Specify a work product the team can test and maintain.
EXECUTIVE ANALYSIS | H. Max Ottesen
AI-generated conceptual editorial illustration. Fictional people and setting; not documentary evidence. Overhead illustration of colleagues converting process cards into a plain internal form prototype.
Choose a work product before choosing an agent. For a maintenance meeting, that might be a reviewable action register tied to the relevant asset and evidence. For a capital request, it might be a clear account of missing documents and unresolved approvals.
These are proposed tools, not accounts of systems I have deployed. The useful question is whether a team can test the output, correct it and keep it working after the demonstration.
Define a bounded job
I would write the intended input, permitted output and approval boundary in plain language. For example: prepare a draft follow-up register from approved meeting notes; preserve the notes; flag ambiguous ownership; make no changes to maintenance status. An uncertain assignment should remain uncertain until someone confirms it.
A builder should also define failure. A fabricated owner, an omitted action and an action attached to the wrong vehicle are different defects. Testing only whether the register looks complete would miss the point.
Original research on a generative assistant in customer support provides evidence that an AI tool can be studied in an actual workflow. [1] Its results do not establish that a newly generated business application is reliable. Different tasks, users and consequences require their own evaluation.
Bring operating knowledge to the design
My Career Vault records operating-system development and hardware/software integration work at IP Security / IP Agency, alongside later fleet and logistics leadership. It supports that business-building background without establishing a software-engineering qualification. [2]
That perspective leads me to ask who owns the process, who can correct its data and who will support the tool when something changes. A generated application with no responsible maintainer should remain an experiment.
Test the uncomfortable cases
Use ordinary cases and cases with missing, contradictory or outdated evidence. Check whether a user can see the source behind an output. Test permissions with a person who should not have access. Confirm that a rejected draft does not become an approved operating record.
NIST’s generative-AI profile recommends verifying generated sources and citations. [3] I would apply that control to the tool’s evidence trail and add application tests suited to its actual functions. A citation check alone cannot establish that an application is secure or correct.
Compare the proposed assistant with a simpler alternative, such as a template, checklist or existing system feature. Include maintenance and review work in the comparison. A tool that needs constant rescue may still be an interesting prototype, but it has not earned operational reliance.
The output worth building is something the team can inspect, use and improve responsibly. The word “build” should raise the standard of proof, because a working business tool has obligations that a polished answer does not.
SOURCES & EVIDENCE
[1] Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond (2023-04-23; revised 2024-11-06). Generative AI at Work. https://arxiv.org/abs/2304.11771. Locator: Version 2 abstract. Supports: An actual-workflow research setting, without importing its productivity estimate. Limitations: Customer-support setting and a particular assistant; not evidence for these proposed logistics applications. No productivity figure is transferred to them. Checked: 2026-10-17.
[2] H. Max Ottesen (2026-10-15). Executive Career Vault and Resume Rules. CAREER_VAULT_SOURCE.pdf, Version 3 Final, sections 5–7, pp. 2–4. Locator: CAREER_VAULT_SOURCE.pdf, Version 3 Final, sections 5–7, pp. 2–4. Supports: The business technology/integration background, not an engineering credential. Limitations: Verified against the governing Vault only. Employer records, underlying calculation definitions and independent audit evidence were not supplied. 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: Verification of generated sources and citations. Limitations: Identifies risks and suggested controls; it does not validate these proposed workflows. Checked: 2026-10-17.
HOW THE CONCLUSION WAS REACHED
Empirical workflow research demonstrates a way to study bounded assistance. The career record supports the author’s relevant perspective. The recommended tests are a proposed acceptance approach, not proof of the proposed tools’ performance.
LIMITATIONS
No application built for this article was deployed or tested. Customer-support research does not validate maintenance software. Security, upkeep, permissions and user adoption need direct evaluation.
© 2026 H. Max Ottesen. All rights reserved.
← All articles & insights