--- schema: 1 type: video title: >- Deploying AI Inside a Bank Is Not an Engineering Problem — Paden Gayle, Applied AI Engineers description: >- A bank's document exception assistant produces a convincing demo, but security, data owners and risk reviewers still cannot approve it. Paden Gayle uses that hypothetical project to show why a working model does not establish a route into production. The missing decisions concern where the system runs, who authorized the documents, who reviews a wrong exception and which existing controls the workflow must satisfy. Drawing on his engineering experience, Gayle makes discovery part of the build: find the people who own those decisions in the first week, then define the smallest useful system they can evaluate. When a customer cannot supply a complete specification, his method is to shadow the work, draft a concrete workflow and shrink it with the control owners. Autonomy starts with the least authority needed to solve the problem. A drafting assistant can gather evidence about errors, overrides and recovery before anyone grants it permission to take further actions. Gayle also outlines an evidence pack covering intended use, data lineage, failure modes, results by segment and accountable controls. Original inputs and outputs belong in that record; rerunning a model later cannot reconstruct exactly what happened. Aggregate performance can conceal a poorly supported document type or user group, so weak segments need explicit restrictions. The closing checklist turns unresolved approval decisions into named owners and next steps. It is a planning aid for deployment, rather than a claim that the illustrative bank project shipped or that a checklist guarantees approval. Speaker info: - https://www.linkedin.com/in/padengayle/ - https://padengayle.com/ Timestamps: 00:00 - A working bank demo that still cannot ship 1:27 - Finding the people who own approval 5:47 - Shadow, draft and shrink the workflow 8:42 - Start with the smallest necessary authority 9:57 - Scoping the hypothetical exception assistant 12:21 - Five files for the evidence pack 13:37 - The first week deployment checklist --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/3ZlQk4-knWM-youtube-d1a028af5b3cc8b5ccd0 Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- canonical: 'https://liquidgold.cc/watch/3ZlQk4-knWM' markdown: 'https://liquidgold.cc/watch/3ZlQk4-knWM.md' llms: 'https://liquidgold.cc/watch/3ZlQk4-knWM/llms.txt' parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-23T22:18:48.105Z' language: null children: [] videoId: 3ZlQk4-knWM durationSeconds: 907 publishedAt: '2026-10-13T15:00:00.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 11 endSeconds: 97 watch: 'https://liquidgold.cc/watch/3ZlQk4-knWM?t=11' heldReason: rights - ordinal: 1 text: null startSeconds: 11 endSeconds: 97 watch: 'https://liquidgold.cc/watch/3ZlQk4-knWM?t=11' heldReason: rights - ordinal: 2 text: null startSeconds: 97 endSeconds: 108 watch: 'https://liquidgold.cc/watch/3ZlQk4-knWM?t=97' heldReason: rights - ordinal: 3 text: null startSeconds: 97 endSeconds: 108 watch: 'https://liquidgold.cc/watch/3ZlQk4-knWM?t=97' heldReason: rights - ordinal: 4 text: null startSeconds: 108 endSeconds: 350 watch: 'https://liquidgold.cc/watch/3ZlQk4-knWM?t=108' heldReason: rights - ordinal: 5 text: null startSeconds: 108 endSeconds: 350 watch: 'https://liquidgold.cc/watch/3ZlQk4-knWM?t=108' heldReason: rights - ordinal: 6 text: null startSeconds: 350 endSeconds: 360 watch: 'https://liquidgold.cc/watch/3ZlQk4-knWM?t=350' heldReason: rights - ordinal: 7 text: null startSeconds: 360 endSeconds: 525 watch: 'https://liquidgold.cc/watch/3ZlQk4-knWM?t=360' heldReason: rights - ordinal: 8 text: null startSeconds: 360 endSeconds: 525 watch: 'https://liquidgold.cc/watch/3ZlQk4-knWM?t=360' heldReason: rights - ordinal: 9 text: null startSeconds: 525 endSeconds: 536 watch: 'https://liquidgold.cc/watch/3ZlQk4-knWM?t=525' heldReason: rights - ordinal: 10 text: null startSeconds: 536 endSeconds: 603 watch: 'https://liquidgold.cc/watch/3ZlQk4-knWM?t=536' heldReason: rights - ordinal: 11 text: null startSeconds: 603 endSeconds: 615 watch: 'https://liquidgold.cc/watch/3ZlQk4-knWM?t=603' heldReason: rights - ordinal: 12 text: null startSeconds: 615 endSeconds: 750 watch: 'https://liquidgold.cc/watch/3ZlQk4-knWM?t=615' heldReason: rights - ordinal: 13 text: null startSeconds: 615 endSeconds: 750 watch: 'https://liquidgold.cc/watch/3ZlQk4-knWM?t=615' heldReason: rights - ordinal: 14 text: null startSeconds: 762 endSeconds: 820 watch: 'https://liquidgold.cc/watch/3ZlQk4-knWM?t=762' heldReason: rights - ordinal: 15 text: null startSeconds: 762 endSeconds: 820 watch: 'https://liquidgold.cc/watch/3ZlQk4-knWM?t=762' heldReason: rights - ordinal: 16 text: null startSeconds: 820 endSeconds: 830 watch: 'https://liquidgold.cc/watch/3ZlQk4-knWM?t=820' heldReason: rights - ordinal: 17 text: null startSeconds: 830 endSeconds: 907 watch: 'https://liquidgold.cc/watch/3ZlQk4-knWM?t=830' heldReason: rights --- # Deploying AI Inside a Bank Is Not an Engineering Problem — Paden Gayle, Applied AI Engineers A bank's document exception assistant produces a convincing demo, but security, data owners and risk reviewers still cannot approve it. Paden Gayle uses that hypothetical project to show why a working model does not establish a route into production. The missing decisions concern where the system runs, who authorized the documents, who reviews a wrong exception and which existing controls the workflow must satisfy. Drawing on his engineering experience, Gayle makes discovery part of the build: find the people who own those decisions in the first week, then define the smallest useful system they can evaluate. When a customer cannot supply a complete specification, his method is to shadow the work, draft a concrete workflow and shrink it with the control owners. Autonomy starts with the least authority needed to solve the problem. A drafting assistant can gather evidence about errors, overrides and recovery before anyone grants it permission to take further actions. Gayle also outlines an evidence pack covering intended use, data lineage, failure modes, results by segment and accountable controls. Original inputs and outputs belong in that record; rerunning a model later cannot reconstruct exactly what happened. Aggregate performance can conceal a poorly supported document type or user group, so weak segments need explicit restrictions. The closing checklist turns unresolved approval decisions into named owners and next steps. It is a planning aid for deployment, rather than a claim that the illustrative bank project shipped or that a checklist guarantees approval. Speaker info: - https://www.linkedin.com/in/padengayle/ - https://padengayle.com/ Timestamps: 00:00 - A working bank demo that still cannot ship 1:27 - Finding the people who own approval 5:47 - Shadow, draft and shrink the workflow 8:42 - Start with the smallest necessary authority 9:57 - Scoping the hypothetical exception assistant 12:21 - Five files for the evidence pack 13:37 - The first week deployment checklist --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/3ZlQk4-knWM-youtube-d1a028af5b3cc8b5ccd0 Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- Canonical: https://liquidgold.cc/watch/3ZlQk4-knWM · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks --- #### Chunk 0: Document excerpt (0:00 to 1:37) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 1: Document excerpt (1:37 to 5:50) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 2: Document excerpt (5:50 to 8:45) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 3: Document excerpt (8:45 to 10:03) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 4: Document excerpt (10:03 to 12:30) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 5: Document excerpt (12:30 to 13:40) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 6: Document excerpt (13:40 to 15:07) > Held: transcript rights do not permit this chunk's text.