Hook: the agency can only staff an accurate, approved request. Production variability is upstream context. The screenshot is the redesigned prototype with sample data, not a measured outcome.

Patrick observed the work across departments, operators, agency managers, and directors. He redesigned earlier planning, allocation, time, and HR flows. The independent AI-native product and build are his. Three-month duration retained from the existing opening draft.

Keep this to a brief definition before explaining responsibilities. Brands place orders with a manufacturer; the manufacturer uses staffing agencies for temporary labor. Orient the audience to the relationship before moving to shared responsibilities. This illustrative image comes from the ShiftForge project assets and does not depict a ShiftForge customer, research participant, or verified temporary worker.

Use the sequence to orient the audience. The product starts with labor demand and carries records forward. It does not configure production lines. Reveal the time labels on the next click. This is a qualitative synthesis of Patrick’s cross-organization workflow observations, not a time-and-motion study. The time labels are directional estimates supplied by Patrick, not measured durations. Each step requires people to compare records, confirm commitments, or route exceptions across organizational boundaries. Traditional separate staffing, scheduling, time, and billing tools make a current cross-party status difficult to assemble. The examples illustrate why manual action accumulates; do not claim every incumbent lacks each feature.

These figures were already in the opening draft and earlier case-study files. Their source records, sample, and dates need confirmation. They describe reported baseline pain, not ShiftForge impact. Remove this slide from interview delivery if the figures cannot be substantiated.

Describe how the process moved from research evidence to a scoped set of AI-assisted and human-owned tasks. Interviews and focus groups surfaced workflow pain points. Process mapping identified an agentic workflow and guided prioritization. UI references informed a design system used to iterate from low to high fidelity. Functional prototyping included building a local LLM setup to verify agent workflows and creating test data to simulate processes. The workspace image illustrates the working environment, not the ShiftForge product interface. This does not claim training a foundation model or deploying the agents to production.

Introduce the three core workflows. Start with plan demand, approval and agency allocation.

Follow the four-stage planning flow from an AI-assisted draft through department approval and AI-assisted matching to agency assignment and publication. The final stage combines two agency actions; after publication, the system locks the week. The manual plan example comes next, followed by the cross-page assistant and Plan approval interface. The legend marks the primary action at each stage; a recommendation is never itself an assignment or approval. Department heads manage their demand; permitted operations roles can manage across departments; agencies alone assign/unassign workers. A changed demand request reopens approval, and workers confirm assignment requests outside this captured state.

The video shows a manager using the Add to plan form to enter multiple positions across three roles, selected weekdays, and Area 1. The shift requirement belongs to the scheduled opening, but a shift selector is not visible in this crop; do not point to one. Authorized department users manage their own demand; operations permissions may span departments. The user confirms the addition, and each department approves its request before agency fulfillment. The visible Add to plan button is a human action, not an approval.

Why this interface: existing operators navigate datasets, not only conversation. The assistant can open across pages to inform or assist with actions. Describe persistent memory as an opportunity if asked; do not claim it is demonstrated without showing it.

Climax: a recommendation does not equal a confirmed worker. The planner selects a worker and days, a manager approves the assignment, and the worker responds to the request. The modal shows the selection; manager approval and later worker confirmation by text or app are not shown in this capture. Agency users can assign and unassign workers; reassignment remains a human action. Immediate undo and permissioned audit amendments are part of the defined design; avoid implying they are visible in this modal. Video replacement: compare candidates, adjust days, select, undo.

Transition from planning and allocation into worked time, exceptions and human review.

Trace the four-step timesheets flow. Worker punches enter the record before the sequence shown here. System checks flag exceptions, and the assistant proposes a fix. The combined Resolve and Approve step represents supervisor resolution and manufacturer or agency approval of changes; it does not collapse their separate authority. The agency submits approved time to invoicing, where the invoice is generated from that record. The action colors show the primary responsibility. This is a proposed workflow, not measured runtime behavior.

Transition from approved time records into invoice reconciliation and the final send decision.

Trace the four-step reconciliation flow. The system drafts the invoice from approved hours and rates; AI flags issues; plant managers resolve them; the agency sends the approved invoice to accounts payable. The four stages are a design sequence, not a capture of one completed invoice transaction. A flag is a prompt to review, not a correction or approval.

The confirmation dialog is a captured prototype state with sample amounts. It shows the total and approved adjustment before the agency sends the invoice to accounts payable. Sending is a separate human action, and the dialog states the invoice cannot be edited after sending. Do not present the sample total as a measured or reconciled result.

Source: https://www.patrickteall.com/work/shiftforge, Impact section (accessed October 13, 2026). Patrick labels these figures simulated data from agent simulations compared with research baselines, with staffing-agent testing still ahead. The source reports ~80% less planning setup, invoices needing correction from 15–30% to 0%, and zero surprise overtime events. Patrick supplied the displayed 0 hours of unexpected overtime during slide review; the published source uses events rather than hours. Present every number as a forecast or simulation, never as measured live-product impact.

Set up the reflection with the unresolved product decision. These are proposed evaluation criteria, not completed research. If the model adds no decision-quality benefit, narrow its role to drafting and explanation.

Supporting source-of-record view. The capture shows roles/status/rates/hours. Patrick’s HR training-tag rule is described; do not claim a certification detail view is shown. Prototype records use sample data.

Gap identified: staffing tools know placements and rates; workforce tools know coverage and hours. Corrections and reconciliation sit between them. This is an internal competitive synthesis across 14 products: sourcing, scheduling, time, approvals, payroll readiness, invoicing, manufacturing fit, staffing OS fit, and agency economics. The four named vendors are representative examples; ShiftForge is the intended design target, not an incumbent measured in the same market scan. Staffing OS strengths: sourcing to invoicing, agency economics and payroll readiness; common gaps: manufacturing fit, UX and implementation. Manufacturing WFM strengths: scheduling, time compliance and labor-cost reporting; gaps: agency economics and staffing-specific invoicing. Recruiting CRM/SMB tools were also reviewed but had less scheduling or back-office depth. Scores on this slide are the presentation synthesis supplied by Patrick; the older archived CSV uses a different scoring scheme.