AI Warehouse Layout Generator: What RackCity's AI Does, and What It Doesn't

We build RackCity, and its AI is called Forky. "AI warehouse layout generator" is a phrase that could mean anything from a chatbot that talks about warehouses to a picture generator that draws something warehouse-shaped, so this page is specific about what ours is: a way of writing a real, editable rack design from a description or a document, with the geometry and the counting done by the same layout engine that runs when you draw by hand.

What the AI Does

Three things, all from the same chat dock at the bottom of the canvas:

• Drafts a layout from a sentence. Building dimensions, clear height, where the docks are, what racking you want, in whatever units you think in. It writes the building, the zones and the rack spec, and the canvas draws them.
• Drafts a layout from a PDF flyer. Upload the landlord's marketing sheet, tick the page with the plan, and it reads the dimensions, the docks and the column grid off the page and drafts the racking in one pass.
• Edits the design you have. "Tighten the aisles to 9 ft." "Make the back zone double-deep." "Add a 30 × 40 office in the north-west corner." It changes the design on the canvas, not a copy of it, and it remembers the flyer you uploaded two turns ago.

How It Works, and Why the Count Is Exact

This is the part worth understanding, because it is what separates a layout generator from a picture generator. The AI does not draw the racking and it does not count the pallets. It writes a design: a structured description of the building (dimensions, docks, columns, walls, offices), the zones (where each block of racking goes and what type it is) and the rack spec for each (frame, beam, levels, aisle, pallet). That description is checked against a strict schema before it goes anywhere — a beam length that isn't a number, a zone outside the building, a rack type that doesn't exist, and the answer is rejected and retried rather than drawn.

What passes the check is handed to RackCity's layout engine, the same code that runs when you drag a zone by hand. The engine places the rows at the aisle width in the design, buries the columns, subtracts the dead bays and the tunnels, and counts the positions. So the count on the header after an AI draft is not the AI's estimate. It is the engine's count of the layout the AI described, and it is exactly as reliable as the count on a layout you drew yourself.

The model behind it is a large language model with a long instruction set about racking: the catalog of frame and beam sizes, how pallets fit on beams, what a flue is, how aisles relate to trucks, how to convert metric inputs exactly rather than approximately. It is the same instruction set on every request, which is what makes its output consistent from one building to the next.

From a Sentence

A good prompt has the building and the racking in it. Enough:

"400 × 250 ft warehouse, 30 ft clear, eight docks on the south wall, 30 × 40 ft office in the south-west corner, back-to-back selective racking with 12 ft aisles, 8 ft beams, 4 levels."

Or in metric, which comes back in metric:

"60 m by 40 m, 8 m clear, 2700 mm beams, 4 levels, 3 m aisles, four docks on the south wall."

Leave things out and it fills them with sensible defaults and tells you what it assumed in its reply: a 12 ft aisle, GMA pallets, 28 ft frames if the clear height allows. The reply is a chat bubble in plain language — what it built and what it guessed — and the design is on the canvas underneath it. The Suggested Prompts button in the dock has starting points in both unit systems.

From a Building Flyer

Drop the PDF on the start screen. RackCity shows you the pages; tick the one with the site plan (the reader always needs a page picked) and, if the sheet is ambiguous, type a note: "the 120,000 sq ft east unit", "clear height is 32 ft, the flyer says eave". It reads the building dimensions and shape, the dock doors along a wall, and the column grid where the drawing shows one, and then drafts the racking for the truck you tell it the building will run. Its reply says what it read and what it assumed, so a wrong reading is visible before you trust the count.

It reads a typical commercial flyer — a clean rectangle with dimensions on it — well. It reads a CAD sheet with a hundred annotations less well, and a hand sketch photographed on a phone worse still. If the building has racking drawn on the plan already, tick has rack and it extracts that layout's spec instead of inventing a new one.

Editing by Asking

Most of the value is in the turns after the first. Every edit is a sentence to Forky, applied to the design as it is:

• "Change the aisles to 3.5 m" — every aisle on the canvas becomes 3.5 m, and the leftover floor is drawn as an open band, not spread into the aisles
• "Make the two rows nearest the back wall double-deep"
• "The flyer said 24 ft clear, not 28 — fix the frames"
• "What's the position count if I go to five levels?" — it answers and offers to make the change

Or don't ask: everything the AI wrote is an ordinary object on the canvas. Drag the zone, open the bay, change the beam in the dropdown. The AI and the mouse edit the same design.

What It Gets Wrong

The layout arithmetic is exact; the inputs are where the errors are, and they are worth knowing by name:

• Notched buildings. A footprint with an L or a cut-out corner is approximated as a rectangle with walls and obstructions inside it. Usually close; check the corners.
• Columns the sheet doesn't show. A flyer with no column grid gets a layout with no columns buried. The count will drop when you add the real grid — paste the bay widths off the architect's drawing and the rows re-place.
• Eave height quoted as clear height. Flyers do this constantly. If the sheet says 32 ft and the sprinklers hang at 29, say so in the note.
• Optimism. Given a bare "fill it with racking", it will. Tell it about the staging strip, the office and the pick line, or add them afterwards, before you quote the number.

Treat the draft the way you would a capable junior designer's first pass: the drawing is competent, the reasoning is stated, and you still walk the building before anyone orders steel.

What It Refuses

Forky is scoped to warehouse layout and says so. Ask it for something outside that and it declines with a short message instead of improvising. Two refusals matter in practice:

• Rack types RackCity doesn't model. Ask for a cantilever layout, drive-in, push-back, pallet flow or mobile racking and it tells you plainly that it can't build that yet, rather than drawing selective racking and calling it something else. If the unsupported type is incidental — "mostly selective, with a cantilever bay for pipe" — it builds the selective part and notes the gap.
• Everything that isn't a warehouse. It won't write your emails or debate your politics. A refused request costs you nothing from the day's allowance.

What It Costs

AI requests are metered by the day and by plan: 3 a day on Free, 20 on Solo ($29 a month) and 100 on Warehouse Pro ($99). A request is one draft or one edit. A request that fails to produce a valid design is retried once behind the scenes and, if it still fails, not charged; a refusal is charged, because it was answered. The counter in the dock shows what's left today.

Frequently Asked Questions