Booth services order extraction turns a stack of pdf forms into a schedule

Operations wants a single answer three days before move-in: which stands need power connected before eight on Tuesday morning. The answer exists. It is spread across 960 signed PDFs sitting in a shared drive, in four different form layouts from two contractors and a venue, and the only way anybody has ever produced it is by opening them.

Booth services order extraction is the step that turns that pile into a table you can sort. Of the five exhibitor documents an organiser routinely collects, this is the one with no standard behind it, and that single fact drives every design decision below.

Four fields make the schedule

Extraction projects fail by trying to capture everything. Start from what somebody will act on.

Line item. What was ordered, normalised to your own vocabulary. The forms will say 1000 watt outlet, 10 amp 120V, and 1000W/120V for the same thing.

Quantity. How many, which sounds trivial and is the field most often lost, because quantity boxes on these forms sit in narrow columns next to two price columns and an empty total column.

Service date. When it is needed, which on most forms is implied by the move-in schedule rather than written, and has to be derived from the show calendar for anything that does not carry an explicit date.

Booth number. Where it goes, and the only field that joins this document to everything else you hold.

Price is tempting and mostly a trap. The contractor's system is authoritative for money, your extracted figure will disagree with it at the margins, and reconciling two price columns you do not control will consume more time than the schedule is worth. Extract price if you have a specific reconciliation you intend to run, and leave it out otherwise, on the same principle that governs any field list built for a downstream consumer.

Why is this document harder than the others?

Compare it with what sits either side of it in the same folder.

Form W-9 has one layout, issued by one authority, revised occasionally, with numbered lines in fixed positions. An ACORD certificate of liability insurance is a standard form with boxes that appear in the same place on every one you will ever see. Both are extraction problems with a known template.

A booth services order has no such anchor. Two published 2026 examples make the point. The Edlen Electrical Exhibition Services forms for the PAA Annual Meeting 2026 lay outlets out as rows with two quantity columns, an advance price column, a regular price column and a total column, grouped by voltage and phase. The Superior Expo Services furniture forms for the National College Fair Houston 2026 use a two-panel layout with tick boxes, item descriptions, a discount price, a standard price and a total, with two independent item tables side by side on one page.

Same document type, same purpose, completely different geometry. And both are reissued each edition with the dates and prices changed, sometimes with the layout adjusted.

Microsoft's guidance on its document models, updated in 2026, describes exactly this failure mode. Document type confidence indicates how closely an analysed document resembles the training data, and a low score points to template or structural variation in the document being analysed. Their remedy is to label a document carrying that variation and add it to the training set, and their general advice is to separate visually distinct document types into different models rather than forcing one model to cover them.

Train per layout and accept the consequence

The consequence is that your extraction estate has a maintenance schedule rather than a launch date.

Two contractors across a portfolio, at three or four form types each, is six to eight layouts. Microsoft's guidance is to include at least five samples of each type in the training data, so a new layout is a small labelling job rather than a project, and the same guidance suggests targeting an estimated accuracy above 80 per cent, with something close to 100 for more sensitive records.

The honest reading of that is a split policy. Booth number and quantity are consequential enough to want the high end, because a wrong booth number sends a crew to the wrong stand and a wrong quantity leaves an exhibitor short. Line item text can live at the lower end, because a human reading the schedule will spot a mislabelled item.

Where an organiser goes wrong is treating the annual reissue as a surprise. Put a calendar entry against each contractor for the week their forms are republished, run twenty documents through, and look at the document type confidence. A drop there is your notice that the layout moved, and it arrives weeks before anybody submits an order at scale.

The normalisation layer deserves its own mention, because it is where the real ongoing work sits and it has nothing to do with the model. Extracted line item text has to map onto a vocabulary you control before the schedule means anything, and that mapping is per contractor, per form, per edition. A 20 amp 208 volt single phase service on one sheet and a 208V/20A drop on another are the same connection needing the same crew. Hold that mapping in a table a coordinator can edit, keep an unmapped bucket that somebody reviews weekly, and resist the urge to make the model do the normalising. Models are good at finding the string on the page. A lookup table is better at deciding what the string means, and a lookup table can be corrected by the person who noticed the problem.

Reconciling booth number against the contract

This is the step that pays for the whole exercise, and it is arithmetic rather than machine learning.

Take 400 exhibitors submitting an average of 2.4 order forms each, which is 960 documents. Extract booth number from every one and join it to the contracted booth list.

On a floor that has been renumbered after a hall change, expect a handful of failures. Suppose 9 booth numbers on the extracted orders do not exist on the current plan, and 14 exist but belong to a different exhibitor than the one who signed the form. Those 23 documents are the entire operational risk in the pile, and they are 2.4 per cent of it.

Finding 23 problems in 960 documents by reading is a two day job that nobody does. Finding them by joining two columns is a query. That is the argument for extraction stated at its narrowest, and it holds even if every other field comes back imperfect.

The mismatches sort into three causes, and the sort matters because the fixes differ. An exhibitor writing last year's booth number is the most common and needs an email. A number that changed after a floor plan revision needs the plan version recorded against the order. A co-exhibitor ordering under the host's number is a different problem entirely, and untangling who counts as the exhibitor at that stand is its own definitional argument before any order can be attributed correctly.

What does the extracted schedule actually give you?

Three things, in ascending order of value.

A stand-by-stand service list, which is what the move-in supervisor wants on a clipboard. Sorted by hall, aisle and stand, with the four extracted fields, printable.

A time-ordered work sequence, which is what the contractor's foreman wants. Group by service date and service type, and the crews can be planned against the actual distribution of work rather than against an assumption that everything happens on day one.

And a set of exceptions worth acting on before anybody arrives. Stands with no order of any kind, which on a show with mandatory power is a phone call. Stands ordering power on a form that predates their upgrade to a larger space. Stands whose orders total more than the space can physically take.

That third output is the one that changes the week. It exists only because the data became queryable, and it is invisible in a folder of PDFs no matter how carefully anybody reads them.

Where this stops

The extracted schedule is a snapshot of what was ordered in advance, and show floors do not respect snapshots. Orders change on site, sometimes substantially, and the contractor's system will hold the truth by Tuesday afternoon while your extracted table still holds Friday's version. Publish it with a timestamp and stop using it once move-in starts.

The authoritative source is also somebody else's. The contractor owns the order, the money and the fulfilment, and your extraction is a read-only shadow of it built for planning. Any process that starts treating the extracted table as the record of what an exhibitor is owed will eventually put your team in an argument it cannot win.

There is a coverage limit too. Extraction only sees the forms that reach you, and on many shows exhibitors order directly with the contractor through a portal you never touch. If half the orders never pass through your document store, the schedule you build is half a schedule, and it will be systematically missing the exhibitors who are most comfortable ordering online. Check your coverage rate before you trust a distribution.

Field level confidence reading, error scoring per field and classification accuracy all sit alongside this and have their own treatment in N26, N25 and N28. This post assumes the document arriving is the one you think it is.

Start with twenty documents and a spreadsheet. Pick the twenty most recent booth services orders in your store, type the four fields out by hand, and join the booth numbers to your contracted stand list. Count the mismatches. If there are more than one or two, you have found the reconciliation that justifies building anything at all, and the exhibitors behind them are worth cross-checking against your portal activity records to see whether they have been struggling with the onboarding process more broadly.