Audit Friendly MODERN ACCOUNTING
EXPLAINER

AI FOR ACCOUNTANTS · PART 9 OF 16

What a token is and why it sets the price of every AI task you run

 

It is not the task that drives the cost - it is the volume of text the model has to read and write to complete it.

A token is roughly four characters of text - a fragment of a word, sometimes a whole short word. When you send a document to an AI model, the model does not read it the way a person does. It converts everything - your prompt, the document, its own reply - into tokens, processes them, and bills you for the total count. The input tokens are what you sent; the output tokens are what came back. Both sides of that exchange cost money, and the rates differ.

For accounting work, that mechanic matters more than it does in most fields because the documents are long and the data is dense. A multi-page audit workpaper converted to plain text can run to an enormous token count before you have typed a single instruction. A multi-entity bank reconciliation with hundreds of line items is similar. The moment you paste either of those into a prompt, you have already consumed a meaningful chunk of the model's context window - the ceiling on how much it can hold at once - and you have set the floor on what that task will cost. If the model also needs to produce a detailed commentary or a structured output, the output token count adds to the bill.

Some accounting tasks have a naturally low token footprint. Classifying a single transaction, drafting a one-paragraph client email, or checking whether one account description matches a chart-of-accounts category - these stay small because both the input and the expected output are short. The math is different for tasks where the source material is inherently large: reading a full general ledger to spot anomalies, summarizing a thick client file, or extracting data from a dense tax return. Those tasks are not necessarily bad uses of AI, but the cost is proportional to the page count, and 'summarize this' is not a free instruction.

Context window size is the related constraint. Models have a maximum number of tokens they can process in one call. If your document exceeds that limit, you either truncate it - and risk losing the figures that matter - or you split it into chunks and run multiple calls, which multiplies cost and introduces the risk that something important falls in a gap between chunks. A reconciliation that spans two chunks, for instance, might have its opening balance in one call and its closing entry in another, and the model in each call sees only half the picture.

The practical adjustment is to be deliberate about what you paste. Extracting only the relevant columns from a reconciliation before passing it to an AI call, or summarizing a workpaper section manually before asking the model to analyze it, reduces input tokens without reducing output quality - often improves it, because a cleaner input produces a more focused response. It also helps to match task size to model tier: a small, inexpensive model handles transaction classification well; a larger, more expensive one earns its cost only when the reasoning task is genuinely complex. The question to ask before running a workflow is not 'can the AI do this' but 'how many tokens does this require, and does the output justify them.'

WORKED EXAMPLE

In practice

A senior accountant is preparing a month-end bank reconciliation workpaper for a mid-size client. The reconciliation covers this month's activity, exported as a plain-text CSV, and she wants an AI assistant to flag any items older than 30 days that remain uncleared.

THE PROMPT
Below is a plain-text CSV of our bank reconciliation for the period just ended. Each row has: date, description, amount, cleared status, days outstanding. Identify every row where cleared status is 'N' and days outstanding is greater than 30. List them as: date | description | amount | days outstanding. Do not include any cleared items. Do not summarize; return only the matching rows. [paste CSV here]

What came back. The model returned a filtered list of uncleared items over 30 days, formatted as requested. A handful of the rows were duplicates that had already been flagged as timing differences in a separate note - the model had no way to know that from the CSV alone and listed them anyway.

How it was checked. The accountant compared the model's output list against the 'timing differences' tab in the workpaper and removed the items already documented, then agreed the remainder to the aging schedule.

A constructed example. The prompt is usable as written; the figures show the shape of a result, not a measured one.

WHEN TO USE IT

 
Filtering or classifying structured data where input rows are short and output is a list.
 
Drafting client-facing summaries from a manually condensed set of notes.
 
Checking whether a single account description matches a defined chart-of-accounts category.
 
Running analysis on a document you have already trimmed to the relevant columns or sections.

WHEN NOT TO

 
Pasting a full general ledger or multi-entity consolidation without any pre-filtering.
 
Expecting one AI call to read a document that exceeds the model's context window limit.
 
Treating AI output as final when source data was split across multiple calls.
The pitfall: The model silently loses context when a document is split across calls - a reconciliation's opening balance and closing entry may land in separate chunks, and neither call knows what the other saw.

WHAT TO TAKE FROM THIS

 
Check the token count of your largest recurring documents before committing to an AI workflow.
 
Tasks that require the model to read a full document cost far more than tasks with short, structured inputs.
 
Splitting large files across multiple calls saves on context limits but adds cost and creates gap risk.
FREE, EVERY WEEK
Get the next explainer in your inbox
Tuesdays: one Build with AI workflow with the prompt, the steps and the verification rule. Thursdays: the explainer and the week's tools.

SPONSORED

Gusto: full-service payroll that posts clean journal entries to your ledger
Federal, state and local filing in all 50 states, native two-way sync to QuickBooks Online and Xero, and published per-employee pricing with no contract. Audit Friendly scored it 74/100 in a fact-checked review; we earn a referral fee if you sign up, and the score is not affected.

QUESTIONS THIS ANSWERS

What is a token in AI billing?

Roughly four characters of text. Models count every token they receive as input and every token they produce as output, and vendors charge for both.

Why do large accounting documents cost more to process with AI?

Because the full text of the document - every line item, heading and label - is converted to tokens before the model begins working. A long workpaper or reconciliation sets a high input-token floor before any analysis begins.

What happens when a document exceeds the context window?

The model cannot process what it cannot see. You must either truncate the document, risking lost data, or split it into multiple calls, which multiplies cost and can leave critical figures in the gap between chunks.

SOURCES

Where this comes from

Read the insights wire →

What the accounting job market is actually asking for.

GO DEEPER

Go deeper

Skills that pay → Which named tools move the number
Software in postings → Every tool ranked by live demand
AF Stack Designer → Design your accounting stack in ten minutes
The insights wire → Cross-dataset findings, refreshed hourly
Accounting and finance job board → The Audit Friendly job board, every posting verified live

IN THIS SERIES

Previously: Why a model invents a number and how to stop it

Next: Why structured output beats a paragraph in accounting work (coming)

An explainer, not a study: it carries no statistics on purpose. Examples are illustrative.

How accountants are using AI, automation and smarter workflows to close faster, audit cleaner, and free up time for real work.

Accounting Stack · Audit Friendly Data · Accounting & Finance Jobs

Audit Friendly · modernaccounting.ai