How Case Size Affects the Quality of AI-Generated Reports

Introduction

This guide explains a simple but important trade-off in Sky AI: the more pages a case contains, the more information the AI has to process at once — and the higher the chance of small inaccuracies or "hallucinations" in the final report.

We do not impose a hard limit on the number of pages in a case. You can process cases of any size. Instead, this guide is here to be honest with you about the trade-off and to give you concrete, practical steps for working with large cases so your reports stay accurate and defensible.

Who This Guide Is For: Anyone who generates reports from case documents in Sky AI — report writers, QA specialists, case managers, assessors, and organization administrators. No technical or AI background required.

What You'll Learn:

• Why case size affects report accuracy (in plain language)
• A practical size scale (XS → XL) mapped to page counts, and what each size means for your review effort
• Three concrete strategies for keeping large cases accurate: splitting into versions, tuning templates with the Category Filter, and targeted review
• Best practices and common mistakes to avoid

The Core Idea: AI Is a Tool, Not a Source of Absolute Truth

The best way to think about Sky AI is the same mental model we use in the report-templates guide: imagine a brand-new employee who is smart, fast, and extremely literal. They have read every page in the case file — but the more pages you hand them at once, the harder it becomes to hold every detail in mind at the same time, keep facts straight, and avoid mixing up similar-looking documents.

Two consequences follow directly from this:

1. The AI is a powerful assistant, not an infallible oracle. It accelerates the work dramatically, but the responsibility for a correct, defensible report stays with you. The more content the AI processes, the more important it is to verify what it produced.
2. The larger the volume of pages, the lower our confidence in every single detail. This is not a flaw unique to Sky AI — it is a fundamental property of how AI systems handle large amounts of information. A fact buried in page 8,000 of a 10,000-page case is simply harder to retrieve and reproduce accurately than the same fact in a 100-page case.

None of this means large cases produce bad reports. It means large cases need more of your attention — especially on factual details, dates, and names.

Why Bigger Cases Mean Lower Confidence

You don't need to understand the internals to use this guide, but a little background makes the recommendations obvious.

When Sky AI builds a report or answers a chat question, it works from the content of your case documents. That content has to fit into the AI's working memory (its "context window") for any single generation step. As a case grows:

• There is more to read and rank. The AI must sift through far more documents to find the pages relevant to each section. The more candidates there are, the more chances there are to miss the right one or pull in the wrong one.
• Content may not all fit at once. When the combined material for a step is very large, only part of it can be used at a time, and the system has to prioritize what to include. Details that don't make the cut can be underweighted.
• More facts must be reconciled. Large cases contain more repeated, updated, or contradictory information (e.g., a diagnosis that changes across visits). The more facts there are to merge, the more room for a subtle error.

The practical upshot: accuracy degrades gradually as size grows — it does not fall off a cliff at a specific page number. That's exactly why we give you a soft size scale rather than a hard limit.

The Case Size Scale

Use the table below as an orientation, not a rule. The page numbers are approximate guidance for how much scrutiny a case deserves — they are not enforced thresholds, and Sky AI will happily process cases well beyond XL.

💡 Key takeaway: Think of L and XL the way you'd think of a task that's grown too big for one sitting — not forbidden, but a signal to slow down, verify more, and consider breaking it into smaller pieces.

What to Do With Large Cases

Here are three concrete strategies, in rough order of how much they reduce risk. You can combine them.

Strategy 1 — Split documents into a separate Version to reduce volume

If a case is very large, you don't have to process all of it as one block. Sky AI lets you create a new Version of a case built from a selected subset of documents. The files are duplicated into the new version, and changes there don't affect the Master version — so you can safely carve a large case into smaller, focused working copies.

How to do it:

1. Select the documents you want to work with.
2. Create a new version from the selected documents (Create New Version → based on the selected documents).
3. Generate your report inside that smaller version, where the AI has far less to process at once and can be more precise.

This is the most effective lever for very large cases: a 10,000-page case split into a few focused 1,000–2,000-page versions gives you S/M-level accuracy on each, instead of XL-level uncertainty across the whole thing.

💡 Tip: Split along natural boundaries — by provider, by time period, or by document type — so each version is coherent on its own.

For step-by-step instructions on moving files into a version, see "How to Add Documents to Version".

Strategy 2 — Tune your report templates to limit what the AI reads

You can reduce the volume the AI processes per section without touching the case itself, by making your report templates more targeted. The main tool for this is the Category Filter on each template block.

By default, every block can draw from every document in the case. On a large case, that means each section wades through everything. Instead:

• Narrow the Category Filter on each section so it only reads the document categories it actually needs. An "Imaging Findings" section should read imaging reports — not correspondence, billing, or legal letters.
• Exclude noise categories (Correspondence, Miscellaneous, Administrative) from analytical sections entirely.
• Split broad sections into focused ones, each with its own tight filter, rather than one giant section that tries to read the whole case.

On a large case, a well-filtered section is both more accurate (less irrelevant content to get confused by) and easier to verify (fewer document types to check against). For the full mechanics of building and filtering templates, see the companion guide "How to Write Effective Prompts for Report Templates."

On a large case, a well-filtered section is both more accurate (less irrelevant content to get confused by) and easier to verify (fewer document types to check against). For step-by-step instructions on building and editing templates, see "How to Create and Edit Report Templates". For the full mechanics of writing the prompts and filters inside them, see the companion guide "How to Write Effective Prompts for Report Templates".

Strategy 3 — Review the AI's output carefully (especially on large cases)

No amount of tuning removes your responsibility to check the result. On L and XL cases in particular, review the report against the sources with extra care, focusing on the things most likely to drift:

• Facts — diagnoses, findings, conclusions. Confirm each important claim against a source document.
• Dates — dates of incident, treatment, assessment, and report. These are easy to transpose or misattribute when there are thousands of pages.
• Names — providers, claimants, assessors, employers. Verify that the right name is attached to the right event.

💡 Tip: In the report-templates guide we recommend adding an explicit rule to your prompts: "If information is not found in the sources, state 'Not Documented' — do not guess or infer." This is doubly valuable on large cases, where the temptation for the AI to fill a gap with a plausible-sounding detail is highest. Anchoring every claim to a named, dated source also makes your own review far faster.

Best Practices

Match your review effort to the case size. A 100-page case needs a light pass; a 10,000-page case needs deliberate, section-by-section verification. Use the size scale as your guide.

Prefer several smaller versions over one enormous case. Splitting is the single most effective way to raise accuracy on very large matters.

Filter aggressively on large cases. The bigger the case, the more a narrow Category Filter pays off — in both accuracy and verifiability.

Verify facts, dates, and names first. These are the details most likely to drift as volume grows, and the most consequential if wrong.

Treat the AI as a fast first draft, not a final answer. The larger the case, the more this matters.

Common Mistakes to Avoid

Mistake 1: Assuming a bigger case is automatically as accurate as a small one

Problem: Trusting a 10,000-page report the same way you'd trust a 100-page one, with the same light review.

Solution: Scale your scrutiny with the case. Use the XS–XL table to decide how much verification is warranted.

Mistake 2: Running everything against one giant case when it could be split

Problem: Keeping a massive case as a single block and accepting lower confidence across the board.

Solution: Create focused versions from subsets of documents (Strategy 1). Smaller working copies restore higher accuracy.

Mistake 3: Leaving every Category Filter wide open on a large case

Problem: Every section reads the entire case, picking up noise and irrelevant content.

Solution: Narrow the filter per section. Analytical sections should read only the categories they need.

Mistake 4: Skipping verification of dates and names

Problem: Reviewing the prose for readability but not checking that dates and names are correct.

Solution: On L/XL cases, explicitly verify dates and names against sources — these are the highest-risk, highest-impact details.

Conclusion

Sky AI can process cases of any size, and we don't stop you from working with very large ones. But we want to be honest about the trade-off: the more pages the AI handles at once, the lower our confidence in every individual detail, and the more you need to verify what it produced.

Key takeaways:

• AI is a tool, not absolute truth — the more content it processes, the more you should check its output
• Bigger cases mean lower confidence per detail — accuracy degrades gradually with size, so scale your review accordingly
• There is no hard page limit — the XS–XL scale is guidance for how much scrutiny a case deserves, not an enforced cap
• Reduce volume when you can — split large cases into focused versions built from selected documents
• Filter templates to limit what the AI reads — narrow the Category Filter per section to cut noise and improve accuracy
• Verify facts, dates, and names — especially on large cases

Next Steps

1. Check the size of your current case against the XS–XL scale and decide how much review it warrants.
2. If it's L or XL, consider creating a focused Version from a subset of documents.
3. Open your report templates and narrow the Category Filter on any section that doesn't need the whole case.
4. On your next large report, do a deliberate pass over dates and names against the source documents.