AI Product Manager

About This Role

K1x turns tax documents (K-1s, K-3s, state schedules, 990s, and a growing range of other filings and

statements) into filing-grade structured data for CPA firms, family offices, and institutional investors.

Accuracy and rework are the first thing our buyers ask about, and the models, prompts, and rules that

produce each extracted value are the product surface that answers them. That surface needs a

product manager fluent in how ML systems behave: how they are evaluated, where they fail, what

they cost, and when to replace a component rather than tune it. This role is the product manager for

the AI team: how its models and services are supplied to every K1x product line, partnering with the

Head of AI on delivery and with the other product managers on where those capabilities land.

What You'll Do

The AI roadmap. A rolling six-to-twelve-month roadmap for extraction coverage, accuracy, and the

platform underneath them, sized against real team capacity, sequenced with dependencies made

explicit, and re-planned when the evidence changes. You represent it in portfolio planning alongside

the other product roadmaps.

The demand queue. Intake and prioritization of new document types, forms, and fields requested by

Product, Tax Content, Sales, and Client Success. You rank the queue with a defensible framework

(RICE, WSJF, Kano, opportunity scoring, voice-of-customer synthesis), make sure each item arrives

with the definition and test data needed to validate it, and explain the ranking to those who did not get

their item first.

Accuracy as a product metric. Own how accuracy is defined and reported for each audience:

executive, customer, product, engineering. Translate model-level measures (per-field precision,

coverage, straight-through rate) into what a user experiences: what we missed, what they had to

touch, how many touches it took to reach a filing-ready result. Own the recurring accuracy report.

Product requirements for model and vendor decisions. Engineering evaluates and recommends

what serves each stage of the pipeline: a frontier model, one of our own models, or a vendor service.

You supply the product side of that decision: which accuracy, cost, and latency thresholds actually

matter to customers, the business case and budget for a change, and the acceptance criteria a

release must clear. You keep the decision record so the reasoning survives.

Capacity and the tax calendar. Filing peaks in September, October, and November drive usage; a

tax-year release lands every January. You plan engineering reserve around the peaks, own scope and

dates for the tax-year release, set the defect-intake service level with Client Success and QA, and

keep proof-of-concept work time-boxed.

The correction loop. Users correct extraction output inside our products. You partner with those

product managers and UX so corrections become a usable signal with field-level provenance, and over time a confidence-driven review experience.

• 4+ years of product management on shipped B2B software, ideally fintech or regtech where a wrong number costs more than a slow one, with at least 2 years owning an AI/ML-powered product surface: document AI or extraction, search and ranking, LLM features, or an ML platform.
• Fluent in the mechanics: capacity planning against a real team, prioritization frameworks (RICE, WSJF, Kano, voice-of-customer synthesis), roadmap sizing, release management, and PRDs an engineer would actually read.
• Understand how ML products fail differently from software: precision and recall trade-offs, evaluation sets, drift, cost per inference, and why "the model got it wrong" is a product question first.
• Have made or shaped build, buy, or replace calls on model or vendor components, and can walk through one that went badly.
• Write clearly and run a tight meeting. Half this job is turning engineers' conviction into a decision document Product, Tax, and Finance can act on.
• Tax-domain knowledge is not required; Tax Content owns the rules and CPAs adjudicate. Curiosity is; the interesting failure modes live in the footnotes.