Skip to content

Straight through processing rate is the onboarding automation number worth reporting

OnboardingUpdated 2026-08-188 min read

In short

Straight through processing rate is the share of documents a pipeline accepts with zero human edits, counted separately for each document type and reported next to the error rate on the accepted pile. A blended rate across all types moves when the document mix moves, which makes it useless for judging whether anything improved.

The board pack says the onboarding pipeline is 90 per cent automated. Nobody in the room asked what the denominator was, and it turns out to mean that 90 per cent of documents reached the extraction step without failing, which every document does, because reaching a step is easy.

Straight through processing rate is the version of that claim you can defend. Count documents received of one type in a period. Count how many were accepted with zero human edits. Divide. The number that comes out is usually much lower than whatever was on the slide, and it is the only automation figure that corresponds to somebody's Tuesday afternoon.

What counts as straight through?

Zero human edits, and zero human confirmations.

The edits part is uncontroversial. If a coordinator changed a value, the document was touched. The confirmation part is where most reported rates get inflated. A pipeline that extracts everything and then shows each document to a person for a single approving click has automated the typing and left the reading, and the reading is most of the cost. Counting that click as automatic makes the rate go up and does nothing for the queue.

Rejections need a rule too. A document that failed classification and went back to the exhibitor never got processed at all, so it belongs in the denominator and out of the numerator. Excluding rejections from the denominator lets a pipeline report a high rate by refusing anything difficult, which is the failure mode you should assume is happening until you check.

Write the definition down in one sentence, put it in the footer of the report, and never change it mid-year.

Report it per document type or it means nothing

Exhibitor onboarding runs on a small set of document types, typically the exhibitor contract, the certificate of insurance, the IRS W-9, the booth services order and the exhibitor profile form. They differ in structure more than any two of them differ from a receipt.

A W-9 is a fixed federal form with the same boxes in the same places every time. A certificate of liability insurance follows an industry layout with a free text remarks box that carries the clause you most need. A booth services order is whatever the general services contractor designed, and it changes between venues, between years and sometimes between halls.

One rate across all five describes none of them. Worse, it is the rate a team will try to improve, and improving a blended rate usually sends the effort at whichever type has the most volume, when the gain is sitting somewhere else entirely.

There is a second cut worth taking, which is by source. The same certificate of insurance arriving as a broker generated pdf and as a photograph of a printed page taken in a car park are two different extraction problems wearing one document type. Splitting the rate by intake channel, even crudely into born-digital and scanned, usually explains more of the variation than anything about the model, and it points at a fix nobody needs a data scientist for: tell exhibitors to upload the file their broker emailed them.

The worked rate, and what the blend hides

Take one month at a show with figures invented so the division is visible.

Exhibitor contracts: 412 received, 264 accepted with no human edit, which is 64.1 per cent. Certificates of insurance: 388 received, 237 clean, 61.1 per cent. W-9s: 301 received, 226 clean, 75.1 per cent. Booth services orders: 655 received, 144 clean, 22.0 per cent. Exhibitor profile forms: 402 received, 338 clean, 84.1 per cent.

Across the five that is 2,158 documents received and 1,209 accepted untouched, a blended straight through processing rate of 56.0 per cent.

The blend spans 62 points, from 22.0 per cent on booth services orders to 84.1 per cent on profile forms. A single figure of 56 per cent describes no document type in the pipeline and points at no action.

Now attach time. If a touched document costs about four minutes of coordinator attention, the 949 touched documents cost 3,796 minutes, or 63.3 hours a month. Lift booth services orders from 22.0 to 45 per cent and 151 more documents pass untouched, worth 604 minutes or 10.1 hours a month. Lift profile forms from 84.1 to 90 and you gain 24 documents and 1.6 hours. The blended rate cannot tell you which of those two projects to fund. The per type rates make it obvious.

Mix changes the blended rate without anything changing

Booth services orders are 655 of 2,158 documents, which is 30.4 per cent of the month's volume and the worst performing type.

Next month the advance order deadline lands, booth services orders rise to 900, and every other type stays flat. Received becomes 2,403. Clean becomes 1,209 minus 144 plus 198, or 1,263. The blended rate falls to 52.6 per cent.

Nothing about the pipeline changed. A 3.4 point drop appeared because a low performing document type became a larger share of the mix, and somebody will spend a week looking for a regression that never happened. Per type rates are immune to this, and a blended rate should be reported with the mix beside it or not at all.

Why is a high rate on its own a bad sign?

Because the rate is trivially improvable by lowering the acceptance bar, and nothing in the number itself pushes back.

Geifman and El-Yaniv (2017) put this precisely in the selective classification setting: a model that can decline to answer trades coverage for accuracy, and the two have to be quoted together. Their headline result reaches roughly 2 per cent top-5 error on ImageNet with high probability while covering just under 60 per cent of the test set. The 60 per cent is the price of the 2 per cent, and reporting either one alone would misrepresent the system entirely.

Straight through processing rate is coverage at document level. So publish it as a pair. The rate, and the error rate measured on the accepted pile against a held out gold set. A pipeline reporting 56.0 per cent straight through with 0.4 per cent field errors on accepted documents is in better shape than one reporting 90 per cent with an accepted pile nobody has ever sampled.

Microsoft (2026) is direct about where the bar sits for this class of work, advising a target estimated accuracy above 80 per cent and close to 100 per cent for sensitive records such as financial ones, with a human review stage added for critical workflows. Exhibitor insurance limits and tax identification numbers sit in that second category, and the correct straight through processing rate for them may be considerably lower than for a profile form.

That has a consequence worth saying out loud, because it runs against how these projects are usually sold. The right target for a document type is set by what a wrong value costs, and for two of your five types the honest answer is that the rate should stay modest on purpose. A show that pushes its certificate of insurance rate from 61 to 88 per cent by relaxing the additional insured check has not automated anything. It has moved a compliance failure from a queue a coordinator works to a spreadsheet the fire marshal reads.

What to do with a 22 per cent document type

Open thirty touched booth services orders and read what the coordinator changed. The answer is almost never spread evenly.

Usually one of four things dominates. The layout changed and the model has never seen it, which is a training data problem and shows up as low document type confidence. One field is systematically wrong, such as a quantity column read from the wrong row, which is a single fix worth most of the gap. The document arrives as a photograph and the character recognition is guessing, which is an intake problem to solve at the upload screen. Or the acceptance rule is too strict, and coordinators are confirming values that were already right.

That last case is common and cheap to fix, and it is worth checking first because it costs nothing. If 40 per cent of the edits on a type are confirmations where the value was unchanged, your threshold is set below where your data says it should be, and choosing it from a risk coverage curve will recover most of that with no model work at all.

Where this stops

Straight through processing rate says nothing about whether the untouched documents were right. A pipeline can accept everything and report 100 per cent while filling your compliance record with wrong insurance limits, and the rate will look better every month as the errors accumulate. The rate is only interpretable next to a measured error rate, and any report that shows one without the other should be sent back.

It is also a poor measure of value early on. The first two months of a document automation project usually show a low rate and a large improvement in something the rate cannot see, such as documents no longer sitting in an inbox for a week. Time to acceptance is worth tracking alongside, because it moves first.

And the rate can be gamed without anybody deciding to game it. Shortening the field list raises it. Dropping a strict validation raises it. Sending harder documents down a manual path outside the pipeline raises it most of all. Freeze the definition, publish the field list version alongside the rate, and treat a sudden improvement as something to explain before something to celebrate. The routing decisions that produce the touched pile, and how to keep that queue affordable, are covered in human in the loop review.

Take last month's onboarding volume, split it by document type, and count how many documents in each type were saved with no edit at all. If you cannot get that count out of the system, that is the first thing to fix, because every other onboarding automation decision needs it.

Questions people ask about straight through processing rate

How is straight through processing rate calculated?
Count documents received of one type in a period. Count how many of those were accepted into the system with no human edit and no human confirmation step. Divide the second by the first. Do it per document type and per period, and record the period, because a rate quoted without a window is uncheckable.
Does a confirmation click count as a touch?
Yes. If a person had to open the document and press accept, a human was in the loop and the document was not processed straight through. Counting confirmations as automatic is the most common way these rates get inflated, and it hides exactly the queue you were trying to shrink.
What is a good straight through processing rate?
There is no cross-industry figure worth quoting, because the rate depends on your document types, your field list and how strict your acceptance rules are. A rate is only meaningful against your own previous period, on the same document type, with the error rate on the accepted pile published beside it.

Related reading

All onboarding operations articles