When It Doesn't Work

Walter Grayson believes the problem was the institution.

Kenji Watanabe believes the problem was Walter.

Diane Reyes believes both of them are partially right and that neither of them listened to her staff.

All three of them are correct, and none of their accounts alone explains why the deployment ended nine weeks in, three weeks before the scheduled conclusion and before the deliverable was complete. The AI’s project timeline shows the deployment failing in week four. The people involved acknowledged it in week nine. The five weeks between those two moments is the most important part of this account.

Walter is 71. He spent twenty-eight years in supply chain management, the last eleven as a director at a regional food distribution company. He retired knowing how supply chains are designed, where they fail, and how to restructure a distribution operation to reduce cost and improve delivery reliability. He is not wrong that he knows these things. He applied to the BGO program because he had expertise he believed a food distribution nonprofit could use. He was correct that the expertise was relevant. He was not correct about how to use it.

Kenji is 24. He finished his operations management degree eighteen months before the deployment. He can build a data model and analyze a distribution network’s cost structure with tools that Walter does not know how to use. He can identify inefficiencies in a logistics system and produce a visualization that an operations committee can read. He is analytically capable and he knows it.

Diane is 52. She has run a food distribution nonprofit in suburban Atlanta for eight years. The organization collects, sorts, and distributes food from grocery partners and food drives to 340 community distribution points across the county. She knows her organization, her staff, and the specific operational constraints of a food distribution model that serves communities with highly variable pickup and delivery access. She asked BGO for help with a warehouse organization and route efficiency analysis. She did not ask for a restructuring of her core distribution model.

The pre-deployment needs assessment identified the warehouse organization and route efficiency analysis as the scope. Walter arrived with both of those in scope and with a structural critique of the organization’s distribution model that he had formed from the pre-deployment materials. He believed the structural critique was the most important thing he could contribute. He expressed it in week one.

The AI’s project timeline shows what happened before anyone said it was happening.

In week four, the session notes from both Walter and Kenji show declining specificity: the notes are shorter, the questions are fewer, and the detail of what was accomplished in each session is thinner than in weeks two and three. This is a pattern the AI monitoring has been trained to flag, because it correlates with deployments where the collaborative engagement is losing momentum without the parties acknowledging it.

In week five, Diane’s communication frequency with BGO coordination drops from twice a week to once. The AI has a baseline for her communication frequency across the pre-deployment planning period. The drop is measurable. The AI flags it as a disengagement signal, a pattern where the receiving institution begins managing the deployment relationship more carefully rather than engaging with it openly.

In week six, the deliverable timeline slips two weeks without a scope amendment request. The project had been on schedule through week five. The two-week slip, unaccompanied by any request for timeline adjustment or scope change, is the clearest signal the AI can read: the project is behind and no one is addressing it directly.

The AI flagged all three signals to BGO coordination. A coordinator reviewed the flags and made one check-in call in week six. The call reached Diane, who said the deployment was proceeding but the team was working through some methodological questions. The call did not reach Kenji separately. It did not reach Walter. The coordinator did not schedule a structured three-party check-in. The deployment continued for three more weeks.

Walter’s account of what went wrong: the institution was not ready to hear what its distribution model actually needed. He had seen this pattern in corporate logistics settings. An organization with an established way of doing things, a loyal staff, and a leadership team that had invested in the current system will resist a structural recommendation from an outside expert. He brought the structural recommendation because it was the right recommendation. The institution’s resistance was the problem, not the recommendation.

Walter is partially right. Diane’s organization did resist the structural recommendation. The resistance was real.

What Walter’s account does not include: the structural recommendation was based on his analysis of corporate food distribution, not nonprofit community food distribution. The two are different in specific ways that matter. Corporate distribution optimizes for cost and delivery speed. Nonprofit community distribution operates with volunteer labor, donated vehicles, community pickup relationships, and distribution points chosen for community access rather than route efficiency. Walter’s expertise in corporate supply chain is not inapplicable to this context. It required translation into the specific constraints of nonprofit food distribution. He did not fully perform that translation before arriving with the structural critique.

Diane’s staff told him this in week two. He heard it as resistance. It was also information.

Kenji’s account of what went wrong: Walter’s analytical framework was correct in principle but the data model he insisted on using was twenty years old in its structure. Kenji built a current-generation route efficiency model. Walter did not trust it because it did not produce output in the format he was familiar with from his corporate logistics work. The better model was not used because the Sage would not accept it. The deployment failed because Walter’s framework was the obstacle.

Kenji is partially right. Walter’s insistence on the familiar data model format slowed the analysis by two weeks. The current-generation model Kenji built would have produced a faster and more adaptable analysis.

What Kenji’s account does not include: the way he communicated his model’s superiority. He was right that the model was technically better. He communicated this in a way that foreclosed the collaboration, not through a single confrontation but through the accumulation of interactions in which his certainty about the technical superiority of his approach made Walter less willing to engage with the analysis at all. The Native who cannot work with a Sage because the Sage’s framework is outdated faces a genuine problem. The solution is not to be right, conspicuously, until the Sage concedes. It is to find the path that incorporates the Sage’s judgment even when the technical tools are the Native’s domain.

Kenji did not find that path. He was right about the model. He was not right about how to use it in this deployment.

Diane’s account of what went wrong: both of them came to solve a problem they had decided in advance was the problem, and neither of them listened to her staff in a way that would have changed their conclusions. The warehouse organization question her staff raised in week three, a specific bottleneck in the receiving process that had nothing to do with route efficiency, was not in either of their analyses. Her staff raised it twice. It was not addressed.

She waited three weeks to tell BGO that the deployment was not working, because she hoped the methodological disagreement between Walter and Kenji would resolve and the deployment would deliver something useful. By the time she made the call, her staff had lost confidence in the deployment and she was managing their frustration rather than managing the project.

Diane is also partially right. She is also partially responsible. The BGO pre-deployment needs assessment did not surface her staff’s resistance to the structural recommendations that were likely to emerge from an outside supply chain expert. This was a gap in the assessment process. It was also a gap in Diane’s preparation of her staff for what the deployment would produce.

The failure belongs to all three of them and to the deployment model’s pre-deployment process. Naming this clearly is more useful than assigning it.

The three failure categories this deployment illustrates are the three categories that BGO’s post-failure operational review identified as the most common sources of deployment failure across the pilot cohort.

Pairing incompatibility: Walter and Kenji had different frameworks for resolving methodological disagreements, and neither of them had the tools to bridge the gap. A good pairing requires the ability to disagree without the disagreement becoming about status or certainty. This pairing did not have that capacity. The pre-deployment matching process did not identify the incompatibility because it assessed complementary skills without assessing collaborative working style. That assessment gap has since been added to the matching protocol.

Institutional unreadiness: Diane’s organization wanted strategic capacity analysis and was not ready to receive a structural critique of its distribution model. The staff resistance was present before the deployment began. The pre-deployment needs assessment did not surface it because it spoke only with Diane, not with the department leads whose buy-in the implementation would require. That structural gap in the needs assessment process has since been added to the pre-deployment requirements.

Expertise mismatch: Walter’s supply chain expertise was applicable to the institution’s context but required translation that the deployment model did not build in time for. The matching process identified his expertise as relevant without assessing whether the Sage had the ability to translate corporate sector expertise into a nonprofit community service context. That dimension has since been added to the matching criteria.

The AI flagged the failure in week four. The deployment ended in week nine. The five-week gap between the AI’s signal and the human acknowledgment of the problem is where the operational revision was most needed.

BGO’s failure detection protocol is now more structured than it was during this deployment. A structured three-party check-in at week three is now a standard deployment component: separate conversations with the Sage, the Native, and the institutional contact, conducted individually before the parties have coordinated their accounts. The AI flag review protocol has been moved from coordinator judgment to a structured response trigger: when the AI produces a multi-signal flag, a structured check-in is required within five days, not a discretionary call. The BGO coordinator’s call in week six was discretionary. It did not surface the depth of the problem because it reached only one party and allowed that party to frame the situation before the coordinator had independent information from the other parties.

The ethical question the AI data raises is not yet fully resolved. The AI had enough signal in week four to recommend intervention. The BGO coordinator made one call in week six. The deployment ended in week nine. What is the obligation of an AI monitoring system that sees a failure developing before the people involved acknowledge it? The current answer is a structured response protocol. The protocol has limits that the people who designed it acknowledge. An AI that flags a problem and a human coordinator who makes a single call and receives a reassuring answer is not the same as an AI that flags a problem and a process that escalates proportionally to the signal strength. That escalation protocol is under development.

Walter, Kenji, and Diane each left the deployment with something.

Walter learned that the translation of corporate sector expertise into a nonprofit context is not automatic and requires explicit attention. He applied to a second BGO deployment three months later, to a nonprofit food co-operative with a logistics challenge that was closer to his direct experience. The BGO matching review flagged the better fit. He is two months into that deployment.

Kenji took a position at a logistics consulting firm that works with food systems organizations. His route efficiency model, the one he built for the Atlanta deployment, is now part of the firm’s standard toolkit for nonprofit food distribution clients. He adapted it based on the Atlanta experience: he added a module for volunteer labor variability and a constraint for community-access pickup locations. The model he built in the deployment was technically superior to what he replaced. The deployment taught him what it needed to be useful.

Diane completed the warehouse organization project with her existing staff, using a vendor assessment process that had been in her organization’s pipeline for two years. The route efficiency analysis was postponed. She is considering applying for a second BGO deployment when the model’s pre-deployment assessment process has been more fully developed. She said this in her post-deployment feedback. BGO published her feedback in its operational review. It is in the operational improvement record alongside the failure.

The failure made the model better. That is the only claim this account makes for it.

What Exists Now, What Is Coming, and What Requires Time

#

BGO pilot deployments are running without a fully developed failure detection and intervention protocol. The AI project management layer tracks deployment progress and flags timeline risks. The structured three-party check-in at week three and the revised AI flag response protocol described here have been implemented following the operational review of this deployment.

Within one to two years, failure detection protocols as a formal deployment component: AI monitoring of session quality, project velocity, and communication patterns generating structured intervention triggers, not discretionary coordinator review.

Within three to five years, prospective matching protocols that reduce pairing incompatibility through AI-informed pre-deployment assessment of communication style and working methodology; institutional readiness assessment as a standard pre-deployment requirement with department-level input, not only executive input.

Walter is two months into his second deployment. The model learned from the first.