Parked 2026-08-21 after the client meeting. They liked run-rate-evaluation.md and the ROI workbook. They got lost in the Sprocket rationalization. This note is the board, not a client deliverable. Engagement context: [[2026-08-04-kwik-trip-ai-architecture-engagement-context]].
Client packet lives in the OI repo:
~/Documents/phdata-projects/organizational-intelligence/clients/kwiktrip/deliverables/engagement/roi-estimate/
What shipped
- Run-rate evaluation (Ben edited: hours freed is the broad figure; financials stay managerial; inputs tagged business / project team / FP&A-HR)
- Per-use-case workbook: Overview + Conservative / Base / Optimistic
- KT AI ROI Guidance transcription
- Sprocket rationalization (source-traced; most yellows unvalidated or EXAMPLE). Keep as an internal worked example, not the thing a requester reads.
Four assets
- FP&A Ground Rules (global). What the agent may assume: 1.3 load, Tampa Bay OEWS, 15% haircut, token unit-cost default. Also what it must leave blank: volume, occurrence, minutes now, any named dollar the business did not give. The run-rate evaluation doc is the human-facing page of this asset.
- Use Case Catalog (global). A view of existing use cases (Excel, SharePoint list, or database). Not a second system of record if OIP already is. Unpriced stays on the list.
- Use Case Discovery Brief (per use case). Working paper. Request details, anecdotes, open questions, who to ask next. What FP&A scrutinizes. Not a "rationalization."
- Use Case Financial ROI Model (per use case). The claim. Positive run-rate is the gate to prioritization, not the price of staying on the catalog.
Five actors
| Actor | Owns |
|---|---|
| Requester | Current work. Volume, occurrence, minutes now, named spend they already pay, qualitative, anecdotes. |
| Requester's manager | Corrects the team's details. Redeploy: keep as capacity, or name the work those hours go to. |
| Project team | Minutes after, count after, any revenue line. Hands the brief on. |
| IT | Systems and run cost (license, tokens, data). |
| FP&A | The claim. Haircut, capital, whether the brief is good enough to prioritize. |
| HR | Loaded rates, if those do not sit with FP&A. |
The agent is not an actor. It takes intake and writes the brief.
Seven intents
lookup · submit · update · request review · process review · approve · reject
These are intents (request types). The mapping exercise is intent mapping.
Three arrival storyboards, not seven:
- Find — lookup, status, what is already out there.
- File — submit new, or update one they own.
- Respond — they were brought in because someone asked. Approve / reject are last beats on this board when the actor is FP&A.
Actions (tool calls against the four assets) come after the storyboards, not before.
Copilot Studio note
Seeding a manager on a named use case is realistic as a Teams Adaptive Card from the agent (Power Automate: post as [[2026-08-03-every-microsoft-copilot-studio-trap|Copilot Studio]] agent, chat with agent, recipient = that manager). Card carries the case, requester inputs, hours freed, and the redeploy question. A "talk this through" button can open a Respond topic with the use-case id.
Catch: the manager must have opened the agent in Teams once. No prior personal chat, no proactive send.
Ground Rules reminder
Default: 1.3 load, Tampa OEWS, 15% haircut, token unit cost. Leave blank: volume, occurrence, minutes now, any dollar the business did not give. Do not invent a number so a use case can "proceed."
Open items (do not lose)
- Org map "10,000+ work orders a year" vs JOB.KT.024 100k–200k (central 150k). Unreconciled.
- CSR catalog rate vs workflow/gap saying senior techs answer the phone.
- Haircut cited as spec/05 §5.2; that section is not in the current spec/05 file.
- EXAMPLE material/capital on the Sprocket sheet. Pull before a client meeting.
- Storyboards not drafted. Blank discovery-brief skeleton not cut. Ground Rules page not written.
Next when this comes back
- Ground Rules page (human + agent).
- Blank discovery-brief skeleton (copy the headings, empty values).
- Three arrival storyboards, six to eight beats each.