Fractional AI Product Manager · Seattle, WA

Hire an AI product manager in Seattle.

Enterprise AI features in the Amazon ISV ecosystem and Microsoft partner software require a different kind of product spec. The functional requirements are the easy part. The compliance spec, audit log, RBAC mapping, and data retention policy are where enterprise deals get blocked.

Boeing technology vendors, Amazon Web Services ISVs, and Microsoft partners building AI features need a PM who writes the compliance requirements alongside the functional ones. Those are not two separate documents. They are one spec.

Fixed engagement, fixed scope. Scope is agreed before work starts, so there are no surprise invoices.

How we work

Tell us about your enterprise AI product challenge.

What enterprise AI PM requires beyond consumer PM.

Consumer AI product management focuses on engagement metrics, model performance benchmarks, and A/B test outcomes. Those matter for enterprise too. But enterprise adds three requirements that consumer products rarely need.

Audit logging is a PM responsibility at enterprise scale. The spec must define what events are written, what each log entry contains, and how long logs are retained. If a procurement team at an Amazon ISV partner asks what audit trail exists for AI feature outputs, the PM must have a documented answer before the deal closes.

Data residency is a PM decision, not just an infrastructure decision. When an AI feature sends user data to a third-party model provider, the product spec must state where that data is processed, whether it crosses a geographic boundary, and what the retention policy is for model inputs. Enterprise customers ask these questions before signing.

RBAC for AI features is more granular than RBAC for standard features. The spec must define which permission level can trigger the AI feature, which can view AI outputs, and which can disable it. Skipping this produces a support ticket six months after launch.

Compliance specs and rollout protocol.

Enterprise AI feature rollout is not a consumer percentage rollout. It requires a different protocol.

Customer advisory group, not a percentage rollout

Rolling out to 5% of enterprise accounts is wrong. You need three to five named enterprise accounts that agreed to beta participation, documented acceptance criteria they will evaluate, and a formal go-live checklist they sign before you proceed.

Acceptance criteria before beta

The advisory group receives the acceptance criteria before the beta starts, not after. Criteria defined after testing has begun are not acceptance criteria. They are rationalization.

Compliance documentation in the spec

Audit log spec, RBAC mapping, data retention policy, and model versioning policy are sections in the product spec, not separate compliance documents produced after the feature ships.

Go-live checklist with sign-offs

The go-live checklist names the people who must sign before general availability: product, engineering, security, and at least one advisory group member. No unnamed approvers.

SLAs for enterprise AI features.

Committing to 99.9% uptime for an AI feature that calls an external LLM API is a liability. OpenAI and Anthropic have both experienced incidents exceeding 30 minutes. The product spec must define what happens when the model provider is down: does the feature degrade gracefully, fail closed, or fall back to a cached result?

The SLA should reflect the fallback behavior. If you have a documented fallback, 99.5% is defensible. Without one, commit to less and document why.

Latency SLAs for enterprise AI features need two numbers: p95 and p99. A p95 under 3 seconds and p99 under 8 seconds is a reasonable starting point for synchronous features. Async features should define a completion time SLA rather than a latency SLA.

The upstream dependency carve-out belongs in the product spec, not just in the enterprise contract. If an LLM provider outage causes your AI feature to go down, the spec should define the notification obligation, the root cause analysis timeline, and the remediation procedure.

How we work.

Fixed scope before any work starts

Discovery call, written scope, fixed price. You approve the scope before we start. Scope changes require a change order. No surprise invoices.

Remote-first, async by default

Shared Slack channel, written updates every Friday, and recorded Loom walkthroughs for every milestone. Works well for Seattle and distributed enterprise teams.

Start in two weeks, not six months

No recruiting cycle, no onboarding ramp. We typically start within two weeks of a signed agreement. The engagement ends with artifacts your team owns outright.

Questions from Seattle enterprise product teams.

What should an enterprise AI feature compliance spec include that a regular feature spec does not?: A compliance spec for an enterprise AI feature adds five sections that a standard feature spec omits. First, an audit log specification: what events are written, what data each log entry contains, and the retention period. Second, RBAC mapping: which permission levels can trigger the AI feature, which can view AI outputs, and which can override or disable it. Third, a data residency section: where AI inputs and outputs are stored, whether they leave a defined geographic boundary, and what the data retention policy is for model inputs. Fourth, an incident classification: what constitutes a material failure, what the notification obligation is to the enterprise customer, and what the remediation timeline is. Fifth, a model versioning policy: what happens to the enterprise customer's workflow when the underlying model version changes.
How do you define the customer advisory group for an AI feature rollout, who should be in it?: An enterprise AI feature customer advisory group should include three to five enterprise customers selected for specific criteria. At least one should be in a regulated industry where AI feature failures have compliance consequences. At least one should have a large user base so that rollout performance at scale is visible before general availability. At least one should have an internal AI governance team that will review your documentation. The group participates in a structured beta: they receive the acceptance criteria before the beta starts, document any failures they observe, and sign off on the go-live checklist before you proceed. Advisory groups built after the feature is already in beta provide feedback too late to change the architecture.
What uptime and latency SLA is reasonable to commit to for an AI feature that calls a third-party LLM API?: For an enterprise AI feature calling a third-party LLM API, a 99.5% monthly uptime SLA is defensible if you implement a fallback path for model provider outages. Committing to 99.9% without a fallback is a liability, because major LLM providers including OpenAI and Anthropic have experienced incidents that exceeded 30 minutes of degraded service. For latency, p95 under 3 seconds and p99 under 8 seconds are reasonable starting points for synchronous AI features. Async AI features, where the output is delivered after a short delay rather than inline, can commit to a completion time SLA instead of a latency SLA. The product spec should define both the happy-path SLA and the fallback behavior, not just the happy-path number.
What happens contractually when your AI feature goes down because your LLM provider had an outage?: Most enterprise SLAs distinguish between service outages caused by the vendor and those caused by upstream provider failures. The standard approach is to include an upstream dependency carve-out in your SLA: if an incident is traceable to a published outage by a named third-party model provider, the downtime counts toward the provider's SLA measurement but not yours. You still owe a root cause analysis and a remediation plan. The product spec should define the carve-out language, the list of named upstream providers it covers, and the notification obligation to the enterprise customer when an upstream incident begins. Leaving this out of the spec means your legal team negotiates it separately for each enterprise contract.

Scope your enterprise AI feature spec.

Tell us which AI feature you are trying to ship and what the enterprise customer requirements look like. We reply within one business day.

Hire an AI product manager in Seattle.

Fixed engagement, fixed scope. Scope is agreed before work starts, so there are no surprise invoices.

What enterprise AI PM requires beyond consumer PM.

Compliance specs and rollout protocol.

Enterprise AI feature rollout is not a consumer percentage rollout. It requires a different protocol.

Customer advisory group, not a percentage rollout

Acceptance criteria before beta

The advisory group receives the acceptance criteria before the beta starts, not after. Criteria defined after testing has begun are not acceptance criteria. They are rationalization.

Compliance documentation in the spec

Audit log spec, RBAC mapping, data retention policy, and model versioning policy are sections in the product spec, not separate compliance documents produced after the feature ships.

Go-live checklist with sign-offs

The go-live checklist names the people who must sign before general availability: product, engineering, security, and at least one advisory group member. No unnamed approvers.

SLAs for enterprise AI features.

The SLA should reflect the fallback behavior. If you have a documented fallback, 99.5% is defensible. Without one, commit to less and document why.

How we work.

Fixed scope before any work starts

Discovery call, written scope, fixed price. You approve the scope before we start. Scope changes require a change order. No surprise invoices.

Remote-first, async by default

Shared Slack channel, written updates every Friday, and recorded Loom walkthroughs for every milestone. Works well for Seattle and distributed enterprise teams.

Start in two weeks, not six months

No recruiting cycle, no onboarding ramp. We typically start within two weeks of a signed agreement. The engagement ends with artifacts your team owns outright.

Questions from Seattle enterprise product teams.

What should an enterprise AI feature compliance spec include that a regular feature spec does not?

A compliance spec for an enterprise AI feature adds five sections that a standard feature spec omits. First, an audit log specification: what events are written, what data each log entry contains, and the retention period. Second, RBAC mapping: which permission levels can trigger the AI feature, which can view AI outputs, and which can override or disable it. Third, a data residency section: where AI inputs and outputs are stored, whether they leave a defined geographic boundary, and what the data retention policy is for model inputs. Fourth, an incident classification: what constitutes a material failure, what the notification obligation is to the enterprise customer, and what the remediation timeline is. Fifth, a model versioning policy: what happens to the enterprise customer's workflow when the underlying model version changes.

How do you define the customer advisory group for an AI feature rollout, who should be in it?

An enterprise AI feature customer advisory group should include three to five enterprise customers selected for specific criteria. At least one should be in a regulated industry where AI feature failures have compliance consequences. At least one should have a large user base so that rollout performance at scale is visible before general availability. At least one should have an internal AI governance team that will review your documentation. The group participates in a structured beta: they receive the acceptance criteria before the beta starts, document any failures they observe, and sign off on the go-live checklist before you proceed. Advisory groups built after the feature is already in beta provide feedback too late to change the architecture.

What uptime and latency SLA is reasonable to commit to for an AI feature that calls a third-party LLM API?

For an enterprise AI feature calling a third-party LLM API, a 99.5% monthly uptime SLA is defensible if you implement a fallback path for model provider outages. Committing to 99.9% without a fallback is a liability, because major LLM providers including OpenAI and Anthropic have experienced incidents that exceeded 30 minutes of degraded service. For latency, p95 under 3 seconds and p99 under 8 seconds are reasonable starting points for synchronous AI features. Async AI features, where the output is delivered after a short delay rather than inline, can commit to a completion time SLA instead of a latency SLA. The product spec should define both the happy-path SLA and the fallback behavior, not just the happy-path number.

What happens contractually when your AI feature goes down because your LLM provider had an outage?

Most enterprise SLAs distinguish between service outages caused by the vendor and those caused by upstream provider failures. The standard approach is to include an upstream dependency carve-out in your SLA: if an incident is traceable to a published outage by a named third-party model provider, the downtime counts toward the provider's SLA measurement but not yours. You still owe a root cause analysis and a remediation plan. The product spec should define the carve-out language, the list of named upstream providers it covers, and the notification obligation to the enterprise customer when an upstream incident begins. Leaving this out of the spec means your legal team negotiates it separately for each enterprise contract.

Scope your enterprise AI feature spec.

Tell us which AI feature you are trying to ship and what the enterprise customer requirements look like. We reply within one business day.