When the Output Is a Sent Email, Not a Draft

Discovery first, then a calibrated curriculum

Week one settled what would be taught. Working sessions covered current and intended Claude usage, target workflows, output schemas, owner approval points and known failure modes across the four capability areas, alongside a review of customer supplied non sensitive prompt samples, schemas and sample documents. Three priority workflows were confirmed in writing as the teaching cases, recorded with the roster, schedule and success measures in a calibration memo the customer accepted.

Ten hours, five sessions, exercises inside the session

Delivered remotely to a cohort of five to ten named participants from product, engineering and operations, with hands on work inside session time rather than scheduled as separate labs.

• Claude fundamentals for the product team: the boundary between deterministic rules, existing product logic and model judgment, and which parts of the four capability areas warrant a model at all.
• Prompt engineering for customer facing communication: holding one individual owner's tone rather than a generic professional register, follow up sequences, priority flagging, and preventing invented availability or unsupported commitments in mail sent under the owner's name.
• Document processing and structured outputs: extraction from invoices, receipts, contracts and forms across PDF, image, Word and spreadsheet inputs, schema conformant JSON for downstream accounting and CRM destinations, poor scans and edge case layouts, and field level validation so low confidence fields surface for the owner rather than filing silently.
• Tool use and workflow automation: tool and function definitions, multi step orchestration with conditional logic across calendar, email, CRM and accounting connectors, idempotency, retries and error handling, and approval steps, audit trails and safe rollback so automated actions stay reversible and under owner control.
• Evaluation, safe deployment and cost control: test sets from de-identified cases, scoring rubrics calibrated against the extraction accuracy Hello Genie publishes, regression testing, prompt injection handling for adversarial content in inbound email and uploaded documents, and model selection and prompt caching against the monthly price tiers.

Reversibility as a design principle, not a feature

Because every automated action is real, the program treated owner control as structural. Approval steps, audit trails and safe rollback were established as patterns for every action the assistant takes, and low confidence extraction was designed to surface for attention rather than file silently. The point was not to make the model more confident, it was to make the system honest about when it is not.

No customer data, written into the contract

Kriv AI did not request, receive, access, store or process personal information relating to Hello Genie users or their customers, mailbox or calendar contents, real invoices, contracts or business records identifying real people or organizations, or production credentials, and held no access to production systems, data stores or connected third party accounts at any point. Every teaching example ran on synthetic or de-identified data the customer supplied or approved, and all exercise work was executed by Hello Genie participants inside their own environment. The engagement also explicitly excluded connector development with third party providers, security certification and penetration testing, and any determination of the accuracy or sufficiency of an extraction, scheduling decision or customer communication produced with Claude assistance. Those remained Hello Genie's responsibility and the contract said so.