The feature roadmap died. Probabilistic systems do not yield to quarterly Gantt charts. The 10x AI-native PM ships rubrics, evaluation frameworks, and FinOps models — not feature lists. They make accountability claims about systems they do not fully control. Below: open projects you can take on with us, the physics of what is changing, and roles we are hiring against.
Real, paid 1-3 week engagements with the Product.ai team. Each one is a problem we are working on at the frontier of product management — and the kind of work we hire against. Pick one that pulls at you and apply.
Last updated
In May 2026, product management at frontier-AI firms is undergoing the deepest paradigm shift the discipline has experienced since the 2010s shift from waterfall to agile. The legacy PM tooling — feature backlogs, sprint planning, A/B testing, OKRs tied to ship dates — assumed deterministic systems where output mapped reliably to input. AI-native products are probabilistic by construction. The same prompt can produce different answers across sessions; the same prompt can produce systematically worse answers as user behavior shifts the distribution. <strong>The 10x AI-native PM treats product management as evaluation architecture, FinOps, and risk posture — not feature scoping.</strong> Senior practitioners do not write user stories; they write rubrics.
Time-bound feature roadmaps are structurally incompatible with probabilistic systems. A roadmap commits to "shipping the X feature in Q3"; a probabilistic system's "X feature" might score 8/10 on the eval one week and 6/10 the next without any code change — because user behavior shifted, because the upstream model was updated, because the retrieval index drifted. The roadmap pretends the system is a thing you build; the system is a thing you continuously verify. Frontier-firm PMs (Anthropic, OpenAI, Cursor, Linear) all converge on a different artifact: the rubric portfolio.
The rubric portfolio is a versioned set of evaluation specifications. Each rubric defines what "good" means for a specific user journey or behavior. The PM's primary work is creating, refining, and weighting rubrics — and explaining to engineering and design what tradeoffs are acceptable. Shipping happens continuously against the rubric, not on a date. The 10x PM is the one whose rubrics teach the rest of the organization what the product is actually trying to do.
Evaluation architecture — the design of rubrics, judges, calibration sets, and decision gates — has become the load-bearing PM craft. It draws from measurement theory (what makes a metric valid?), epistemology (what counts as evidence?), and adversarial reasoning (how does this rubric fail?). The PM who can describe "why my rubric for chat-answer-quality has a 17-point Likert checklist instead of a 5-point scale, and what failure mode I added the eighth checklist item to detect" is operating at the new frontier. The PM who still thinks in user stories is doing 2018 work.
Anthropic's public stance on evals, OpenAI's evaluation framework, and Cursor's eval-driven release cadence all converge on the same physics: the team with the better rubric ships the better product. This is why senior AI-native PMs spend disproportionate time with the LLM-judge designers, the labeling team, and the calibration engineers — the rubric is upstream of every shipped behavior. The PM who delegates eval design to engineering has lost the seat.
AI products have variable, non-trivial, distribution-sensitive cost-per-query. A PM who specs a feature without modeling cost-per-interaction is shipping a financial mystery. The new craft requires fluency in tokenomics, model-routing economics, caching strategy, and the inference-cost-vs-quality tradeoff curve. Cursor's 2026 pricing-model evolution, Anthropic's prompt-caching pricing, and Replit's agent-vs-completion model pricing all reflect the same physics: cost is a product surface, not a finance back-office concern.
The 10x AI-native PM has internalized this. They will trade 200ms of latency for $0.02 of cost savings on a high-volume feature, or accept a 30% cost increase to escape a quality cliff that hurts retention. They build cost-quality dashboards alongside the eval dashboards. They argue with engineering about whether to use Haiku, Sonnet, or Opus on a per-route basis. The PM who treats inference cost as someone else's problem is producing financial liability the company will discover later.
Probabilistic systems require a new risk posture. The PM cannot guarantee output behavior the way a 2018-era PM could guarantee a button works when clicked. They can guarantee bounded behavior — distributions of outcomes, calibrated failure modes, escalation paths when the model disagrees with the user. They cannot guarantee deterministic correctness on every interaction. This is "accountability without total control," and it requires articulating risk in probabilistic language: "in the 95th percentile of inputs, this system produces X; in the long tail, it produces Y; here is what we do when Y."
The senior AI-native PM spends meaningful time on adversarial review, red-teaming, and brand-safety calibration. They write the rubric for "what the system must never say" with the same care as the rubric for "what the system should usually say." They distinguish capability (can the model do X?) from policy (should it?), and they own the second axis as their accountability surface. PMs who can't make this distinction become liabilities the moment the model exhibits the long tail in front of a customer.
The Engineering-Product-Design triad has not disappeared, but the power dynamics have shifted. The PM still owns the rubric and the risk posture. Engineering owns the model surface, the eval infrastructure, and the deterministic gates. Design owns the calibrated trust UI — how the system communicates its own uncertainty. The 10x AI-native PM operates with engineering and design as peers, not subordinates; they work shoulder-to-shoulder on the rubric and the eval, with design contributing the user-facing trust calibration in real time.
The strongest filter in 2026: a candidate who can show you a working rubric portfolio they shipped, with eval artifacts and FinOps modeling alongside. A polished PRD beats a polished resume; a working eval framework beats a polished PRD. The AI-native PMs who win are the ones whose rubrics other companies copy. The ones who lose are still authoring user stories in a tool that pretends features ship on dates.
Most of our best people came through projects, not interviews. If a project pulls at you and the trial goes well, the role conversation follows.
Twelve-minute Hireflix video, async. Then a 30-60 minute chemistry call. Then a paid 1-3 week project alongside the team. We will know within a week whether to move forward.