Working with AI When the Problem Is Still Taking Shape

Begin with a purpose you can explain

A starting description needs enough context to orient the conversation: the outcome sought, the people affected, and the constraints already known. “We spend too much time chasing project updates and want more dependable handoffs” is enough to begin exploring. It does not yet establish the cause or prescribe the solution.

Consider an everyday illustration: planning a vacation. You may know your dates and budget, need to investigate local driving requirements, and have preferences you only articulate when someone asks about past trips. Discussing one destination may reveal a transport dependency that changes which hotel makes sense. Different parts of the same plan need different kinds of help.

The same principle can guide an enterprise conversation. When the destination itself feels constrained by current routines, use Possibilities to explore a different outcome. Select a question worth pursuing, then examine the conditions that would make it feasible.

Choose the help that fits the question

The familiar language of knowns and unknowns offers a useful starting lens. The distinction among known knowns, known unknowns, and unknown unknowns predates current AI tools and appeared in Donald Rumsfeld’s February 2002 briefing. [1] Here, a fourth state—knowledge or experience not yet articulated—is included to help shape the conversation.

The following mapping is RePerspective Labs’ proposed working aid. It is not a validated assessment of someone’s knowledge, a maturity model, or an automatic assignment of work to AI.

The third state is shorthand for experience that has not yet been made explicit in this discussion. It can also point toward knowledge held elsewhere in the organization, although another person’s knowledge is not literally your own. Ask whose perspective is represented before labeling a gap.

Apply the lens to a specific question or assumption. An entire initiative will usually contain several kinds of uncertainty. A statement initially treated as known may need investigation when its source or applicability is challenged.

Let AI propose the map—and keep it open to correction

Letting AI suggest the first breakdown can reduce the preparation needed to begin. Ask for a small set of connected questions and the assumptions behind them. Review that map before expanding it. A missing stakeholder, an incorrect boundary, or a premature solution can shape every branch that follows.

There is research behind using decomposition in AI interactions. Zhou and colleagues’ least-to-most prompting study examined breaking problems into simpler subproblems and using earlier answers in subsequent steps. Its experiments concerned specified reasoning tasks; they do not validate an enterprise discovery method or establish that an AI-generated map is complete. [2]

In practical use, the map should distinguish what someone reported, what a source supports, what has been inferred, and what remains unresolved. The distinction matters because a plausible explanation may still be incorrect. NIST’s Generative AI Profile identifies confidently erroneous content as a characteristic risk. [3] A claim should move toward “established” through suitable evidence, not merely because the assistant repeats it convincingly.

People should challenge the proposed framing early. AI can help formulate an interview question; it cannot substitute for the colleague whose experience is missing. Some questions will be resolved through a record, a conversation, or a small observation without further AI help. Nor does understanding a task grant permission to execute it. Authority remains a separate operating decision.

Follow the connections as understanding changes

Illustrative enterprise scenario. A team wants to reduce handoff delays while keeping consequential decisions with people. It begins with an impression that information is scattered. The cause has not been established.

An initial conversation might produce four connected questions:

• Outcome: What makes a handoff useful to the receiving team, and where does the current delay begin and end?
• Readiness: What do experienced colleagues check before accepting the work?
• Records: Where are the latest decision, its reasoning, and any changed dependencies recorded—and which source is authoritative?
• Exceptions: What situations would require a person to reconsider an otherwise routine handoff?

Suppose an interview reveals that an experienced coordinator checks for an unresolved dependency before treating an item as ready. That check was absent from the initial description. Drawing out experience has now created an investigation: how often is this dependency missing, and does it explain delays?

The team examines a defined sample of recent handoffs. Imagine that the sample suggests people use different meanings of “ready,” even when the relevant records are available. The next action changes: agree and test the readiness criteria with both teams before investing in another summary or integration.

If the evidence instead points to inaccessible or inconsistent records, knowledge access becomes the priority. Both paths remain hypotheses until checked. Their value is that a finding in one branch changes what happens elsewhere.

The method therefore needs a return path. Update the shared picture after a useful discovery. Merge duplicate questions, retire assumptions that no longer hold, and record which dependency changed the proposed action. Repeated smaller assessments should improve the whole decision.

The AI Puzzle explains the wider discipline of connecting operating capabilities. This Insight addresses the earlier discovery work: how to develop that picture while the questions themselves are still emerging.

Stop when the next commitment can be judged

Discovery needs a useful stopping point. Agree the decision and the time available before branching further. Continue exploring when an unresolved question could materially change the outcome, feasibility, or acceptability of the next step.

Pause the conversation when the team can describe one bounded action, identify who owns it, name the remaining uncertainty, and specify the evidence that would support proceeding, revising, or stopping. The strength of evidence should match the consequences of that commitment.

For the handoff example, an initial trial might compare elapsed handoff time, the share accepted without clarification, and total preparation and review effort against a baseline for comparable work. Check omissions and rework as well as speed. These are proposed tests, not promised results. Measurement provides further measures to adapt to the operating outcome, including first-time-right completion and human oversight effort.

When a question becomes a design decision about knowledge, coordination, or authority, use the Autonomy Map to examine those connections. It can help locate the next operating question; choosing to retain human approval can remain an appropriate outcome.

Try the conversation with a starter prompt

Paste the prompt below into your preferred AI assistant. Use an approved tool and share only material appropriate for it. You can start with a rough description; the prompt does not require a completed business case or a starting quadrant.

It is self-contained and does not require access to this website. It guides the conversation; it cannot ensure that the assistant will follow every instruction or produce reliable conclusions. Review what it proposes.

Help me work through a situation that is still taking shape. My aim is a clearer understanding and a practical next step.

Start by asking what I want to improve, for whom, and in what context, unless I have already told you. I can begin with a rough description or say “not sure.”

After that first answer:

1. Propose a small, provisional map of the main questions and dependencies. Base it on what I share. Label assumptions and invite me to correct the map before developing it further.

2. Use four kinds of help as appropriate: apply and check established knowledge; investigate a recognized gap; ask questions that draw out experience or unstated assumptions; explore alternative perspectives and possible missing considerations. You do not need to use all four. Do not claim to know everything I know or to have discovered every blind spot.

3. Ask one question at a time. Offer a few tailored options plus Other, Not sure, and Skip. When a choice calls for judgment, give a tentative recommendation and explain why. Let me combine options or answer freely. Wait for my reply.

4. Keep my statements, source-supported facts, your inferences, and unresolved questions distinguishable. Cite sources you actually accessed and flag what you cannot verify. Ask whose knowledge or which record could help. Do not treat a fluent answer, a confident label, or my agreement as verification.

5. When an answer reveals a dependency, examine it only if it could change the outcome, feasibility, or next decision. Explain which other question it affects. Update the map when evidence changes it; do not keep expanding branches without a reason.

6. After roughly five answers, or earlier if enough is clear, summarize the current purpose, what we have learned, the most important uncertainty, and one practical next action. Identify who should be involved, what evidence to gather, and what would lead us to proceed, revise, or stop. Ask whether to investigate further or close with that next step.

Keep this exercise to discussion, research, and preparation. Do not send messages, change records, or commit resources without my separate authorization.

If you want the conversation explicitly grounded in RePerspective Labs’ operating dimensions, the Autonomy Map’s “Use this map with your AI” section provides its own guide-based starter. Use whichever starting point suits the work; neither requires you to know every question in advance.