Capture the reason for a decision without turning every past practice into policy.
EXECUTIVE ANALYSIS | H. Max Ottesen
AI-generated conceptual editorial illustration. Fictional people and setting; not documentary evidence. Illustrated technicians recording context from a manual at a safe maintenance training bench.
When a delivery requires a special sequence, the instruction alone may not be enough. A later team also needs to know which constraint required it, who approved the exception and whether the circumstances still apply. That is the kind of knowledge I would try to preserve.
This article proposes a controlled knowledge workflow. It does not claim that a deployed AI system has captured an organization’s expertise or improved its performance.
Preserve context before generating prose
NASA’s published description of its lessons-learned training emphasizes complete, valid lessons grounded in understood causes. [1] That is a useful analogy for operating knowledge: retain the event and reasoning rather than treating a memorable answer as a universal instruction. It is not evidence of AI effectiveness.
For a hypothetical warehouse exception, I would record the relevant source, constraint, decision-maker and review date. Distinguish an approved procedure from a temporary workaround. If the cause is unknown, leave it unknown rather than asking a model to finish the story.
An assistant could be tested on preparing draft summaries and identifying contradictions between documents. The process owner would decide whether to change the formal standard. A searchable draft must remain visibly different from an approved instruction.
Retrieval does not confer authority
The original retrieval-augmented generation research combines retrieval with generation for knowledge-intensive language tasks. [2] It offers a technical approach to bringing documents into an answer; its benchmark results do not establish that a company knowledge tool will retrieve current policy or honor employee permissions.
I would therefore show the originating document, version and effective date beside the answer. If two approved documents conflict, the tool should surface the conflict for the responsible owner. It should not choose whichever wording produces a smoother response.
Test what the system should withhold
Start with one process and a deliberately limited document set. Test current and superseded instructions, ambiguous queries and questions outside the approved scope. Include records the test user must not see. Check both the answer and whether the underlying retrieval crossed a permission boundary.
NIST’s generative-AI profile identifies confidently false content as a risk. [3] For this use, I would treat an unsupported instruction as a substantive defect even when it resembles something an experienced employee might say.
Ask frontline staff whether the reference helps on a difficult case. Record where they still need judgment or clarification. A knowledge tool should make it easier to locate the right evidence and person; it should not conceal an unresolved question under a confident summary.
The aim is a repeatable way to preserve and review operating judgment. Experience remains essential, including the experience needed to recognize when yesterday’s exception should not become tomorrow’s rule.
SOURCES & EVIDENCE
[1] Jesse Leitner / NASA NESC Academy (Page updated 2024-08-16; talk date not shown). A Structure (and Examples) for Capturing Lessons Learned. https://nescacademy.nasa.gov/video/6e3615a5fbfc477fa65f92244597777c1d. Locator: Published talk description. Supports: The published lessons-learned structure and cause/context emphasis. Limitations: Only the page description was used, not a claim to have watched the video. Aerospace lessons are an analogy, not an AI knowledge-transfer evaluation. Checked: 2026-10-17.
[2] Patrick Lewis et al. (2020-05-22; revised 2021-04-12). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. https://arxiv.org/abs/2005.11401. Locator: Abstract. Supports: The retrieval-plus-generation research approach and limited benchmark scope. Limitations: Research benchmarks, not evidence of reliable access control, current company policy, or correct operational advice in a deployed company knowledge system. Checked: 2026-10-17.
[3] Chloe Autio et al. / NIST (2024-07-26). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf. Locator: NIST AI 600-1, section 2.2 and action MS-2.5-003. Supports: Confidently false generated content. Limitations: Identifies risks and suggested controls; it does not validate these proposed workflows. Checked: 2026-10-17.
HOW THE CONCLUSION WAS REACHED
The NASA description supports preserving context, while retrieval research establishes a technical approach rather than operational authority. Those premises support testing provenance, versioning and permissions as design requirements. They do not prove successful expertise transfer.
LIMITATIONS
No private knowledge base was accessed or tested. NASA video footage was not reviewed; only its published description was used. Retrieval accuracy, access control and knowledge-transfer outcomes remain unverified.
© 2026 H. Max Ottesen. All rights reserved.
← All articles & insights