The Twenty Minute VCWhy Token Maxing is Failing Enterprise Startups | Legora CTO
CHAPTERS
AI tooling spend: opportunity cost beats token cost
The conversation opens with a framing question: how much of a developer’s salary is worth spending on AI tooling. Jacob argues the right lens is opportunity cost in a competitive market—small efficiency gains compound and can outweigh direct token costs.
Building an engineering org in 2026: constant change, fewer priors
Jacob describes why building teams and orgs “now” differs from even two years ago: productivity, team size assumptions, and processes are all in flux. He views his lack of big-company priors as an advantage—iterate on org design like product.
Code is cheap now—product and review become the bottlenecks
Jacob breaks software creation into product definition, code writing, and review/merge. With AI compressing the code-writing phase, bottlenecks shift to product work (figuring out what to build) and review (ensuring it’s correct, secure, and aligned).
Why AI code review isn’t working yet—and what ‘real’ review should be
Legora uses AI review bots, but Jacob calls the space rough and incomplete. He argues effective review should focus less on reading lines and more on architectural impact, security boundaries, and strategic trade-offs—areas where humans still matter.
The future engineer: systems designer + ‘agent meta-engineer’
Jacob predicts engineers move one abstraction layer up: designing systems and orchestrating AI to implement details. He also highlights a growing need for teams that make agents effective through guardrails, tooling, and feedback loops—analogous to DevEx for humans.
Legora’s reality: >50% of code is AI-generated, but security risk rises
Jacob shares that Claude/Cursor-generated output accounts for more than half of their code changes, outpacing any individual engineer. This increases velocity but heightens security concerns, requiring continued human review and better risk-scoring approaches.
Faster PM work and postmortems: AI accelerates prototyping and incident response
AI changes day-to-day operations beyond coding: PMs can prototype earlier and validate with users before involving engineering, while incident agents can analyze logs/telemetry and draft postmortems rapidly. The result is front-loaded learning and faster operational cycles.
Design and ‘taste’ in an AI world: consistency over button debates
Jacob argues detailed design debates over feature-level UI may shrink because prototyping is cheap, but design language and consistency still matter. On “taste,” he believes differentiation comes from having an opinionated stance rather than converging into AI-generated sameness.
Copying is easy; enterprise edge cases are hard—and lawyers consume slower than you build
Jacob explains why fast copying doesn’t change Legora’s strategy: the last-mile enterprise requirements (RBAC, auditing, unhappy paths, scale quirks) are the real moat. He also notes a mismatch between AI/product speed and human adoption speed in legal, requiring careful rollout and education.
Internal vibe coding in enterprises: build vs buy and the rise of AI enablement teams
Jacob describes using “vibe coding” to create internal tools—from migration helpers to potentially HR/payroll components—because customization is often the real cost of SaaS. He proposes enterprises will need dedicated internal AI systems/enablement teams to rebuild workflows from first principles.
PM role and the new bottleneck: why product work stays central at Legora
While some argue PM and engineering converge, Jacob says in Legora’s context PM time is most valuable on customer discovery and synthesis, because product work is the bottleneck. PMs should prototype enough to reduce handover costs, but not spend large amounts of time engineering.
Models, latency vs quality, and open source: routing matters more than raw cost
Jacob explains Legora uses multiple models and re-evaluates frequently, optimizing for performance and latency rather than cost (for now). He’s bullish on open source for sovereignty and competition, but worries about duopolies and the lack of strong EU/US open-source alternatives.
Scaling engineering: DevEx as a force multiplier, hiring realities, and Europe vs US dynamics
Jacob reflects on under-hiring and under-investing in developer experience early, noting DevEx becomes more valuable as AI amplifies each engineer’s output. He compares hiring in Europe vs the US (risk tolerance, loyalty, equity understanding) and emphasizes low-ego cultures for integration and acqui-hires.
Token maxing is failing: measure outcomes, not usage—and IDEs will be reinvented
Jacob criticizes token leaderboards that incentivize waste (“token maxing”) and recommends demos/hack days that reward output and effectiveness. He also argues the current IDE form factor will die, replaced by interfaces focused on architecture and agent execution rather than reading code lines.
Big bets and the ‘crazy’ prediction: lawyers move one layer above contracts
In quick-fire and closing topics, Jacob stresses reinvention as the biggest threat, and offers a forward-looking view of legal work: like engineering, lawyers may operate one abstraction level up from drafting text—focusing on negotiation stance and acceptable risk while AI handles language details.
