**Bottom line:** The analysts making the noise can't actually stop this — your model-validation function can, and it will do it quietly by keeping the pilot "not yet validated" until next quarter has slipped. Lock validation into a dated, itemized production sign-off this week, or your cautious board will "prudently defer" the rollout for you. --- ## Stakeholder Map *Assumptions stated plainly:* I'm treating "analysts fear redundancy" as the loud, visible signal — real, but not the binding constraint. I assume the risk department contains an independent model-risk/validation function (standard for a regulated bank this size, even if lean or partly outsourced) that holds the production sign-off pen, and that fraud reports up through risk — so risk *leadership* sponsors while model validation *inside* risk stays independent. I assume the card network is a major scheme (Visa/Mastercard) running its own auth-level fraud scoring. Customers are the ultimate beneficiaries and victims of false declines, but they aren't a group you negotiate with this week — their interests are carried by branch staff (experience) and compliance (fairness), so serving those two serves them. | Stakeholder group | Overall stance | Priority to address | | --- | --- | --- | | Model risk / validation team (risk dept, 2nd line) | Neutral | High | | Fraud analysts (front-line ICs) | Resistant | High | | Compliance | Neutral | High | | Senior leadership & the board | Neutral | Medium | | Risk leadership / fraud strategy owner (sponsor) | Supportive | Medium | | IT & security | Supportive | Medium | | Branch staff | Neutral | Medium | | Card-network partner | Supportive | Low | --- ## Group Detail ### Model risk / validation team (risk dept, 2nd line) - **Gains day-to-day:** A properly documented, validated model in the inventory means a clean exam story and no scramble when a regulator asks how automated customer-impacting decisions are controlled. - **Fears losing:** Independence and cover. Their entire value is being the brake; a fast rollout threatens their identity and risks making them the function that "waved through" an unvalidated Tier-1 model that later produced biased declines or an exam finding. - **Proof to earn trust:** A validation plan with firm dates (not open-ended), a challenger-model benchmark, disparate-impact testing on the *decline* population, and explainability documentation they can defend to an examiner. - **Fair participation:** Bring them in now as co-authors of the go-live gate, not last-minute approvers. Let them write the exact validation checklist so they own the standard rather than just enforcing it. - **If it fails / mitigation:** If the model misbehaves post-launch, they carry the exam finding. A documented validation trail *before* rollout is precisely what protects them — so early rigor is their insurance, not your delay. Frame it that way. - **First conversation + notes:** *"Tell me the exact conditions under which you'd sign this for production, and let's put a date on each one."* Bring: the pilot performance pack, draft model documentation, a proposed validation timeline, and an offer of a dedicated analyst + IT resource to close gaps fast. ### Fraud analysts (front-line ICs) - **Gains day-to-day:** The model auto-clears the noise — low-score approvals, textbook declines — so their queue shifts to the ambiguous, high-value cases where human judgment actually wins. Less mind-numbing false-positive triage, more real investigation. - **Fears losing:** Relevance and headcount. That they become rubber-stampers, that good pilot numbers become the excuse to cut the team, that their pattern-recognition expertise gets commoditized into a score. - **Proof to earn trust:** Pilot data showing the model routes *more* complex cases to them, not fewer people needed; a published rule that analysts can override the model with every override logged and reviewed (human-in-the-loop, not human-replaced); a respected senior analyst seated on the model-tuning group. - **Fair participation:** Their calls on pilot edge cases feed model retraining — their judgment literally trains the system. Give them visible credit and a standing seat in the tuning cadence. - **If it fails / mitigation:** If it flops they get blamed for "resisting" and morale craters. Early co-ownership makes a miss a shared learning rather than their sabotage. - **First conversation + notes:** *"I want you tuning this model, not competing with it — where is it getting cases wrong today?"* Bring: pilot route-to-human stats, the override-logging design, and a named seat on the tuning group. ### Compliance - **Gains day-to-day:** Lower fraud losses with *no* spike in complaints or discriminatory decline patterns is a regulatory win they can point to — fewer disputes, cleaner fair-treatment posture. - **Fears losing:** Control over customer-harm exposure. That automated declines hit certain segments disproportionately (UDAAP / fair-treatment risk), that a declined customer gets no explanation, that SAR obligations blur once decisioning is automated. - **Proof to earn trust:** Decline-rate breakdown by customer segment showing no disparate impact; a documented customer-explanation and redress path for false declines; confirmation that confirmed fraud still routes into the existing SAR workflow. - **Fair participation:** A standing seat on the model-governance forum and sign-off authority over the customer-facing decline and appeal language. - **If it fails / mitigation:** A discriminatory-decline pattern or complaint spike lands on compliance in the next exam. Early bias testing plus a redress path means they've already controlled it before rollout. - **First conversation + notes:** *"What would you need to see to be comfortable this doesn't create a fair-treatment or complaints problem?"* Bring: segment-level decline data from the pilot, a draft customer-appeal flow, and SAR-routing confirmation. ### Senior leadership & the board - **Gains day-to-day:** A defensible 30% fraud-loss reduction they can report *without* a matching rise in complaints or regulatory risk — a controlled win in a conservative house. - **Fears losing:** Reputational and regulatory safety. A headline about wrongly declined customers, an exam finding on ungoverned AI, being the board that approved an AI system that went wrong. - **Proof to earn trust:** Clean second-line sign-off (validation + compliance), pilot results showing false declines held flat, and a clear kill-switch/fallback if the model fails. - **Fair participation:** A concise pre-read and a decision framed as "controlled rollout with defined gates," not "trust the AI." - **If it fails / mitigation:** A public failure is their reputational hit. A staged rollout with a fallback and documented governance lets them approve with a safety net. - **First conversation + notes:** (via the sponsor) *"Approve the rollout plan and its gates, not the algorithm."* Bring: a one-page risk-and-controls summary, the gate criteria, and the fallback design. ### Risk leadership / fraud strategy owner (sponsor) - **Gains day-to-day:** Hits their fraud-loss target and modernizes the function — their bet pays off. - **Fears losing:** Credibility if the rollout stalls at the board door, or if their own validation team embarrasses them. They're caught between their independent second line braking and their own delivery goal. - **Proof to earn trust:** Pilot on track and second-line objections surfaced early — not sprung at the board meeting. - **Fair participation:** Co-own the go-live gate so the standard being met is theirs. - **If it fails / mitigation:** Their name is on it. Early de-risking of validation and compliance protects their credibility directly. - **First conversation + notes:** *"Your validation team is your biggest rollout risk, not the analysts — help me get them to a dated sign-off."* Bring: the validation gap list and a proposed timeline. ### IT & security - **Gains day-to-day:** A real-time scoring service they build and own, a modern MLOps capability, and fewer brittle manual rule hacks to maintain. - **Fears losing:** Stability and on-call sanity. Real-time latency in the authorization path, a model outage taking down approvals, PCI/data exposure, becoming the 3am owner of a black box. - **Proof to earn trust:** Latency budget met in the pilot (sub-100ms in the auth flow), a rules-based fallback when the model is down, and clear data-handling inside PCI scope. - **Fair participation:** They own the deployment architecture and the fallback design — not handed a model to just "plug in." - **If it fails / mitigation:** An outage or breach lands on them. Monitoring and a fallback built now prevent the 3am call. - **First conversation + notes:** *"What's your latency and fallback design if the model goes down mid-authorization?"* Bring: pilot latency numbers, integration points, and a fallback-rule proposal. ### Branch staff - **Gains day-to-day:** Fewer angry customers if false declines stay flat, plus a clear script and hotline when a card is declined — less improvising at the counter. - **Fears losing:** Face with customers. Being blamed for declines they can't explain or override, with no fast path to rescue a legitimate customer. - **Proof to earn trust:** A real-time override/verify path, a decline-explanation script, and pilot data showing false declines didn't rise. - **Fair participation:** Feed the false-decline cases they see at the counter back into tuning; a branch voice in the pilot review. - **If it fails / mitigation:** They eat the customer anger. An override path and script issued before rollout protects the relationship. - **First conversation + notes:** *"When a customer's card is wrongly declined, what do you need to fix it in two minutes?"* Bring: draft override path, decline script, and pilot false-decline numbers. ### Card-network partner - **Gains day-to-day:** Lower fraud and chargebacks on their rails; a bank using their tooling well. - **Fears losing:** Little. Mostly wants alignment between the bank's model and its own auth-level scoring, and no rule violations. - **Proof to earn trust:** Confirmation the model complies with network rules and doesn't conflict with network-level scoring or liability terms. - **Fair participation:** A technical sync on how the two scoring layers interact. - **If it fails / mitigation:** Low exposure. Early technical alignment avoids duplicate declines or rule friction. - **First conversation + notes:** *"How does your auth-level scoring interact with our model so we don't double-decline?"* Bring: model placement in the auth flow and a list of network-rule questions. --- ## The Silent Blocker - **Group most likely to block silently:** The **model risk / validation team inside the risk department** — the independent second line that holds production sign-off. - **How the resistance shows up:** No open objection. Instead: "we're supportive, we just need to validate properly," an ever-lengthening documentation request list, "let's run the pilot one more quarter to be safe," and sign-off conditions that quietly expand each time you meet them. Death by prudence, not by veto. - **Why they resist:** Pure incentive asymmetry. They get **zero credit** for the 30% fraud reduction — that's the business's win. But if the model produces a biased decline pattern or an exam finding, *their* independence is what failed. Delaying is career-safe; approving fast is not. A rational validator slow-walks. - **Formal argument they would use:** *"We fully support strengthening fraud controls, and the pilot results are encouraging — so this isn't about whether, only about sequencing. This is a Tier-1 model driving automated, customer-impacting decisions, and independent validation isn't complete: we haven't finished conceptual-soundness review, disparate-impact testing on the decline population, or the challenger benchmark. Deploying ahead of validation sign-off would be a finding waiting to happen at our next exam. Given our regulatory posture, the prudent path is to extend the pilot one more quarter and validate properly. We'd be doing the board a disservice to recommend otherwise."* Every sentence is reasonable, and in a conservative bank it wins the room. - **Not to be confused with the loud objector:** The **fraud analysts** are the loud objector — they'll voice the redundancy fear openly and often. But they don't control the rollout gate; they can grumble, not stop. Fix them for adoption and morale, not for approval. The validation team says less and holds the pen. --- ## Three Conversations This Week 1. **Model validation head** — *"I'm not asking you to lower the bar, I'm asking you to put dates on it. Tell me exactly what closes each gate, and I'll resource it this week."* Convert open-ended caution into a dated, itemized sign-off before it reaches the board. 2. **Senior fraud analyst lead** — *"I want you tuning this model, not replaced by it. Your calls become its training data, and every override is logged and reviewed."* Turn the loud fear into ownership and secure real adoption during the pilot. 3. **Compliance lead** — *"Show me what redress path and segment-level decline cut would make you comfortable — before this hits the board, not after."* Pre-empt the fair-treatment gate so it never becomes the board's reason to defer.