Operational Task Mapping
A continuous activity with three visible touchpoints across Weeks 1-3 plus a workshop touchpoint. Selection conversation with the CEO picks 3-4 functions to study. Per-function task work surfaces a Function Task Brief per lead. The workshop touchpoint converts the master task table into a ranked candidate set, then records a candidate-specific readiness decision and defined next move for each serious finalist.
Week 1: Selection conversation 60 min, remote, Brent + the CEO (plus anyone the CEO invites)
Pick 3-4 functions to study, with criteria written down, so the selection traces to documentation the CEO can hand to anyone who asks why these functions and not others. The CEO decides who else (a VP Risk lead, etc.) joins the conversation.
- Goal
- Pick 3-4 functions to study, with criteria written down, so the selection traces to documentation the CEO can hand to anyone who asks why these functions and not others.
- Facilitation moves
- Open with the survey synthesis and CEO interview signal (drag points named, areas of concern, comfort distribution). Walk the four criteria as evaluation lenses, applying each to the candidate functions. Capture rationale in writing during the conversation. Agree on observation opt-ins for 2 of the 4 functions. Assign async-vs-facilitated pre-work mode per lead (see Engagement modes card). Send lead-notification emails within 24 hours.
- Success indicators
- The CEO and any others she invited agree on the 3-4 functions. Written rationale exists for each. Each function has an assigned pre-work mode. Leads have been notified and confirmed. Observation invitations have been sent for 2 of the 4 functions.
- If drift, repair
- Function selection feels political rather than criteria-driven: revisit the criteria explicitly; if a function does not satisfy any criterion, surface that to the CEO. The CEO defers selection without reaching agreement: schedule a 30-min follow-up; do not proceed to Touchpoint 2 with ambiguous selection. Lead expresses concern or push-back on being selected: 15-min call with the lead before Touchpoint 2 begins to address concerns.
Four selection criteria
- Volume of work in the function. Does this function consume meaningful capacity that AI could reshape? Look at FTEs allocated, hours per week on routine tasks, throughput numbers Brent has from the Vendor AI Audit and the foundation pillar scores.
- Drag points named in the executive survey or CEO interview. Has this function been surfaced as a friction point in Sprint One's earlier work (survey responses on Q4 about what changed in the past 12 months, Q5 task naming, or CEO interview Q3 about a recent stall-or-move-forward moment)?
- Governance risk profile against the client authority record. Does this function trigger heightened review or approval under the client's policy? Functions with high governance complexity may need more careful pilot design and benefit from being mapped early.
- Strategic relevance to a Program of Work. Does this function tie to a named Program of Work in the the strategic plan strategic plan (Annual AI Plan, Copilot rollout, member-service initiatives, etc.)?
Multi-branch scope decision
If the credit union operates across multiple branches, functions may operate identically across branches, or they may have meaningful branch-level variation. Decide here whether the OTM work is institution-level (default) or accounts for branch-level differences for one or more functions. Default to institution-level unless the CEO flags branch variation as material for a specific selected function.
Lead-notification email template
Suggested copy. Adapt per lead.Observation-invitation template (for 2 of 4 functions)
Suggested copy. Adapt per function.Weeks 2-3: Materials and method what each functional lead receives and how the work is shaped
Each lead receives a packaged set of materials: the Function Task Brief template, the three-way tagging rubric, a one-page method packet drawn from the LIMRA AI Automation Identification Framework, and a framing message. Brent does NOT pre-populate function-specific task lists. The lead does the task decomposition themselves using the framework and rubric provided.
View the printable Function Task Brief template →
- Goal
- Equip each functional lead to break their function down into 12-18 atomic tasks and apply the three-way tag (Automate / Augment / Preserve Human) to each. Brent provides the framework, rubric, and template; the lead provides the function-specific operational knowledge.
- Facilitation moves
- Send the materials within 24 hours of the selection conversation (per U2). Include the lead-notification email's framing about what to complete async vs in the calibration session. For leads assigned to facilitated mode, schedule the guided pre-work session before sending the template (they complete it together with Brent, not alone).
- Success indicators
- Each lead has received the template, the rubric, and the method packet within 24 hours of selection. Each lead knows their pre-work mode (async vs facilitated). Each lead has a scheduled calibration session date.
- If drift, repair
- Lead doesn't respond within 3 business days: follow up directly (call, not email). Lead pushes back on the task-decomposition framework as too abstract: switch to facilitated mode and walk them through the first 3-4 tasks together. Lead's function appears too narrow for 12-18 tasks: revisit selection rationale; consider whether the function should be re-scoped or merged with an adjacent function.
Three-way tagging rubric
For each task, the lead applies one of three tags. The borderline subsection addresses tasks that sit between two tags or shift over time.
Automate
Remove the human; the agent runs this task end-to-end with exception-only human review.
Decision question: "Could the agent be the system of record for this task, with humans intervening only on exceptions?"
Worked CU examples populate from U8 (LIMRA-adapted research + Brent's review).
Augment
Keep the human in the loop; AI assists. The human decides; AI drafts, retrieves, classifies, summarizes, or suggests.
Decision question: "Does the value come from a human's judgment, with AI making that judgment faster or better-informed?"
Worked CU examples populate from U8.
Preserve Human
Keep fully human; AI deliberately stays out of the path.
Decision question: "Is this task one where AI presence would erode something Septapod is trying to preserve (member trust, mission integrity, judgment under uncertainty)?"
Worked CU examples populate from U8.
Borderline tasks
When a task sits between two tags: apply the decision question for the more conservative tag first (Preserve Human, then Augment, then Automate). If the answer is uncertain, default to the more conservative tag and capture the uncertainty in the lead's notes. Brent and the lead revisit borderline tags during the calibration session; tags can be revised based on conversation. The borderline subsection populates with specific worked examples drawn from U8.
Function Task Brief template
Each lead's owned artifact for their function. The brief is a standalone HTML page intentionally distinct from the wizard's app register, designed as a printed business document with serif typography, generous margins, masthead, and signature block.
→ Open the Function Task Brief template (opens in new tab; print to PDF for the lead's records)
Weeks 2-3: Engagement modes how Brent engages with each lead individually
Two decisions Brent makes per lead during or right after the selection conversation: which pre-work mode the lead uses (async or facilitated), and whether the function gets a 30-minute observation layer (2 of 4 functions, opt-in).
- Goal
- Match each lead to a pre-work mode they'll actually complete, and surface 2 functions where Brent observes a frontline doer to feed the calibration session with tacit-knowledge probes.
- Facilitation moves
- Default each lead to async mode. Switch to facilitated mode when the lead's function is unfamiliar to them at the task level, the lead's schedule rules out 60-90 min of focused async work, the lead's thinking is known to unlock better in conversation than alone, or the lead specifically requests it. For observation: nominate the 2 functions during the selection conversation; send observation invitations within 24 hours; brief the doer before observation begins.
- Success indicators
- Each lead has an assigned pre-work mode and knows what to expect. For async-mode leads: the template has arrived. For facilitated-mode leads: the guided session is scheduled. 2 of 4 functions have confirmed observation invitations.
- If drift, repair
- An async-mode lead doesn't complete pre-work by the calibration session date: convert that lead's calibration session to a facilitated pre-work session followed by an abbreviated calibration. A nominated observation function declines: substitute the next-priority function. Fewer than 2 functions consent to observation: the calibration session for non-observed functions uses enhanced interrogation prompts (longer drag-point probing, more open-ended workflow walking) in place of observation probes.
Pre-work mode per lead
Observation layer (2 of 4 functions)
30 minutes shadowing a frontline doer in their natural work flow, BEFORE the lead's calibration session. Brent watches for keyboard shortcuts, workarounds, system-switching patterns, decision rhythm, and the moves the lead would not think to write down. Notes feed the calibration session as specific probes.
Weeks 2-3: Calibration session 45-60 min per lead, with Brent
The synchronous moment where Brent and the lead interrogate the lead's pre-work together, refine tags, surface drag, and draft the synthesis paragraph. Five phases.
- Goal
- Convert the lead's pre-work into a finalized Function Task Brief with accurate tags, surfaced drag points, captured tacit knowledge, and a synthesis paragraph the lead signs off on.
- Facilitation moves
- Open with a 5-minute review of the lead's draft. Read it back to confirm understanding. Then interrogate (15-20 min) using the prompts below. Then validate tags (10-15 min): apply the decision question for each task, push back when something feels mis-tagged. Then draft the synthesis paragraph together (10 min). Then lead sign-off (5 min). For observed functions, weave observation notes into the interrogation phase as specific probes.
- Success indicators
- The lead's Function Task Brief is signed off and exported to PDF. Brent has populated the master task table from the brief. Tags feel defensible: the lead can explain each one. Drag points are named (not just inferred). Tacit knowledge surfaced: at least 2-3 things the lead didn't write down but mentioned during conversation.
- If drift, repair
- The lead pushes back on a tag Brent wants to change: revisit the decision question for that tag; if the lead's reading is reasonable, accept it and note the variance for the workshop. The session runs long: drop the synthesis paragraph drafting (Brent writes a v0 after, lead reviews async). The lead seems guarded about a task: shelve it; sometimes the right move is to come back to it after building trust. A task the lead tagged turns out to not actually exist in the function: drop it; document the surprise.
Session structure (45-60 min)
- Opening review (5 min). Brent reads the lead's draft Function Task Brief back to them: function summary, 3-4 most prominent tasks, lead's initial tag rationale. The lead corrects anything Brent misread.
- Structured interrogation (15-20 min). Brent walks through the tasks using interrogation prompts (below). For observed functions, observation notes become specific probes (e.g., "I saw you switch between two systems several times on that inquiry; what was that?"). The lead surfaces drag points and tacit knowledge that didn't make the pre-work template.
- Tag validation (10-15 min). For each task with a non-obvious tag, Brent and the lead apply the decision question for that tag together. Tags get refined. Borderline tasks are flagged.
- Synthesis paragraph drafting (10 min). Together, Brent and the lead draft 3-5 sentences naming the function's strongest AI candidates and the principles that should guide deployment in this function. Drafted live in the document, not after.
- Lead sign-off (5 min). The lead reads the full brief back. Any final corrections. The lead signs the brief. Brent confirms it's now their owned artifact.
Interrogation prompts (from AI Use Case Discovery Whitepaper, adapted)
- Drag prompts: "When in your week is the most frustrating moment in this task? What's happening?" / "What's the workaround you've built that wasn't in the documentation?"
- Tacit-knowledge prompts: "If you stopped doing this tomorrow, who would notice first? How would they know?" / "What's the thing you do that the new hire would take six months to learn by watching?"
- Tag-pressure prompts: "If we automated this completely, what would break that you couldn't see today?" / "If we deliberately kept AI out of this, who would be served better, and who would be served worse?"
- Borderline prompts: "Why isn't this one easier to tag?" / "What changes about this task in a year? In five?"
Per-lead session notes
Weeks 2-3: Master task table Brent's view across all functions; lives as a Google Sheet
As each calibration session completes, Brent populates a master task table that aggregates all functions' tasks into one sortable, filterable view. The table is the synthesis artifact that converts per-function briefs into a ranked candidate set the workshop opens with.
- Goal
- Maintain a master task table that aggregates 50-70 tasks across the 3-4 mapped functions, tagged on the three-way decision, scored on impact and feasibility, sortable by composite score, with the top tier visible as the finalist set for the workshop's readiness decisions.
- Facilitation moves
- Create the Sheet from a Brent-maintained template before Touchpoint 2 begins. Populate per function as each calibration session completes (typically within 24 hours). Score impact (1-5) and feasibility (1-5) per task using the anchored criteria; composite = impact × feasibility. After the last calibration session, transcribe the three-way tag totals and the top 5 candidates into the prototype inputs below so the Diagnosis Summary populates.
- Success indicators
- The Sheet URL is in the input below. All 50-70 tasks present with function, tag, impact, feasibility, composite. The top tier (composite of 16 or higher, typically) is visible as the candidate set. Three-way tag totals transcribed. Top 5 candidates transcribed.
- If drift, repair
- The Sheet falls behind the calibration sessions: catch up the same day; the master table is the workshop's input and cannot be incomplete on the day of the workshop. Scoring feels arbitrary: re-anchor against the 5-point criteria; document Brent's reasoning in the lead's notes column so the exec team can challenge it at the workshop. The top tier ends up smaller or larger than expected (under 3 or over 10): surface in the calibration synthesis paragraph. Sometimes the function genuinely has few candidates, and that's data.
Sheet structure (columns)
Function | Task | Tag (Automate / Augment / Preserve Human) | Impact (1-5) | Feasibility (1-5) | Composite (1-25) | Drag notes | Observation notes | Lead's notes | Brent's notes
Sample rows are intentionally NOT shown here in the prototype. The Sheet is the source of truth; this card holds the column structure plus the link and the transcribed summary data Brent fills in after scoring.
Scoring approach
Impact (1-5): anchored to volume × strategic relevance × drag burden. 1 = low-volume task with no strategic tie and no drag. 3 = moderate volume OR moderate strategic relevance OR meaningful drag. 5 = high-volume task tied to a named Program of Work with significant drag.
Feasibility (1-5): anchored to data readiness × vendor availability × governance complexity × technical lift. 1 = no data readiness, no off-the-shelf vendor, complex governance, heavy technical lift. 3 = mixed. 5 = data is clean and accessible, off-the-shelf vendor exists, governance authority is clear from the AI Policy, lift is light.
Composite: impact × feasibility (1-25). Top tier (composite of 16 or higher, typically) becomes the first-pilot candidate set.
Sheet URL
Three-way tag totals across the master table
Transcribe from the live Sheet after the calibration sessions complete. These feed the Diagnosis Summary's Master task table reference section.
Top 5 candidate tasks (by composite score)
Transcribe the top 5 ranked candidates from the live Sheet. These open the workshop's candidate-review segment.
Function lead summary (for the Diagnosis Summary)
Capture each lead's name, function, and calibration completion status. Feeds the Diagnosis Summary's Function Task Briefs section.
Workshop touchpoint: candidate-review segment 60-90 min within the Sprint Two workshop
Inside the workshop, the candidate-review segment converts the master task table into candidate-specific readiness decisions. Three phases stay inside the existing envelope: lead presentations (12-15 min), exec team scoring (20-25 min), and finalist readiness and disposition discussion (15-20 min, with the AI-fit review integrated into the finalist discussion).
- Goal
- Surface the per-function task work to the exec team, narrow to serious finalists, and record the next responsible learning move for each finalist without forcing an operational pilot.
- Facilitation moves
- Brief each lead before the workshop on the 3-minute presentation format (function summary, top 2-3 candidate tasks, the lead's tag rationale, one drag point that motivated the candidate). Have the master task table and per-function briefs visible. Brent pre-scores impact and feasibility; the exec team adjusts during scoring. The group then examines AI fit and readiness evidence. The portfolio decision, proposed owner's acceptance, and policy approval remain separate records.
- Success indicators
- Each lead presented confidently. The exec team engaged with the scoring criteria rather than faction or politics. Every serious finalist has a source-backed readiness profile, effective disposition, blocking conditions, and next move. The workshop can close with no operational pilot.
- If drift, repair
- A lead reads from notes: Brent picks up the thread with a question. Scores converge too quickly: ask each exec to score independently, then surface the deltas. The group vote names an owner who has not accepted: mark ownership blocked. Evidence or authority is missing: choose sandbox first, prepare first, or hold/stop rather than treating enthusiasm as readiness.
Segment structure (60-90 min)
- Lead presentations (12-15 min total, 3 min per lead). Each lead presents their function's mapped work to the executive team. Format: function summary (30 sec), top 2-3 candidate tasks with tags and rationale (90 sec), one drag point that motivated the candidate (30 sec), Q&A with the exec team (30 sec). Brent moderates time.
- Exec team scoring (20-25 min). The exec team reviews the top tier from the master task table. For each candidate, the team applies the impact and feasibility anchors (see Scoring template card) and either confirms Brent's pre-scoring or adjusts. Deltas between Brent's pre-score and the team's adjusted score are surfaced explicitly as data, not litigated.
- Finalist readiness and disposition (15-20 min). For each serious finalist, capture the six readiness dimensions, separate authority records, blocking conditions, and one of four dispositions: limited operational pilot, sandbox first, prepare first, or hold/stop. Impact, feasibility, and AI-fit evidence remain visible inputs but cannot bypass a readiness gate.
Materials in the room
- Each functional lead has their Function Task Brief (printed PDF or open laptop) for reference during their presentation.
- Brent has the master task table visible (Sheet projected or shared screen) with the top tier highlighted.
- Exec team has the scoring template (next card) visible, either projected, printed, or shared screen.
- A designated recorder captures the evidence and stated decisions. The client authority map supplies the portfolio decision authority; the group discussion itself does not create that authority.
Workshop touchpoint: scoring template 5-point impact x 5-point feasibility with anchored criteria
Two dimensions, five points each, composite = product (1-25). The exec team applies this rubric to the top-tier candidates during the workshop's scoring phase. Brent's pre-scoring (captured in the master task table) is the starting point; the team adjusts using the anchors below. Top tier (composite of 16 or higher, typically) becomes the serious finalist set for separate AI-fit and readiness review.
- Goal
- Give the exec team a defensible scoring rubric they can apply in 20-25 min. Each score level has a concrete anchor description so the rubric survives Architect-Dc scrutiny.
- Facilitation moves
- Display the rubric prominently during the scoring phase. Walk through one candidate together as a calibration exercise before the team scores in parallel. When the team's score diverges from Brent's pre-score, surface Brent's rationale (captured in the master table) and let the team decide whether to adjust.
- Success indicators
- The exec team can articulate why each top-tier candidate scored where it did. Score deltas (Brent vs team) are explicit and addressed. The final top-tier ranking after team scoring is an input to the readiness review, not a readiness or authorization decision.
- If drift, repair
- Scoring feels arbitrary or rushed: pause and walk one candidate through the anchors out loud as a calibration exercise. Team scores converge too quickly on Brent's pre-scores (suggests deference): ask each exec to score independently before discussing. Wide score divergence between execs on the same candidate: that's data, not a problem. Surface the divergence and ask each exec what their anchor was.
Impact (1-5): volume x strategic relevance x drag burden
| Score | Anchor |
|---|---|
| 1 | Low-volume task with no strategic tie and no drag. Examples: rare ad-hoc requests with no pattern. |
| 3 | Moderate volume OR moderate strategic relevance OR meaningful drag. Most tasks land here by default; reserve 1 and 5 for clear outliers. |
| 5 | High-volume task tied to a named Program of Work in the the strategic plan strategic plan with significant drag (lead-reported friction, observed workarounds, capacity that could be redirected). |
Feasibility (1-5): data readiness x vendor availability x governance complexity x technical lift
| Score | Anchor |
|---|---|
| 1 | No data readiness (data is scattered, dirty, or inaccessible), no off-the-shelf vendor solution, complex governance (heightened policy approval required), heavy technical lift (custom build). |
| 3 | Mixed: some dimensions ready (e.g., good data, available vendor) but others harder (e.g., governance complexity or technical lift). Most candidates land here by default. |
| 5 | Data is clean and accessible, an off-the-shelf vendor solution exists, governance authority is clear from the client policy and authority record with no additional approval triggered, and lift is light (configuration rather than custom build). |
Composite (1-25)
Impact x Feasibility. Top tier (composite of 16 or higher, typically) becomes the serious finalist set. Below 8 is generally not worth pursuing as an early pilot. 8-15 may be worth pursuing later but not now. The composite does not establish readiness.
Workshop touchpoint: AI fit and learning-quality gate 10-15 min inside the existing finalist review
Impact and feasibility identify valuable work the credit union can plausibly deliver. This gate asks a different question: is the finalist a good match for current AI, and can a short pilot produce evidence strong enough to support a real decision? Apply it only to the top 2-3 finalists, not the full task inventory.
- Goal
- Keep an attractive, purchasable idea from becoming the first pilot when its output is difficult to verify, its essential context is tacit, or useful feedback will arrive too late.
- Facilitation moves
- Name how the candidate entered the conversation: a technology capability, an operational problem, or a practice the team wants to improve. Then cross-check the other two. Walk the six dimensions quickly. Any weak dimension needs a concrete mitigation or the candidate is reshaped or held.
- Success indicators
- Each serious finalist has a named learning advantage or a named reason to reshape or hold it. The team can explain what evidence would change the next-move decision.
- If drift, repair
- The team treats vendor availability as proof of AI fit: return to verifiability and feedback speed. Scores become a second ranking contest: stop scoring and name the decisive unknown. A finalist is strategically important but weak for a first pilot: keep it on the roadmap and select a faster-learning pilot first.
Verifiability
Can the team tell whether an output is correct without creating a second full workflow?
Context accessibility
Is the needed context documented and available, or held in people's heads and local workarounds?
Feedback density
How quickly will the pilot reveal whether the output helped, harmed, or changed nothing?
Semantic stability
Do the people involved share a stable definition of a good result?
Source trustworthiness
Are inputs factual and clean, or filtered through incentives, self-reporting, or unresolved quality issues?
Pattern availability
Is this a common, well-represented task for current AI, or rare and proprietary?
Workshop touchpoint: candidate-specific readiness decisions pre-filled from existing evidence, confirmed inside the current 60-90 min segment
Create one record for every serious finalist. Brent pre-fills the known evidence before the workshop from the master task table, Function Task Briefs, AI-fit review, policy-authority map, and existing conversations. The workshop confirms gaps and chooses the next responsible move. This replaces the forced single-pilot output without adding participant time.
- Goal
- Produce a source-backed readiness profile and defined next move for each serious finalist. Valid outcomes are limited operational pilot, sandbox first, prepare first, and hold or stop.
- Facilitation moves
- Review what is already known, then ask only about evidence or authority that can change the disposition. Keep the portfolio decision, proposed owner's acceptance, and policy approval as separate records. Treat the effective disposition as the truth in the UI, synthesis, and export.
- Success indicators
- Each finalist has six dimension records, evidence sources, unknowns, status, next actions with owner and date, an effective disposition, and a pause or stop condition. No overall readiness score is calculated. A blocked condition prevents a limited operational disposition.
- If drift, repair
- Discussion runs long: stop debating scores and record the unresolved evidence as a next action. A group vote names an absent owner: mark owner acceptance unresolved. A candidate is attractive but blocked: choose sandbox first or prepare first. A legacy selection record exists: preserve it as history, but do not treat it as readiness.
Disposition boundaries
- Limited operational pilot: requires confirmed purpose and evidence, accepted ownership, real capacity, affected-worker confirmation, documented scope/data/systems/vendor/human oversight, applicable governance approvals, fallback/incident/stop path, and evaluation/monitoring evidence.
- Sandbox first: requires a named owner, purpose, boundary, learning question, and safe stop. It does not authorize real data or operational use.
- Prepare first: records the readiness work, owner, evidence, and date needed before reconsideration.
- Hold or stop: records why the candidate leaves the active path and what, if anything, would justify reopening it.
Authorization boundary: Sprint Two readiness never grants real-data, operational-use, or scale authorization. Those permissions are separate, sequential Sprint Three decisions loaded from the client's policy and authority record.
Workshop Planning
The on-site workshop in Sprint Two, structured to hold what the exec team can only do together: strategic identity broadening, the Playing to Win cascade, and the candidate-review segment from Step 1's Touchpoint 3 (lead presentations, scoring, and candidate-specific readiness decisions). Pre-work and post-work bracket the synchronous workshop. The workshop runs as one on-site trip in one of two formats: a single 4-5 hour block, OR ~6 hours split across two consecutive days.
Pre-Workshop: Async Collection (Week 4)
Embryonic Issues Surfacing 15 min async per person, individual written responses
Four prompts sent to each exec individually during Week 4. Written responses collected async. Brent synthesizes anonymous themes into the pre-read package. Individual responses produce more honest answers than live group discussion, and the themes inform the workshop's strategic conversations naturally.
Challenge Statement Builder
Source: PAIR Workshop Facilitator's Guide. Brent drafts v1 pre-workshop from Steps 1-3 data; team refines live in ~15 min.
Template: "How might [organization] use AI to [desired outcome] without [key constraint]?"
On-Site Workshop (Week 5, 4-5 hours single block or ~6 hours over two consecutive days)
What Business Is Septapod Really In? 25 min, opens the strategic conversation
One broadening question before the cascade. Without it, the cascade tends to produce incremental answers anchored in current operations. With it, real strategic options become visible.
Pick the framing that fits the CEO's style best:
- "What business is Septapod really in? Not what does Septapod do, but what does Septapod contribute to?"
- "For what purpose does Septapod do what it does? What are the ultimate benefits members get that go beyond the financial products?"
- "Strip away the products and services Septapod offers today. What would members lose if Septapod stopped existing tomorrow that they couldn't easily replace with another financial institution?"
Tighe's worked example: a public library broadened from "access to information" to "solutions to society's information needs." That single reframe opened creative and community-library strategic options that had been invisible. Septapod's equivalent might surface community wealth-building, financial wellbeing, or trust-based relationships with money.
Playing to Win Cascade for AI
Source: Playing to Win (Roger Martin), adapted for AI strategy context
Post-Workshop Follow-up
PAIR Activity Selector
Source: PAIR Workshop Facilitator's Guide / PAIR Guidebook
Brent selects PAIR activities based on workshop outputs and proposes to the CEO in a 30-min follow-up call. These activities shape how Sprint Three pilots are facilitated.
Direction Synthesis
Aggregated view of what Sprint Two produced: operational task mapping, strategic choices, and candidate-specific readiness decisions. Print or copy as markdown for the strategic direction document and board briefing. The synthesis can close with no operational pilot.
Governance Readiness Assessment: the supervision lens
A component of the Governance Readiness Assessment · applies to any pilot where AI acts without a person approving each action
The first pilots a credit union runs are typically AI assistance: a person stays in the loop and approves the output before it goes anywhere. Those pilots still follow the client's policy and authority record. A smaller set have AI take an action on its own, such as sending a member message, routing or approving a request, or making an access decision. The client adaptation must load any heightened approval for those actions from its own governing sources.
For any function where AI would act without a person approving each action, answer three questions in writing:
- Who supervises it? Name the person or system that watches what the AI does and can stop it. Define what they watch for and how often they look.
- Where is it boxed in? Name the limits on what the AI can reach and do: which systems, which records, which dollar amounts, which member segments. Everything outside the box is off-limits by default.
- Which actions need a person to approve before they run? List the actions the AI may never take on its own (moving money above a set amount, closing or freezing an account, denying an application, reporting a member to a credit bureau). Those actions stop and wait for a person.
An answer of "not decided yet" on any of the three means the pilot is not ready to run without a person approving each action. Run it with a person in the loop until all three are answered.
Source: Adapts the agent-as-insider-threat frame from Google DeepMind's AI Control Roadmap (v0.1, June 2026) to credit union scale. The full framework's threat taxonomy and control levels are built for organizations deploying autonomous agents with a security team behind them. The three questions here carry the usable core for a CU running its first pilots.
What Sprint Two Delivers
Four durable outputs · the summary sections below are working artifacts that compose them
- Strategic Direction for AI. Where AI gets investment and where it does not, across learning, communications, capacity building, and implementation. Built from the Playing to Win cascade and the ranked task mapping.
- Governance Readiness Assessment. Shows where the AI Policy's existing authorities are operational and where they need support before pilots begin. For any pilot where AI would act without a person approving each action, it carries the supervision lens (who supervises it, where it is boxed in, which actions wait for a person).
- Board Briefing. A 1-hour delivered session that establishes the recurring format for ongoing board AI conversations.
- Candidate-specific readiness decisions. Each serious finalist has a six-part evidence profile, separate authority records, one of four dispositions, and a defined next move. Sprint Three separately authorizes any use of real data, operational exposure, or scale.