Hire an AI product manager in Portland.

Questions from Portland product teams.

What does a bias testing protocol look like in an AI product spec. What groups are tested and what thresholds apply?

A bias testing protocol in an AI product spec has three components. First, the group definitions: the spec names the specific demographic groups, user segments, or product categories the bias test covers. For a Nike digital product generating AI content about athletic performance, the groups include gender, age range, and sport category at minimum. Vague language like 'diverse users' is not a spec. Second, the test set design: the spec defines the test cases for each group, how they are generated (real user data, synthetic data, or a combination), and what sample size is required per group for the result to be statistically meaningful. A test set of 20 cases per group is not sufficient for a consumer product at Nike's scale. Third, the threshold: the spec states the maximum acceptable difference in outcome quality across groups. If AI-generated content scores 15 percentage points lower in quality rubric evaluations for one demographic group than another, the feature does not ship. The threshold is calibrated to the specific task and the brand risk. A consumer-facing AI feature at Columbia Sportswear carries different brand risk from a bias failure than an internal logistics tool.

How do you spec carbon-aware inference scheduling as a product requirement?

Carbon-aware inference scheduling is a product requirement when the feature has flexibility in when inference runs. It is not applicable to synchronous features where the user is waiting for a response. For async features (AI-generated content queued overnight, bulk recommendations processed in a batch job, background personalization updates), the spec defines three things. First, the carbon-aware scheduling signal: the spec names the data source the scheduling system uses to determine grid carbon intensity (Electricity Maps API is the standard). Second, the deferral window: the maximum time the job can be deferred to run during a lower-carbon grid period. A 6-hour deferral window is typical for non-urgent batch jobs. Third, the fallback: if the job has not run within the deferral window, it runs regardless of carbon intensity. The spec states the deferral window explicitly and confirms with engineering that the batch infrastructure supports time-shifted scheduling. Products at Vacasa or Columbia Sportswear that publish sustainability commitments can include carbon-aware inference as a verifiable practice, which requires a spec, not just an intention.

What does a public AI disclosure for a consumer-facing AI feature need to say?

A public AI disclosure for a consumer-facing feature needs five elements. First, what the AI does: a plain-language description of what the AI feature produces (recommendations, generated text, image edits) and how it produces it (based on which inputs). The description is accurate and complete, not a marketing summary. Second, what the AI does not do: the scope limits. If the AI generates product descriptions but does not make pricing decisions, the disclosure says so. Third, how to override or dismiss: if the AI feature can be turned off or its output ignored, the disclosure tells the user how. Fourth, what data is used: if the AI feature personalizes based on user history or preferences, the disclosure says what data it uses. Fifth, what to do if the output is wrong: a contact path for the user to report an AI output that is incorrect or harmful. For Nike or Columbia Sportswear, the disclosure does not need to be lengthy. A 150-word disclosure at the first use of the feature, with a persistent link to a full AI practices page, satisfies most consumer AI transparency expectations.

How do you balance responsible AI requirements with shipping speed?

Responsible AI requirements add time when they are retrofitted. They add minimal time when they are built in from the start. The practical approach is to distinguish between requirements that must be in version one and requirements that can be in version two. In version one, the spec must include: user control over AI outputs (the user can dismiss or override), an accurate disclosure at first use, and a basic bias test on the initial feature dataset. These are not optional. In version one, the spec can defer: carbon-aware inference scheduling (applicable only to async features with a batch component), a full bias audit across all demographic groups (a targeted test on the highest-risk groups is sufficient for v1), and a published AI practices page (a brief in-product disclosure is sufficient initially). This distinction is explicit in the product spec. The responsible AI requirements are sequenced, not ignored. At a Portland product team where the brand risk of an AI failure is high, the v1 requirements are non-negotiable. Skipping them to ship a week faster creates a recovery cost that is larger than the time saved.