AgentsAutonomy & tool use
IMB, Noblis and AeroCore describe AI governance rules and two production agents at the Midwest AI Summit
Three Midwest AI Summit talks argue that AI boundaries and named human owners should come before code: IMB on what its FaithBot refuses to do, Noblis on accountability and new technical debt, and AeroCore on two narrow production agents.
From the Practical AI podcast episode of November 5, 2026 and cited sources: Before writing any code, IMB drafted an AI charter. [2] [3]
SOURCES[2] Transcript: Abundant Intelligence at the Midwest AI Summit… share.transistor.fm[3] Christian Post: 600,000+ using FaithBot christianpost.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- Practical AI podcast episode, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 5, 2026 podcast episode published
- Covered
- Nov 5 edition
- Evidence
- 10 sources across 8 sites; transcript sourced
- 1Practical AI episode pageshare.transistor.fm
- 2Transcript: Abundant Intelligence at the Midwest AI Summit…share.transistor.fm
- 3Christian Post: 600,000+ using FaithBotchristianpost.com
- 4AviTrader: AeroCore–Delta agreementavitrader.com
- 5Fortune: MIT report on 95% of pilotsfortune.com
- 6Eventbrite: Midwest AI Summiteventbrite.com
Reach: not measured yet
Before writing any code, IMB drafted an AI charter. The charter let it turn down a request for an AI counseling chatbot without debate. It also set limits on FaithBot: the bot won't pray, write sermons or give financial advice, and it points users in crisis toward people. Don Barger says FaithBot has about 2.5 million users in about 50 languages. That is his own figure. An April 2026 report counted more than 600,000 English-language users. [2] [3]
Noblis CIO Mark Lay says accountability can't be handed to AI. A named person has to own data protection, output quality, agent monitoring and outcomes. Over a year, Noblis reviewed more than 1,000 datasets and narrowed them to about 115–120 critical ones, each assigned to one of about 30 stewards. Lay also warns of new technical debt. In one case, a vendor upgrade broke every app teams had built on its platform. [2]
AeroCore runs two agents on Prediction Guard's Agent Forge. The first is a text-to-SQL agent in Microsoft Teams that reaches the database through MCP and returns its query along with the results. The second is a daily vision check that compares aircraft tail numbers in technicians' photos with what they typed into the app. Mismatches go to a human for review. After a month with no incorrect work orders, the month-end reconciliation meeting was cancelled. [2] [7] [8]
The evidence is speakers' accounts of deployed systems, not measured evaluations.
Read the full assessment
For teams, the takeaway is that these organizations credit three things for getting AI into daily use: refusals decided in advance, named human owners and tightly scoped compare-and-flag checks.
Executive brief
FaithBot is a faith chatbot run by the International Mission Board (IMB), a 181-year-old missions agency. Don Barger, IMB's director of innovation and AI, said in a talk that it now has about 2.5 million users across about 50 languages. In April 2026 Christian Post reported 600,000+ English-language users. The episode is Part 2 of Practical AI's Midwest AI Summit coverage and has three talks: IMB on AI boundaries it set before writing any code, Noblis on accountability for AI, and AeroCore on two production agents. The common thread is governance that comes before the build. The talks give their results as anecdotes, not measured evaluations.
What changed and event timeline
AeroCore wins Delta contract
Delta and Endeavor Air named AeroCore their exclusive engine foam-wash provider under a 10-year agreement. This is the operation the AeroCore agents now track.
FaithBot launches
IMB launched FaithBot with partner Chipp.ai, partly to handle the volume of messages it receives worldwide.
The "95% of pilots fail" report
MIT NANDA's report said most generative-AI pilots stall before scaled adoption. The episode's host later cites it as the contrast to AeroCore.
FaithBot passes 600,000 English users
The report counted only English users and said the tool also runs in 25 other languages.
Midwest AI Summit
The summit took place in Indianapolis under the theme "Abundant Intelligence."
Episode published
Practical AI released the Barger (IMB), Mark Lay (Noblis) and Darren Taillie (AeroCore) talks.
Capabilities and access
- FaithBot (faithbot.tools) is free and paid for by sponsors. No model names are disclosed.
- AeroCore built two tools on Prediction Guard's Agent Forge, which connects agents to external systems through MCP:
- a text-to-SQL agent, available inside Microsoft Teams;
Read the full section
- FaithBot (faithbot.tools) is free and paid for by sponsors. Users can stay anonymous or sign in to keep chat history. Barger says each of about 51 languages uses a different LLM with its own prompts and worldview framing (transcript 11:09). No model names are disclosed.
- AeroCore built two tools on Prediction Guard's Agent Forge, which connects agents to external systems through MCP:
- a text-to-SQL agent, available inside Microsoft Teams;
- a daily check that uses a vision model to verify work orders.
- No model versions are named for either AeroCore tool.
Technical analysis for researchers and developers
- Text-to-SQL: Database access goes through MCP servers instead of giving staff direct database credentials.
- Vision check: At 08:15 each day, a scheduled agent compares aircraft registration numbers in technicians' photos with what they typed into the app.
- Evaluation: No accuracy, false-positive or latency figures are given for either tool.
Read the full section
- Text-to-SQL: Database access goes through MCP servers instead of giving staff direct database credentials. Each answer returns the SQL query along with the results, so people can audit it (transcript 36:35).
- Vision check: At 08:15 each day, a scheduled agent compares aircraft registration numbers in technicians' photos with what they typed into the app. A mismatch raises a flag for a human to review. The team repeatedly tightened the prompt so the model returns only the number, with no guessing in prose. The same pass now also flags smoke/fume notes and work orders that cross midnight and look implausibly long (transcript 41:09–43:34).
- Evaluation: No accuracy, false-positive or latency figures are given for either tool.
Claims and evidence
- FaithBot scale: Barger's 2.5 million users and 50–51 languages (transcript 09:54) are self-reported.
- Noblis data inventory: Lay says Noblis reviewed more than 1,000 datasets over a year and narrowed them to about 115–120 critical ones with about 30 data stewards (transcript 22:05–23:18).
- AeroCore results: Taillie says the month-end reconciliation meeting (about 5–6 person-hours) was cancelled after a month with no incorrect work orders (transcript 44:28).
Read the full section
- FaithBot scale: Barger's 2.5 million users and 50–51 languages (transcript 09:54) are self-reported. The only outside figure is 600,000+ English users plus 25 other languages in April 2026 (Christian Post).
- Noblis data inventory: Lay says Noblis reviewed more than 1,000 datasets over a year and narrowed them to about 115–120 critical ones with about 30 data stewards (transcript 22:05–23:18).
- AeroCore results: Taillie says the month-end reconciliation meeting (about 5–6 person-hours) was cancelled after a month with no incorrect work orders (transcript 44:28). This is a single month, not an independent measurement.
- Possible bias: The host, Daniel Whitenack, is CEO of Prediction Guard, and AeroCore and IMB are both Prediction Guard users (Latent Space).
Context and prior work
- The 95% statistic: The host contrasts AeroCore with MIT NANDA's finding that 95% of generative-AI pilots stall.
- AeroCore's business: The company reports it completed its 10,000th foam wash in December 2025 (PR Newswire).
- Earlier talks: Lay's "leash" framing and the deterministic guardrails he mentions refer back to earlier summit talks (Part 1).
Read the full section
- The 95% statistic: The host contrasts AeroCore with MIT NANDA's finding that 95% of generative-AI pilots stall. The same report found the best returns in back-office automation, which is what AeroCore built (Fortune).
- AeroCore's business: The company reports it completed its 10,000th foam wash in December 2025 (PR Newswire).
- Earlier talks: Lay's "leash" framing and the deterministic guardrails he mentions refer back to earlier summit talks (Part 1).
Limitations, safety and contested findings
- Sermon writing: Barger says FaithBot refuses to write sermons and lectures users who ask (transcript 13:33).
- Crisis handling: Many FaithBot users disclose abuse or self-harm.
- Prayer: He removed the bot's offer to pray for users after noticing it during testing (transcript 11:31).
Read the full section
- Sermon writing: Barger says FaithBot refuses to write sermons and lectures users who ask (transcript 13:33). Yet he was quoted in April saying pastors use it as a research assistant to "ethically prepare" sermons (Christian Post). The two may be compatible (research help vs. ghostwriting), but the line isn't documented.
- Crisis handling: Many FaithBot users disclose abuse or self-harm. Barger says live human takeover of flagged chats was possible at 5,000 users but not at 2.5 million (transcript 12:57).
- Prayer: He removed the bot's offer to pray for users after noticing it during testing (transcript 11:31).
- Vendor lock-in: Lay reports that a vendor upgrade broke every app teams had built on the platform (transcript 27:13).
Business and practitioner implications
- IMB used its charter to turn down a request for an AI counselor without debate.
- Lay's view is that an accountable human must always sit above the AI.
- Apps built by individual departments, vendor version changes, reversal costs and token costs all add up.
Read the full section
- Decide refusals in advance. IMB used its charter to turn down a request for an AI counselor without debate.
- Accountability needs owners. Lay's view is that an accountable human must always sit above the AI. In practice that means named owners for data, output quality, agent monitoring and outcomes.
- Watch new technical debt. Apps built by individual departments, vendor version changes, reversal costs and token costs all add up.
- Start with narrow checks. AeroCore's quickest wins were tightly scoped: look up a record, compare two values, flag mismatches for a human.
The source trail.
Sources (10)
Practical AI episode page
share.transistor.fmTranscript: Abundant Intelligence at the Midwest AI Summit Part 2
published_transcript
share.transistor.fm