AI at the Decision Table: How Teams Keep Ownership of Decision
Keeping the manager’s personal AI assistant open to challenge
When I bring my personal AI assistant, Seraphim, into a working discussion, I also bring my authority as the person leading the work. A colleague may hear the assistant’s answer as a direction I already endorse. Questioning it could then feel like questioning me.
That is a leadership risk I want to take seriously. A person can remain responsible for an action while having little influence over the question, the evidence or the assumptions behind it. Asking that person to approve the final recommendation does not necessarily repair the imbalance.
Decision ownership should include the power to change the question before anyone approves an answer.
Two published conversations with my colleagues Khun Pin and Khun Ploy provide a practical starting point. They show people clarifying circumstances, directing the assistant and identifying what still needs checking. These are edited records of our own work, not controlled evaluations. My concern here is what leaders can learn from those interactions.
Let the Person Doing the Work Change the Question
In one exchange, Pin asks Seraphim: “Please give me short steps, each explaining what to do and how.”
It would be easy to treat that as a request for shorter text. I see a more useful leadership question in it: can the person carrying the work forward inspect the proposed action?
Pin is discussing a work-permit case. She clarifies that the foreigner is the sole director and that no other authorized person can sign for him. I restate the scope of the discussion. Her request for practical steps follows those clarifications. The record shows her helping shape the question and the form of the response; it does not establish that the proposed route was legally correct or accepted.
Turning an explanation into steps can expose its dependencies. Which action is possible with the information already available? Which depends on an unverified condition? Who has the authority to decide whether that condition has been met?
Those questions give a request for simplicity more substance. A shorter answer is useful only if it preserves the conditions that matter. Removing an unresolved prerequisite to make the response easier to read would leave the colleague with a clearer instruction and a poorer basis for acting.
A leader can start by letting the colleague describe the circumstances directly and correct the assistant’s summary. Then ask the colleague to explain which part of the proposed action still does not fit the case.
This also places a limit on delegation. Knowing the operating circumstances does not make someone qualified to certify every legal or technical claim. Give colleagues influence over the problem they understand, and obtain specialist review for questions that exceed their expertise.
Keep a Reported Requirement Attached to Its Source
In another exchange, Ploy reads a carrier’s message in Thai. It requests product details for genus and species fields in shipping paperwork. Seraphim asks whether US Customs requested the information or the carrier is collecting it for its own processing. Ploy identifies the document as the carrier’s. The excerpt ends with a proposed review of the original messages and shipping documents. That review is not shown.
What remains unresolved is the basis for the requirement. Ploy’s account does not establish that no official requirement exists.
Consider two possible summaries of a similar situation. “The carrier requested this information” identifies who made the request. “Customs requires this information” adds a claim about its authority. Moving from the first statement to the second requires evidence.
That is an easy change to miss when a conversation is compressed into a summary. If the source disappears, the team may start solving a more authoritative version of the problem than the evidence supports. An assistant can then produce a coherent answer to a question the team has not established.
The practical response is to preserve the original wording and ask what supports the claimed obligation. A carrier may be conveying an official requirement, applying its own process or asking for clarification. The underlying message and relevant documents help distinguish those possibilities.
The team can authorize obtaining that evidence while leaving the substantive response undecided. A meeting that ends with a precise verification task has produced something usable. It does not need a definitive answer to every question it raises.
Make an Objection Change the Discussion
Inviting people to challenge AI is only the beginning. What happens after someone objects tells them whether that invitation matters.
If a colleague identifies a missing fact, reopen the recommendation. Restate the corrected premise and assess whether it changes the proposed action. The original recommendation may still stand, but the reason for keeping it should address the objection. If the issue remains unresolved, preserve it in the handoff.
In the Ploy exchange, I ask the assistant to use English for me and repeat the instruction after it responds in Russian. This is a small, visible example of directing the assistant’s participation. Substantive oversight asks more: can a colleague’s correction change the question, the evidence required or the action being considered?
The manager has a particular responsibility when the assistant is their own. Ask for the detail that does not fit, and give a specific correction a visible response. Do not interpret the absence of an objection as evidence that everyone understands or agrees.
Research supplies a reason to make checking explicit. Lee and colleagues found that greater confidence in generative AI was associated with less reported critical thinking among surveyed knowledge workers. This was a self-report study, not evidence that AI causes declining thinking ability. The study did not test the separate concern I raise about a manager’s own assistant.
Define What the Conversation Authorizes
Human presence alone tells us little about the quality of oversight. In a meta-analysis, Vaccaro, Almaatouq and Malone found that human–AI combinations performed better than humans alone on average, but worse than the better-performing human-only or AI-only condition. Results varied across tasks and settings. The finding cautions against treating participation itself as proof of a better decision.
For a team leader, the actionable question is what the person involved can actually do. Can they access the source material? Challenge a premise? Seek another opinion? Pause an action whose conditions have not been met?
Approval should be equally specific. Obtaining a document, preparing a response and authorizing its submission are different acts. A useful closing instruction identifies which one is permitted and what must be checked before moving further.
For example, a team might authorize a colleague to request the original message and clarify its basis while reserving approval of the substantive response. That gives the colleague a defined task and tells them which decision remains open.
NIST’s voluntary AI Risk Management Framework calls for clear responsibilities, executive accountability for AI risk decisions and defined roles for human oversight. In a small team, the same person may perform several roles. The distinction still matters: supplying context, checking evidence and accepting responsibility for action involve different judgments.
Responsibility becomes more credible when the person assigned the work also has access to the evidence and permission to escalate an unresolved issue.
Record the disagreement as well as the next action
A polished meeting summary can make uncertainty disappear. If it records that the team agreed to proceed, someone taking over the work still needs to know what proceeding meant.
I recommend a short handoff that preserves what was reported, what remains unverified and what action has been authorized. For a carrier query, an illustrative record could look like this:
Reported request: The carrier has asked for genus and species information.
Still to establish: The exact request and the basis for it.
Authorized action: Obtain and review the original message and relevant documents.
Decision still open: What response to provide, subject to verification and approval.
Responsibility: Name the person obtaining the evidence and the person authorized to approve the response.
This is a proposed example, not a record of what happened after Ploy’s exchange.
Keep material disagreement in that note. If someone thinks the request has been misunderstood, record the objection and what would resolve it. Otherwise, the summary may quietly convert an open question into apparent agreement.
Before handing over the work, ask the colleague to explain the action and its limits in their own words. A useful answer should make clear which evidence supports the action and which new fact would require a pause or another decision. If the colleague cannot explain that, revisit the handoff.
Pin’s request for usable steps and the unresolved basis for the request in Ploy’s exchange offer concrete places to begin. They make it possible to examine whether people can shape the reasoning before it becomes an instruction.
The colleague carrying the work forward should have enough evidence to explain the action and enough authority to pause it if a crucial premise fails.
References
Prudchenko, E. (2026). AI at the Decision Table: Working session with Khun Pin [Edited video and English transcript]. Article and recording notes; published English captions.
Prudchenko, E. (2026). AI at the Decision Table: Khun Ploy [Edited video and English transcript]. Article and editorial notes; published English captions.
Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, Article 1121, 1–22. https://doi.org/10.1145/3706598.3713778.
Vaccaro, M., Almaatouq, A., & Malone, T. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. Nature Human Behaviour, 8, 2293–2303. https://doi.org/10.1038/s41562-024-02024-1.
Tabassi, E. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0) (NIST AI 100-1). National Institute of Standards and Technology. GOVERN 2.1, 2.3 and 3.2. https://doi.org/10.6028/NIST.AI.100-1.
The published working sessions and editorial notes provide the source context for both episodes. They are first-party edited excerpts; the opening of the Ploy edit is reordered and a later transition omits intervening material. They do not establish regulatory correctness, successful applications or shipment outcomes.
The NIST AI RMF Core offers additional guidance on assigning roles and documenting human oversight.
Eugene Prudchenko is a venture architect based in Bangkok and a director at Burakorn Partners. He develops practical AI systems and operating workflows across ventures and cross-border business operations. His writing examines how people work with personal AI assistants while retaining judgment, responsibility and control.
