How I Plan With AI

Most AI usage starts too late. By the time the model shows up, the team has already picked a direction, collapsed the uncertainty, and chopped the work into tickets. AI can still help draft, summarize, and speed things along, but the highest-leverage moment is already gone. The expensive part of most work was never the typing. It was deciding what the work actually is.

I would rather put AI to work before execution starts. I use it to widen the planning surface, surface missing constraints, pressure-test the shape of the problem, and make the decision record easier to inspect. I do not treat the model as an authority. I treat it as a tireless planning partner that can hold a lot of context, run pass after pass, and come back with its uncertainty laid out.

That distinction matters more than it looks. Use AI as an answer engine and the work tends to get more confident and less grounded. Use it as an investigation layer and the work gets more humble, more explicit, and easier to govern. The whole thing earns its keep by producing better questions before it produces a plan.

I usually start by pulling the planning problem apart from the writing problem. Rather than ask for a plan right away, I have the system map the current state: known facts, unknowns, constraints, dependencies, likely failure modes, and the evidence I would need before committing to a path. That keeps the model from jumping to a tidy sequence before it has earned one.

The work often runs in narrow passes: one for architectural risk, one for product ambiguity, one for operational cost, one for whatever would make the plan impossible to verify. The aim is narrow: catch the usual failure where one polished answer quietly hides five unresolved assumptions, without drowning in a bureaucracy of prompts.

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

The most valuable model output is often the one you do not want to hear. If one pass calls the work straightforward and another says it hinges on an untested integration, that contradiction is signal, not noise. The planning process should hold on to those conflicts long enough for a human to resolve them.

This is where AI improves planning without pretending to replace judgment. It can compare interpretations, surface missing definitions, and show where different parts of the system are pulling toward different priorities. It can also catch a subtle, dangerous planning smell: the plan looks easy only because the hard part got quietly renamed.

The human's job is to decide which contradiction actually matters. Some conflicts are real blockers, some are tradeoffs, and some are just the model getting confused. The win is that they are visible before the team starts building on top of an invisible assumption.

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

I like plans that break the work into bounded packages. Each one should have a reason to exist, a clear owner, and a verification path. AI helps here by keeping the plan mechanically consistent: do the packages actually cover the acceptance criteria, do the risks map to mitigation steps, does the verification really test the stated outcome?

This matters even more for agentic systems. Once the work involves models, tools, memory, retrieval, or autonomous execution, the plan needs more than feature acceptance. It needs behavioral checks, drift checks, trace review, and a way to learn from failures. A good planning artifact already knows how it will be tested.

AI planning does not earn its keep by removing approvals. It earns it by making approvals better 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 calling the work done.

The model can prepare the gate. It can summarize the evidence, flag the 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 starts behaving in ways nobody expected. Memory here does more than store what happened; it holds you accountable for the reasoning over time.

It also changes the emotional texture of planning. A team can move fast without pretending every uncertainty is settled. The plan can say plainly what is known, what is provisional, and what gets checked next. The work starts to feel less like a pitch and more like a controlled operating loop.

Used well, AI does not water down the rigor. It makes rigor cheaper, giving teams a way to investigate more, document more clearly, and move faster without turning speed into guesswork.