How an AI employee reads a 300-page document.

The unit of work in the roles we build for is not a message or a ticket. It is a document, and usually a long one. A commercial credit file runs a few hundred pages once you count financial statements, appraisals, and correspondence. A quality dossier for a production batch is the same. So is the paperwork behind a disputed freight invoice. When we say an AI employee reviews documents, the honest question is: how does software actually read 300 pages?

The tempting answer is the one the model vendors made easy: context windows are enormous now, so paste the whole thing in and ask. We tried it. It produces answers that sound right, cost too much, and cannot be defended, because when a reviewer asks where a claim came from, the trail points at 300 pages, which is the same as pointing at nothing.

Why the one-big-prompt read fails

Three problems, in increasing order of severity. Attention: models degrade on very long inputs, facts in the middle get skipped, and two statements 200 pages apart that contradict each other are rarely caught. Cost: paying for all 300 pages on every question makes the per-task economics ugly for work that runs all day. Auditability, the fatal one: a conclusion drawn from an undifferentiated blob has no citation. In regulated review work, a finding you cannot trace is not a finding. It is a liability.

Index first, read second

So our employees read the way a trained reviewer reads, which is to say they do not start at page one. The first pass builds an index: the page map of the document, its sections and exhibits, where the tables sit, which attachments exist, what the defined terms are. The index is cheap to build, and it turns a 300-page blob into a structured object the employee can plan against.

The second pass is deliberate. The employee works a checklist that comes from the role and the process map, and for each item it decides which sections bear on the question, pulls those pages, and extracts what it needs. A debt service coverage question pulls the financial statements and the loan terms. A packaging deviation question pulls the batch record section and the relevant spec. Most of the document is never fed to the model for most questions, and that is a feature, not a shortcut.

• →Index: map the sections, tables, exhibits, and defined terms of the document.
• →Plan: for each checklist item, choose the sections that bear on it.
• →Pull: read those sections and extract findings, each recorded with the page span it came from.
• →Cross-check: when two sections disagree, pull both and flag the conflict instead of picking one silently.
• →Journal: write every step, pull, and citation into the task journal as it happens.

Page-level citations are the product

Every finding the employee submits carries page citations. Not a document name. Pages. When the file reaches the approver, each line of the summary links to the exact page that supports it. The approver does not re-read 300 pages to sign off. They spot-check the pages behind the findings that matter, which is how experienced reviewers already supervise junior ones.

A finding without a page number is an opinion. A finding with one is evidence.

What the journal makes auditable

Because reading happens as explicit pulls, the journal records not just what was read but what was not. An auditor can see that the employee pulled the appraisal and the covenant schedule and never opened the marketing addendum, and can judge whether that coverage was right for the task. When a miss happens, and misses happen, the journal turns it from a mystery into a diagnosis: the index missed the section, or the plan skipped it, or the extraction misread the table. Each of those has a different fix, and you cannot fix what you cannot see.

The honest limits

Index-first reading is only as good as the index, and real documents fight back. Scans with poor OCR. Tables that lie about their own headers. Exhibits referenced but never attached. Page numbers that restart three times inside one PDF. When index confidence is low, the employee reads more, not less, and says so in the journal. And the approval gate stays regardless: the point of the architecture is not that the employee never misreads. It is that when it does, a named person with page-level evidence in front of them is positioned to catch it.

Reading is the whole job in these roles. Getting it right is unglamorous work about page maps and pull plans and citation plumbing. It is also the difference between an AI that produces confident summaries and an employee whose work survives a review.