Lead Product Manager, Agentic AI

About The Role

We're not looking for a traditional PM who writes Jira tickets and hands them off, and we don't need a backend engineer who wants to stay hidden behind the API. We're standing up a small, highly autonomous pod to build a net-new agentic AI product inside our Tax & Accounting division, and we need someone who lives at the intersection of product strategy, tax domain logic, and AI capability.

You'll be the center of gravity for AI execution in the pod: embedded with domain experts to surface complex tax logic, and using agentic tooling to prototype workflows that engineering will harden into production.

One thing to be explicit about up front: WK T&A already ships mature tax calculation engines that encode a lot of the deterministic logic in this domain.
Your job is to orchestrate agents around that engine — not to reimplement it inside a prompt. Candidates who understand why that distinction matters will do well here.

• Prototypes and evals are the spec. Instead of long PRDs, you'll ship high-fidelity agent prototypes and the eval suites that define "done." Prototypes prove value; evals define correctness. Engineering scales what survives both.
• Own the eval framework. Tax software has to be right. You'll design prompts, define tool-use constraints, and build the eval harness (with engineering) that measures accuracy, catches hallucinations, and traps logic failures before they reach a client return.
• Bridge the translation gap. Decompose ambiguous tax jobs-to-be-done into structured agent architectures — multi-agent orchestration, state management, human-in-the-loop decision points — that engineering can build against.
• Own the unit economics. Balance model accuracy against latency and inference cost. A 20-step multi-agent loop might solve the problem; you decide whether it can ship at a commercially viable gross margin, and design cheaper paths when it can't.
• Design for graceful degradation. Enterprise users have zero tolerance for infinite spinners or confidently wrong answers. You'll design explicit fallbacks: when the agent hits ambiguity, the system steps down to deterministic rules from the calc engine, or routes to a human — visibly and predictably.
• Partner with tax SMEs. You are not expected to be a tax expert. You'll work side-by-side with our internal domain experts, translating their judgment into system prompts, tool contracts, and eval criteria. Building trust with SMEs is part of the job; they are collaborators, not a resource to be mined.

Ultimately, you own the journey from opportunity identification through validated prototype, evaluation performance, and production launch. Success is measured by customer adoption, task completion accuracy, and business impact, not document output.

This is a small pod, but it isn't a solo role. You'll partner with:

• Engineering
• UX Design partner
• Tax domain SMEs

We call this out because agentic AI in a regulated domain isn't a one-person job. If any of those capabilities are missing, identifying the gap and helping build the right team around the product is part of the role.

• AI power user. You don't need to write production Python from scratch, but you are a daily practitioner of agentic tooling. You've orchestrated real work using agentic development environments such as Claude Code, Cursor, Copilot, or similar tools.
• Transcript debugger. When an agent gets stuck or hallucinates, your first move is to open the raw transcript. You can tell whether the failure was a bad prompt, a missing tool, an eval gap, or a genuine model limitation — and you know which of those to fix first.
• Systems thinker. You understand context windows, RAG, model routing, tool design, and — crucially — when a deterministic rule or a call into an existing engine is the right answer instead of an LLM call.
• Comfortable in the gray. Traditional software is deterministic; agents aren't. You have a track record of shipping UX that handles AI uncertainty honestly — fallbacks, confidence signals, human-in-the-loop validation — without hiding the seams from users who need to trust the output.
• Commercial pragmatism. You…