Test Predictions Against a Real Baseline.

Use an agreed simple baseline and leakage-safe time or group splits. Hold out representative data; compare decision costs and task-specific metrics rather than relying on a single accuracy percentage.

Match metrics to the task: forecast error and service costs, precision/recall for rare events, calibration for risk scores or ranking quality for recommendations. Agree thresholds with the decision owner before release.

Scope data-quality findings, a baseline, reproducible training/evaluation code and a feasibility report. Production serving, automated decisions, new labeling programs and recurring retraining are separate deliverables unless agreed.

Name an owner for data checks, live performance review and model version rollback. Agree retraining triggers, access to fresh labels, acceptance tests and support terms; retraining is not an automatic outcome guarantee.

Budget depends on ML scope, source readiness, integrations, evaluation effort, privacy controls and support. Agree deliverables and exclusions in a written quote; provider, hosting, license and ongoing operations costs are separate line items.

Start after access, sample data and reviewers are available. Agree discovery, pilot evaluation, acceptance and release gates. Dates depend on prerequisites and findings; no universal launch deadline is promised.

Agree repository access, intellectual-property terms, third-party licenses and account ownership in the contract. Handover the agreed source, configuration, evaluation evidence and operating notes; support and future changes need explicit terms.

Map authorized data flows, identity/access controls, retention, logging and deletion before implementation. Review selected processors and configurations with the data owner; the service does not itself establish regulatory approval or certification.

Use the contact page with a non-sensitive workflow summary, current tools, constraints and a decision owner. Review related service pages through Services; provide confidential samples only through an agreed channel.