actually benefit the client.
The harder challenge is how to reward the work that doesn’t fit neatly into a ticket: understanding a client’s problem, designing the architecture, or deciding what to build in the first place. CJ explained that as models get better, he spends less time writing code directly and more time translating customer context into a format the model can work with. A compensation system built around “delivering tasks” must carve out room for that kind of work.
Alex shared that Tenex has grown to 75 people, with a client base split roughly evenly between Fortune 2000 enterprises and mid-market companies. The company started out as an AI-powered software development agency; today, an increasing number of clients are turning to them for company-wide AI transformation support. Jump to the compensation discussion →
MetaHarness: Keeping the Entire Project in the Repo
CJ broke down Tenex’s MetaHarness: a shared set of skills, project documentation, and a CLI wrapped around the development process. Planning and decisions are committed to Git alongside the code. When an agent picks up a task again, it can read not just what was built, but why the team decided to build it that way in the first place.
Tenex’s own article on the AI-native SDLC illustrates this concept as a shared project record for both humans and agents, integrating views across Linear, GitHub, and client updates:
Source: Tenex, Building an AI-Native SDLC.
A valuable distinction here is between using agents simply to write code faster versus making the scaffolding around the code legible—including planning, acceptance criteria, decisions, and proof of work. Tenex’s post also acknowledges the trade-off: for small teams managing a single project, this level of structure can easily feel like over-engineering. CJ explains the workflow →
Kill My SaaS: The Cost of Owning What You Build
In our Kill My SaaS episode, we evaluated 72 submissions designed to replace our speaker management software. The team that won in the eyes of our human judges ranked only 19th on our AI judge leaderboard. It served as a stark reminder: verifying whether software satisfies a checklist of use cases is fundamentally different from judging whether people will actually want to use it.
I brought up this project to Alex as a real-world case study for why a business might choose to build its own software. His counterargument: the higher upmarket a customer sits, the higher the cost and risk of ripping out and replacing an off-the-shelf system. And even with a smaller tool, someone still has to maintain it. Do the cost savings truly justify the ongoing effort?
My response was that the return on building in-house goes beyond the bill. Our team had endured a mediocre product for far too long. By owning the workflow, we could deliver a vastly better experience for speakers and operators, iterating in minutes rather than waiting months on a vendor. That upside has to be weighed against the maintenance overhead Alex highlighted. Watch that segment →
The Case For and Against Vertical AI
Alex is broadly bearish on vertical AI products: if every enterprise expects a unique, bespoke solution, how do you maintain software-like margins? And as frontier models keep improving, how much of the underlying work will simply be absorbed by general-purpose tools?
I pushed back with examples like Harvey and OpenEvidence. Even if the underlying models are accessible to everyone, proprietary domain-specific data, tailored user interfaces, system integrations, and established user habits still matter tremendously.