Abundant Intelligence at the Midwest AI Summit Part 2

The episode’s central argument is that responsible AI starts with clear limits, not just technical capability. Don Barger said the 181-year-old International Mission Board built trust and openness to change by asking whether AI use was responsible. Almost four years ago, eight people spent a couple of days drafting an AI charter before writing code; its boundaries have since helped the team decide which projects to accept or reject. The IMB uses AI for drafting, editing, summarizing, research, translation, training, learning support, and code generation, with humans involved. Barger said this has shortened product development from years to a prototype in a day and shipment within months.

The IMB will not have AI impersonate a person in ministry, make final decisions about people, monitor people without their knowledge, or operate autonomously in high-risk situations. It also reserves pastoral care, counseling, prayer, discipleship, crisis support, and other soul care for human relationships. AI may help prepare support or screen and encourage people, but Barger rejected an AI counseling chatbot when human counselors were unavailable, arguing that the wrong intervention could be worse than none. His advice was to decide what AI should never do before building with it.

Donnie described Faithbot, a Christian chatbot designed to answer from a faith perspective, as having about 2.5 million users after roughly a year and a half, with growth by word of mouth and no paid advertising. It supports 51 languages, with a different large language model and setup for each language and worldview. Donnie removed the ability to offer prayers because he believes AI cannot pray and would only mimic its training data. Faithbot directs users toward human support, particularly in crises, and can recommend help services or crisis centers. It is programmed not to be hateful, coercive, or abusive, not to provide medical diagnoses or financial advice, and to discourage pastors from using AI to write sermons.

Users can sign in to retain Faithbot chats or remain anonymous; chats were not archived initially. Donnie can review flagged critical conversations and take over from the AI, but said that approach was feasible with 5,000 users, not 2.5 million. The service is free, not for sale, and supported by voluntary sponsors. He urged organizations to assess trust alongside feasibility, establish firm limits early, and recognize that clear boundaries can make development easier by narrowing what teams need to build.

The governance discussion emphasized matching safeguards to the likelihood and consequences of errors. Deterministic guardrails are appropriate when risks are high, and a person should remain accountable for AI actions, data protection, output quality, agent behavior, and outcomes. One organization identified more than 1,000 potentially relevant data sets, then narrowed its focus to about 115–120 operationally critical sets across roughly 30 data stores. Data owners need to understand data quality, timeliness, and access protections; controls in the system that stores data may not be enough for generative AI tools. Participants also warned that decentralized applications can be difficult to govern and create technical debt when tools or dependencies change. Mark Lay said organizations should weigh a solution’s value against its cost and whether IT can maintain it; more output is not valuable if the output is not worth the expense.

Darren Talley showed how those principles translated into daily work at ArrowCore, which uses patented nucleated-foam technology to wash jet engines and remove carbon buildup and fouling. The company analyzes engine performance before and after washes to assess efficacy, track trends, and investigate predictive maintenance and underperforming aircraft. It operates about 11 domestic sites and overseas sites in Japan, Riyadh, and Brazil, while its wash trucks have no digital components. Talley and one other employee make up the data analysis group, and database access is limited because data can be deleted too easily. Repeated data requests had created a bottleneck, with analysts spending as much as 40 hours a week responding, including to 55 separate requests for the number of washes in Atlanta the previous night.

ArrowCore used Agent Forge, MCP servers, and a Teams agent to let employees ask text questions about work orders and receive results with the underlying query. Moving the agent into Teams drove adoption and cut complex requests from 30–60 minutes to a couple of minutes; one query took about three seconds. A separate vision-model agent reviews photos submitted through a Power App, comparing aircraft details and work orders with technicians’ entries. It can read aircraft numbers, but is prompted to return required numbers without guessing aircraft identity. Mismatches and poor or incorrect photos go to people for review rather than being resolved automatically.

The system sends leadership a noon digest of the prior 24 hours of work orders across commercial and military sites, checking for issues such as smoke, fumes, abnormal procedures, unusual text, and work spanning midnight that may appear to last 36 hours when it likely lasted four. Flags are tracked for team review, with a human process to confirm, address, or dismiss them being built. Talley said work-order checks found no errors before a monthly finance, data, and operations review was cancelled, avoiding five to six person-hours, while vision verification reduced correction work. He said future models might gather information from trucks and photos or verify valve, temperature, and pressure settings without adding computers or complexity to trucks.