# AI-Assisted Capital-Plan Sequencing

**Version:** 2.0
**Date:** 2026-08-20
**Supersedes:** v1.0 (2026-08-16), which described the method with no plan attached.
**Type:** Method, with the property improvement plan it was run on

---

## Why this version has a building in it

Version 1.0 said a brand's Property Improvement Plan arrives as scope and never as
a sequence, and described turning one into the other. That is true and it is
abstract. What it left out is the part that decides whether the schedule is worth
anything: **almost every duration in a capital plan is an estimate, and estimates
compound.** A twelve-phase sequence built on twelve guesses is a guess with a
Gantt chart drawn on it.

This version runs the method on a real plan — **Westmark Fairbanks**, 400 rooms,
against a PIP issued April 2025. It is written around the one thing that turned
out to matter most, which no methodology document predicted: the property had
already finished one floor, and that floor was the only measurement in the
building.

---

## 1. The three inputs

Everything the sequencing needs already exists at the property:

1. **The scope document — the PIP itself.** What must be done, to which spaces, to
   what standard.
2. **The labour picture.** Which trades exist in-house, their true capacity, and
   when the labour year swells and shrinks — including any collective-agreement
   rules about seasonal movement of staff.
3. **The demand calendar.** Daily occupancy and revenue history, at least one full
   year. This is the input most properties have and least use.

If any of the three is missing, build it first. A sequence computed on a guessed
demand calendar is a guess wearing a schedule.

## 2. The constraint that shapes everything

At this property the demand calendar does not merely inform the sequence, it
deletes most of the year from it.

**Rooms cannot come out of service between 1 May and 30 September.** The
property's commercial commitments over the summer are large enough that the
inventory is spoken for; the specific terms are confidential and are not the point
here. The operational consequence is the point:

> **Roughly seven working months a year, for 400 rooms, and the PIP wants most of
> them touched.**

Add the rooms already out of order for other reasons and the working inventory
shrinks again before any phase starts.

**The generalisable rule: find the blackout before you sequence, not after.** A
solver handed twelve months will produce a beautiful plan that cannot be executed.
The blackout is not a complication to the model — it is most of the model.

## 3. The method

**Step 1 — extract the scope into work units.** AI reads the PIP and returns a
structured register: one row per unit of work, with the space it applies to, the
trade it needs, and whether it is a per-room repetition or a one-off. This is the
step that turns prose into something a solver can hold.

The register immediately sorts the work into two kinds, and they schedule
completely differently:

| | Example from this plan | Character |
|---|---|---|
| **Ball rollers** | switch, outlet and plate swap — on the order of 3–4,000 units | easy per unit, enormous in aggregate, parallelisable |
| **Long campaigns** | 200 shower conversions — demo, plumbing, tile, glass | hard per unit, sequential, trade-constrained |

Mixing them in one average is how capital plans become fiction.

**Step 2 — batch into phases.** Rooms come out of service in batches, not
individually. The batch size and the maximum rooms offline at once are business
decisions, not engineering ones: they trade renovation speed against sellable
inventory. On this plan the model was run at **twelve phases, batches of 33 rooms,
21 days per batch, and a hard ceiling of 72 rooms offline at any moment.**

**Step 3 — solve against the real calendar.** This is a constraint problem, not a
spreadsheet: minimise displaced revenue subject to the blackout, the offline
ceiling, trade availability and phase durations. A CP-SAT solver does it in
seconds and, more importantly, will tell you when there is **no** feasible
schedule — which is itself the answer, and one a spreadsheet will never give you.

**Step 4 — price each phase in displaced revenue,** not in rooms. A room closed in
February and a room closed in June are not the same cost, and the demand calendar
already knows the difference.

## 4. What actually decided this plan

The methodology above is sound and it is not what mattered most.

**One floor was already finished.** Floor 8 — 35 rooms, 8.75% of the building —
had been completed before the sequencing work began. The owner's own description
of it was *"we rushed it though."*

That single fact outranks the rest of the document, for one reason:

> **Every duration in a capital plan is a planning estimate. A completed floor is
> a measurement.**

Walking floor 8 against the PIP list, item by item, is an afternoon's work. It
answers the question the entire twelve-phase schedule is built on — *how many
calendar days from start to first room back in service?* — with an observation
instead of an assumption. Everything downstream scales off that number.

**It also answers a question nobody wants to ask.** If the floor was completed
before or without the approved PIP in hand, some of it may not count, and it may
need doing twice. That is a bad answer, and it is far cheaper to have it in
August than in the middle of phase 6.

**The rule: before sequencing anything, find the work already done and measure it.
Calibrate the model on your own building before you trust it about the rest of
your building.** Do this before phase zero. It outranks everything else in the
method.

## 5. What this method cannot do

- **It cannot price the work.** It sequences scope; the costs come from the
  contractor and the costbook.
- **It cannot see condition.** A PIP lists what the brand requires, not what the
  building actually needs, and the two diverge.
- **It is only as good as the demand calendar.** Where the occupancy history is
  back-filled or synthetic, say so in the output, because the phase pricing
  inherits that uncertainty entirely.
- **The offline ceiling is a business ruling, not an optimisation.** The model will
  happily go faster by closing more rooms. Whether that is acceptable is not a
  question the solver is allowed to answer.

## 6. Transfer

The method carries to any property with a scope document, a roster and a demand
calendar. What does not carry is the shape of the constraint. This property's
blackout is a summer one driven by tour and group business; a city hotel's may be
midweek, a resort's may be inverted.

For a property that has not opened, there is no demand calendar to solve against
and this method does not apply yet. What applies instead is the discipline from
section 4, in advance: **decide now which work will be measured rather than
estimated, and make sure the first batch is instrumented so it becomes the
calibration for everything after it.**

---

*Version 2.0. Superseded versions stay listed rather than disappearing. Figures
are operational; no term from any commercial agreement appears in this document,
and no guest or staff information appears in it.*