It is Tuesday morning, the project is chasing a topping-out milestone fourteen weeks out, and the superintendent has the trade foremen in a room with a wall of sticky notes and, this time, a screen. The AI pull-planning tool has read the master CPM and staged a candidate pull plan: the sequence of handoffs from the last milestone backward, each trade's work mapped onto the wall, the obvious conflicts pre-flagged. It looks finished. It would be easy to walk the room through the staged plan, ask for nods, and call it the pull plan. But a pull plan is not a schedule the AI drew; it is a chain of promises the trades make to each other, the drywall foreman promising the framing will be done and inspected by the date the painter is counting on, and a chain of promises nobody actually made is not a pull plan, it is a wall of sticky notes that will not hold. This lesson is about using AI to stage the candidates and flag the conflicts fast, so the room spends its ninety minutes on the work only the room can do: making the real commitments, surfacing the real constraints, and walking out with a promise from every trade. The AI stages the plan; the trades make the promises, because in the Last Planner System the promise is the product, and a promise the AI drafted but no one made is worth exactly nothing.

What Last Planner Pull Planning Is and Why It Works

The Last Planner System is a production-planning method built on a simple, hard-won insight: the people who do the work should plan the work, and they should plan it as reliable promises to one another rather than as directives handed down from a master schedule. Pull planning is the technique at its heart. The trades gather, start from a target milestone, and work backward, each trade defining what it needs from the trade before it and promising what it will hand to the trade after it, so the plan is pulled from the milestone rather than pushed from the start, and every activity in it is a commitment a foreman made out loud, in the room, to the people depending on it. The plan that results is dense with handoffs, and each handoff is a promise: this scope, done and ready, by this date, for you.

This works because of the lesson the recovery chapter established and this one completes: people deliver on commitments they made, not on plans handed to them. A master CPM can sequence the work logically, but the foreman who never agreed to the dates feels no ownership of them, and the schedule slips because no one promised it would not. A pull plan inverts that: the foreman who promised the framing by Thursday, in front of the painter who is counting on it, owns that date, and the social weight of a promise made to a peer is what makes the pull plan reliable where the handed-down schedule is not. The Last Planner System measures this reliability directly with Percent Plan Complete, the PPC, the fraction of the week's promises that were actually kept, which turns promising into a tracked, improvable discipline rather than a hope.

So pull planning produces a different kind of plan than the CPM produces, not a logic network computed from durations and dependencies but a web of human promises made by the people who will keep them, and its reliability comes from the promising, not the drawing. That distinction is the whole key to using AI here, because the AI is good at the drawing and cannot do the promising, so the question the lesson answers is how to let the AI accelerate everything about the plan except the one thing that makes it work, which is the commitment the trades make to each other in the room.

What the AI Stages: Candidates From the CPM and Real-Time Conflict Flagging

The AI's contribution to a pull-planning session is real and it is mechanical. Working from the master CPM, a tool like Touchplan with AI assistance can stage a candidate pull plan before the session: it reads the logic network, identifies the activities between the current state and the target milestone, and lays them out as a first-draft wall of handoffs, sequenced and dated against the CPM, so the room starts from a populated board rather than a blank one. This is the generative engine in the drafting role the program has named again and again, the same role it plays drafting a submittal log, an RFI response, or a recovery option: it produces a candidate fast, turning the tedious setup of transcribing the schedule onto the wall into a starting point the room can react to.

The second contribution is during the session, and it is the more valuable one: as the trades adjust the plan, move a handoff, compress a duration, re-sequence a stretch, the AI can flag constraint conflicts in real time. When the foreman pulls the inspection earlier to make a date work, the tool can flag that the inspector's lead time will not allow it; when two trades claim the same work area in the same week, it can flag the collision; when a compressed duration violates a cure time or a required sequence, it can surface that immediately rather than three weeks later when the conflict bites. This is the analytic engine doing what it does well, holding the whole web of constraints and checking the room's moves against it faster than a human facilitator scanning a wall of notes could, so the conflicts surface while the room can still resolve them.

The AI stages the candidate pull plan and flags the constraint conflicts in real time, but the promise behind every handoff is made by the trade that will keep it, because in the Last Planner System the promise is the product and a promise the AI drafted but no one made will not hold.

Both contributions are genuine accelerants, and both are squarely on the mechanical side of the line: staging the candidate is transcription and layout, flagging the conflicts is checking moves against constraints, and neither is the promising. The AI hands the room a populated, conflict-aware board so the ninety minutes are not spent transcribing the schedule or hunting for collisions, which leaves the room free to spend its time on the work only the room can do, which is making the commitments. The value is that the AI does the setup and the checking; the danger, the one the rest of the lesson guards against, is mistaking the staged, checked board for the pull plan, when it is only the candidate the room turns into a pull plan by promising against it.

The Session Is Where Promises Are Made, Not Where a Plan Is Handed Down

Here is the heart of the lesson. The pull-planning session exists to produce commitments, and a session that walks the room through an AI-staged plan and collects nods has not produced commitments, it has handed down a plan, which is the exact failure pull planning was invented to cure. A nod to a plan someone else drew is not a promise; a promise is a foreman saying, in their own words, what they will do and by when, having reasoned about whether they can actually do it, in front of the trades who will depend on it. The difference is not ceremony, it is ownership: the foreman who reasoned to the date and said it out loud owns it, and the foreman who nodded at the AI's date does not, so the nodded plan slips for the same reason the handed-down CPM slips, because nobody promised it.

This means the facilitator must use the AI-staged board as a provocation, not a conclusion. The staged plan is a strong first draft that gives the room something concrete to react to, which is truly useful because reacting to a draft is faster than building from nothing, but the facilitator's job is to make the trades own each handoff, to push past the nod to the real promise: can you actually hand the framing to the painter by Thursday, what do you need to make that true, what are you promising. The candidate's dates are the AI's proposal; the plan's dates are the trades' promises, and the session is the act of converting the first into the second, which happens only if the facilitator refuses to let the room ratify the draft and instead makes each trade commit in its own voice.

The recovery lesson set this up exactly: an AI-drafted recovery is a proposal that becomes a recovery only when the trades commit to its levers in a session like this one, and this lesson is that session. The pull plan is where the proposals, the recovery's crashed crews, the master schedule's sequence, the AI's staged handoffs, become commitments the trades actually make, with their constraints surfaced and resolved. A pull plan whose handoffs were ratified rather than promised is the drawn recovery without commitments behind it, the fiction that fails twice, and the facilitator's central discipline is to ensure the room leaves with promises it made, not a plan it approved.

The AI Flags the Constraint, the Room Resolves It

The real-time conflict flagging is the AI's most useful trick in the session, and it has its own version of the same boundary. When the tool flags that a compressed duration violates a cure time, or that two trades collide in a work area, or that an inspection cannot happen as early as the plan now wants, it has surfaced a constraint, and surfacing it early, while the room can still act, is the value. But flagging a constraint is not resolving it, and the resolution is the room's work, because resolving a constraint means a trade agreeing to do something different, start later, sequence around the collision, find the labor to hold the compressed duration, and that agreement is a commitment only the trade can make.

So the flag directs the room's attention exactly as a metric-as-signal directs the competent person's attention behind the life-safety gate: it says here is a conflict, look here, and the room looks and decides. The facilitator brings the flagged conflict to the trades involved, and they work out the resolution, the painter agreeing to start a day later, the two trades agreeing to split the area by week, the framer committing to the second crew that holds the compressed duration, and the resolution is a new set of promises that replace the conflict. The AI cannot make those promises any more than it can make the original ones, so its role ends at surfacing the conflict and begins again at checking the room's resolution for new conflicts, in the loop that makes the session productive: the AI flags, the room resolves and re-promises, the AI re-checks.

This loop is where the AI earns its place in the session, because it lets the room iterate fast, trying a resolution and immediately seeing whether it creates a new conflict, which a human facilitator scanning the wall could not do at the speed of the conversation. But the loop only works because the room is making real promises at each step, not approving the AI's adjustments, so the facilitator keeps the loop honest by ensuring each resolution is a commitment a trade actually made, not a move the AI suggested and the room let stand. The AI flags and re-checks at machine speed; the room resolves by promising, and the productive session is the two working together with the line between them held.

PPC Measures Promise Reliability, It Does Not Make the Promise

Percent Plan Complete is the Last Planner System's core metric, the fraction of the week's promises that were kept, and AI makes it easy to track: the tool knows what was promised and can record what was delivered, computing the PPC and trending it, and even analyzing the reasons promises failed, the variances, to surface patterns in why the plan breaks down. This is useful, and it is metric-as-signal in the form the program has taught throughout: the PPC is a signal of how reliable the promising is, a number that directs attention to where reliability is low, not a verdict that the plan is good or a substitute for the promising itself.

The discipline is to read the PPC as a measure of the promising, not as the promising. A high PPC means the trades are making promises they keep, which is the goal, but it is a result of good promising in the session, not a thing the AI can produce by tracking; a low PPC signals that promises are failing, which directs the room to ask why, were the promises unreliable, the constraints unresolved, the commitments not real, but the fix is better promising in the next session, not better tracking. So the PPC tells the room how its promising is going and where it is failing, and the room responds by promising better, which keeps the metric in its proper place as a signal that informs the human work rather than a number that replaces it.

The variance analysis the AI can offer is the same: when it surfaces that a particular trade's promises fail most often, or that constraints of a particular type keep breaking the plan, it has directed attention to a pattern, and the room uses the pattern to promise better, getting the unreliable trade's constraints surfaced earlier, building the recurring constraint into the make-ready work before the session. The AI's analysis finds the pattern; the room's response is to improve the promising, which is the human work the pattern informs, so the PPC and its variance analysis serve the promising discipline rather than substituting for it, the metric-as-signal boundary holding here exactly as it does behind the other gates.

The Applied Problem: Facilitate a 90-Minute AI-Augmented Pull-Plan Session

Here is the exercise. Design and facilitate a ninety-minute pull-planning session for a real upcoming milestone, using AI to stage the candidates and flag the conflicts while the room makes the promises. Before the session, have the tool stage the candidate pull plan from the master CPM, the handoffs laid out and dated, the obvious conflicts pre-flagged, so the room starts from a populated board. In the session, run the discipline that converts the candidate into a pull plan: walk the milestone backward, and at each handoff push the responsible trade past the nod to the real promise, what they will hand off, by when, what they need to make it true, surfacing the constraints, and use the AI's real-time flagging to catch the conflicts the room's moves create and bring each one to the trades to resolve by re-promising.

Produce two things. First, the session design: the AI's pre-session staging and in-session conflict flagging, the facilitator's discipline for making each handoff a real promise rather than a ratified draft, the constraint-resolution loop where the AI flags and the room resolves by re-promising, and the close where each trade leaves with an explicit commitment and the week's promises are captured for PPC tracking. Second, the commitment analysis: why a pull plan is a chain of promises and not a schedule, why the AI's staged board is a candidate the room turns into a plan by promising against it rather than a plan the room ratifies, why the constraint flag is a signal the room resolves rather than a resolution the AI supplies, and why the PPC measures the promising without making it, with the rule that a session which collects nods on the AI's plan has handed down a plan and not produced commitments.

The deliverable is the session design and the commitment analysis, and the lasting product is a pull-planning practice that uses the AI to spend the room's ninety minutes on promising instead of transcription and conflict-hunting, so the room starts from a populated, conflict-aware board and leaves with a real promise from every trade. The professional who masters this gets the AI's speed in staging and checking, which is real, without the failure of a ratified draft that nobody promised, because the session design places the staging and the flagging with the AI, where its speed is the value, and the promising and the constraint resolution with the room, where the commitment lives, so the pull plan the project runs on is a chain of promises the trades actually made, which is the only kind of pull plan that holds.

Key Takeaways

• The Last Planner System rests on the insight that the people who do the work should plan it as reliable promises to one another, and pull planning is the technique: the trades start from a target milestone and work backward, each promising what it will hand to the next, so every activity is a commitment a foreman made out loud to the people depending on it.
• Pull planning works because people deliver on commitments they made, not on plans handed to them (the lesson the recovery chapter set up and this one completes): the foreman who promised the date in front of the painter owns it, where the foreman who nodded at the CPM's date does not, and Percent Plan Complete (PPC) measures this promise reliability directly.
• The AI stages the candidate pull plan from the master CPM before the session (a tool like Touchplan reads the logic network and lays out the dated handoffs) and flags constraint conflicts in real time during the session (a compressed duration violating a cure time, two trades colliding in a work area, an inspection pulled earlier than its lead time allows), both mechanical accelerants on the right side of the line.
• The danger is mistaking the staged, conflict-checked board for the pull plan: it is only the candidate the room turns into a pull plan by promising against it, so a session that walks the room through the AI's plan and collects nods has handed down a plan, the exact failure pull planning was invented to cure.
• The facilitator uses the staged board as a provocation, not a conclusion, pushing each trade past the nod to the real promise (can you actually hand this off by then, what do you need to make it true, what are you promising), because the candidate's dates are the AI's proposal and the plan's dates are the trades' promises, and the session is the act of converting the first into the second.
• Real-time conflict flagging has the same boundary: the AI surfaces a constraint (the value is surfacing it early, while the room can act), but resolving it means a trade agreeing to do something different, a commitment only the trade can make, so the loop is AI flags, room resolves by re-promising, AI re-checks, at machine speed with the line held.
• PPC is metric-as-signal: it measures how reliable the promising is and directs attention to where reliability is low, but a high PPC results from good promising in the session, not from better tracking, and a low PPC is fixed by better promising next time, not by better measurement, so the metric and its variance analysis serve the promising discipline rather than substituting for it.
• The artifact: design and facilitate a 90-minute AI-augmented pull-plan session (AI stages and flags, the room promises and resolves, each trade leaves with an explicit commitment captured for PPC), and analyze why a pull plan is a chain of promises the room makes against the candidate rather than a plan it ratifies, because a pull plan whose handoffs were ratified rather than promised is the drawn recovery without commitments, the fiction that fails twice.