The Hidden Cost of Tickmarks: Why Audit Annotation Is Ready for AI

Ask an experienced auditor what they spent last Tuesday doing and you may hear something like this. Opened 40 invoices. Highlighted the approval signature on each one. Added a tickmark. Updated the legend. Cross referenced each invoice number back to the population listing. Noted the exceptions. Saved the file. Moved to the next control.

Eight hours of work. Zero analytical insight produced.

This is the annotation layer of audit testing, and it is one of the largest and least discussed consumers of audit hours. It is also the part of the job most ready to be handed to a machine.

What annotation actually involves

Annotation is everything that turns raw evidence into a defensible workpaper. It includes marking what was examined and what was found, applying tickmarks and maintaining the legend that explains them, cross referencing items back to a population or a trial balance, labeling and indexing evidence files, noting the date and source of every document pulled, and writing the brief explanatory comments that let a reviewer follow the preparer’s path.

None of it requires professional judgment. All of it is required.

That combination is exactly what makes it so expensive. The work cannot be skipped, because a workpaper without annotation is not auditable. But it also cannot be delegated to a junior resource without creating quality problems, because inconsistent annotation generates review notes, and review notes generate rework.

Why it consumes so much time

The volume compounds in ways that are easy to underestimate.

A single control might have 25 sampled items. Each item might require three to five annotated attributes. That is roughly 100 annotation actions for one control. Multiply by the number of key controls in scope, then by the number of testing cycles per year, then by the number of entities in a multi-location environment. Mid-sized companies routinely cross into tens of thousands of annotation actions per year.

Format variation makes it worse. An approval might appear as a signature on a scanned PDF, an email thread, a workflow status in a ticketing system, or an initial in the corner of a photographed document. A human has to look at each one, find the relevant element, and mark it. There is no shortcut.

And it happens under deadline pressure. Annotation tends to land at the end of fieldwork, when the calendar is tightest and fatigue is highest. That is precisely when consistency drops.

The downstream cost nobody budgets for

Here is the part that makes annotation more expensive than the hours suggest.

Most review notes are annotation notes.

Walk through a typical review cycle and look at what the reviewer actually wrote. Tickmark used but not defined in the legend. Cross reference points to the wrong schedule. Evidence file not labeled with the date pulled. Attribute tested but not documented. Conclusion references an item that does not appear in the supporting detail. Inconsistent tickmark usage between two sections of the same workpaper.

These are not findings about risk. They are findings about documentation hygiene. Yet each one triggers the same cycle. The reviewer writes the note. The preparer stops other work to clear it. The reviewer re-reviews. If the clearance is incomplete, the loop runs again.

A workpaper that takes three review rounds to clear often does so because of accumulated small annotation issues, not because of a disagreement about the testing conclusion. The actual conclusion may have been correct from round one.

So the true cost of annotation is the preparer hours, plus the reviewer hours spent catching annotation problems, plus the preparer hours spent fixing them, plus the reviewer hours spent confirming the fixes, plus the schedule slippage that pushes everything toward the deadline.

Why AI is a natural fit

Annotation has three properties that make it unusually well suited to AI.

It is rule based. What makes a valid tickmark, what a cross reference must point to, how evidence should be labeled, these are defined by methodology, not judgment. The rules can be stated explicitly.

It is repetitive. The same actions repeat across thousands of items with minor variation. That repetition is exactly what humans do poorly over long sessions and what machines do identically every time.

It is pattern recognition on documents. Finding an approval signature, locating an invoice number, identifying a date field, matching a vendor name across two documents. Modern models handle this across varied formats without needing a template for each one.

An AI system reading a scanned invoice can locate the approval, mark it, extract the invoice number, match it to the population listing, apply the correct tickmark from the legend, label the file with its source and retrieval date, and record the attribute result. Then repeat the process identically 10,000 times.

The consistency benefit matters more than the speed benefit

Speed is the obvious gain. Consistency is the more valuable one.

When three different preparers annotate three sections of the same testing file, you get three slightly different approaches. Different comment styles. Different levels of detail. Different interpretations of the same tickmark. A reviewer has to normalize all of it mentally before evaluating anything.

A machine applies one standard everywhere. The same tickmark means the same thing in every workpaper. Cross references follow the same convention. Evidence labeling follows the same format. Every attribute gets documented, because the system does not get tired at 9pm on the day before the deadline.

That consistency is what removes review notes before a reviewer opens the file. The completeness and hygiene issues that generate most notes simply do not occur.

What the reviewer gets instead

This changes the nature of review in a useful way.

A reviewer opening a machine annotated workpaper is not hunting for missing cross references. The annotation is complete and consistent by construction. What remains is the question that actually deserves senior attention. Does the evidence support the conclusion, and is the conclusion right?

That is a better use of experience. It is also a faster review, because the reviewer is reading for judgment rather than scanning for omissions.

What still requires a human

AI annotation does not mean unreviewed annotation.

Someone has to define the methodology the system follows, including tickmark conventions, cross reference standards, and documentation requirements. Someone has to validate that the system applies them correctly, ideally by running parallel manual annotation for a cycle and investigating every discrepancy. Someone has to review exceptions, because an AI that cannot find an approval signature may be looking at a genuinely missing approval or may be looking at a document it misread.

And someone has to own the conclusion. That never moves.

A reasonable starting point

Annotation is a good first AI use case precisely because it is low risk. It does not change your testing methodology. It does not alter sample selection or the attributes you test. It does not reach the conclusion. It changes who performs the mechanical documentation of work that is being done the same way it always was.

That makes it easy to validate and easy to explain to an external auditor. Run it in parallel for one cycle. Compare the machine annotation to the manual annotation item by item. Document the comparison. Then rely on it.

The payoff shows up in two places. Testing finishes faster, and review rounds drop.

The broader point

Every profession accumulates work that was necessary when it started and merely habitual once technology moved. Audit annotation is in that category. It exists because workpapers must be auditable, and for a long time the only way to make them auditable was a person with a highlighter and a legend.

That is no longer true. The requirement has not changed. The method for meeting it has.

Auditors were trained to evaluate risk and exercise judgment. Spending a third of the testing cycle on tickmarks was never the point.

© 2026 Audagic. All rights reserved.