How I Plan With AI

Most AI usage still begins too late. A team has already chosen the direction, collapsed the uncertainty, and turned the work into tickets. At that point AI can help draft, summarize, and accelerate, but the highest leverage has already passed. The expensive part of most work is not typing the answer. It is deciding what the work actually is.

My preferred use of AI begins before execution. I use it to widen the planning surface, expose missing constraints, pressure-test the shape of the problem, and make the decision record easier to inspect. The model is not treated as a source of authority. It is treated as a tireless planning partner that can hold a large amount of context, run multiple passes, and return with structured uncertainty.

That distinction matters. When AI is used as an answer engine, the work tends to become more confident and less grounded. When AI is used as an investigation layer, the work tends to become more humble, more explicit, and easier to govern. The system is useful because it produces better questions before it produces a plan.

I usually start by separating the planning problem from the writing problem. Instead of asking for a plan immediately, I ask the system to map the current state: known facts, unknowns, constraints, dependencies, likely failure modes, and the evidence needed before committing to a path. This keeps the model from jumping to a tidy sequence before it has earned one.

The work often runs in narrow passes. One pass looks for architectural risk. Another looks for product ambiguity. Another looks for operational cost. Another looks for what would make the plan unverifiable. The point is not to create a bureaucracy of prompts. The point is to avoid the common failure where one polished answer hides five unresolved assumptions.

A good AI planning pass should leave behind artifacts that a human can audit: assumptions, decision points, candidate approaches, open risks, and what evidence would change the recommendation. If the model cannot explain what would falsify the plan, the plan is not ready.

The most valuable model output is often the uncomfortable one. If one pass says the work is straightforward and another pass says the same work depends on an untested integration, the contradiction is not noise. It is information. The planning process should preserve those conflicts long enough for a human to resolve them.

This is where AI can improve planning without pretending to replace judgment. It can compare interpretations, surface missing definitions, and show where different parts of the system imply different priorities. It can also catch a subtle but dangerous planning smell: the plan is easy only because the hard part has been renamed.

The human role is to decide which contradiction matters. Some conflicts are real blockers. Some are tradeoffs. Some are just model confusion. The value is that they are visible before the team starts building against an invisible assumption.

The end product is not a slide or a memo. It is an execution artifact. A strong plan should name the objective, the scope, the non-goals, the acceptance criteria, the likely risks, the sequencing, and the evidence required to call the work complete. It should be clear enough that another person can disagree with it in specific terms.

I like plans that break work into bounded packages. Each package should have a reason to exist, a clear owner, and a verification path. AI is useful here because it can keep the plan mechanically consistent: do the packages cover the acceptance criteria, do the risks map to mitigation steps, and does the verification actually test the stated outcome?

This is especially important for agentic systems. If the work involves models, tools, memory, retrieval, or autonomous execution, then the plan needs more than feature acceptance. It needs behavioral checks, drift checks, trace review, and a way to learn from failures. The planning artifact should already know how it will be tested.

AI planning is not useful because it removes approval. It is useful because it makes approval more informed. The right operating model keeps human gates at the moments where judgment actually matters: before committing to the plan, before expanding scope, and before declaring the work complete.

The model can prepare the gate. It can summarize evidence, highlight unresolved choices, and show what changed since the last decision. It can also keep a record of why a decision was made, which matters later when the system behaves differently than expected. Memory is not just recall; it is accountability for reasoning over time.

This also changes the emotional texture of planning. A team can move quickly without pretending that every uncertainty is resolved. The plan can say what is known, what is provisional, and what will be checked next. That makes the work feel less like a pitch and more like a controlled operating loop.

Used well, AI does not make planning less rigorous. It makes rigor cheaper. It gives teams a way to investigate more, document more clearly, and move faster without turning speed into guesswork.