|
EXPLAINER |
AI FOR ACCOUNTANTS · PART 1 OF 16
What a model context window means for your workpaper
When a lease file is longer than the window, the model is not reading the whole thing - and it will not tell you what it missed.
Every large language model reads text through a fixed-size opening called a context window. Think of it as the model's working memory: whatever fits inside gets processed together; whatever falls outside simply does not exist for that session. For a short memo or a trial balance export, this rarely matters. For a 400-page lease file, an audit workpaper drawing from dozens of source schedules, or a multi-entity consolidation package, the window is a real constraint - and the model will answer as if the constraint is not there.
The mechanism is worth understanding precisely. A context window is measured in tokens, which loosely correspond to word fragments. A dense financial document - tables, footnotes, defined terms, cross-references - consumes tokens faster than plain prose. When a practitioner pastes or uploads a file that exceeds the window, most tools silently truncate: they keep the beginning, drop the end, and return an answer based on the partial text. Nothing in the output signals that the final sections, which happened to contain the variable-rate provisions and the lessee's purchase options, were never read.
This bites hardest in exactly the situations where practitioners reach for AI help. A lease abstraction across a large retail portfolio. A revenue recognition workpaper that pulls from a master service agreement, several amendments, and side letters. A tax memo that references a long regulatory ruling. In each case, the practitioner's implicit assumption - that the model has read the whole file - may simply be wrong. The model does not hallucinate here in the traditional sense; it answers accurately from what it saw. The problem is the gap between what it saw and what the file actually contains.
Chunking is the standard workaround, and it works, with caveats. The file gets split into sections small enough to fit individually, each section gets processed, and the results get assembled. The catch is that a clause in an early section that modifies a term defined much later in the document will not be connected unless both sections are in the window at the same time. Good chunking strategy for accounting work means splitting on document structure - by lease schedule, by amendment, by entity - rather than by raw page count, so that related provisions travel together.
The practical rule is to match the unit of work to the window, not the other way around. Before running a model over a large file, identify the discrete question: not 'summarize this lease' but 'what are the renewal options and notice periods in Schedule B?' Feed the model the section that contains the answer, confirm the section is complete, and verify the output against the source. When the question genuinely requires the whole document - a full-file consistency check, for instance - break it into overlapping segments and reconcile the outputs manually. The window is a tool specification, not a defect; knowing it changes how you design the task.
On the firm side, Kaufman Rossin named Gautam Anne as its new Chief Information Officer, with a remit that includes AI strategy and automation - a signal that larger firms are formalizing the infrastructure decisions that determine, among other things, how document processing pipelines handle exactly these constraints.
WORKED EXAMPLE
In practice
A lease workpaper for a large retail portfolio. The source file is a single PDF containing the master lease form, property schedules, and several amendments. The practitioner needs to extract renewal option terms for each property for the current month's lease accounting remeasurement.
What came back. The model returned the renewal options with their term lengths, the required notice period, and a no-default condition, with exact clause citations. It flagged that one clause referenced an exhibit not present in the pasted text. The notice period it extracted matched the schedule; a condition in a later amendment that capped rent increases during renewal had not been included in the chunk and was therefore absent from the output.
How it was checked. The practitioner compared the extracted renewal terms line by line against the physical schedule and then ran the relevant amendment separately to confirm the rent cap condition before entering figures into the lease accounting workpaper.
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
| WHEN NOT TO
|
WHAT TO TAKE FROM THIS
| Check whether your tool confirms how much of an uploaded file it actually processed. | |
| Split large files by document structure, not page count, so related clauses stay together. | |
| Verify every model output against the source section, not just the summary. |
SPONSORED
QUESTIONS THIS ANSWERS
What is a context window in plain terms for an accountant?
It is the maximum amount of text a model can read and reason over in a single session. Anything beyond that limit is not available to the model when it generates its answer.
Will the model tell me if it only read part of my file?
Usually not. Most tools truncate silently and return an answer based on the portion they received. The output gives no indication that sections were dropped.
Does chunking a file solve the context window problem?
It helps, but only if related provisions land in the same chunk. A clause that modifies a term defined elsewhere in the document will not be connected if the two sections are processed separately.
SOURCES
Where this comes from
What the accounting job market is actually asking for.
GO DEEPER
Go deeper
IN THIS SERIES
Next: Why agents need approval gates (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 · Remote A&F Jobs
Audit Friendly · modernaccounting.ai