What the review has to produce
A literature review does three jobs: it maps what is already known, it shows where sources disagree, and it states what is still open. A summary of each paper does none of those jobs on its own.
Decide the output before you collect anything. A matrix plus a one-page synthesis is usually enough. A list of paper summaries is not, because the reader cannot see which claim each paper supports.
Build the source matrix first
Create one row per paper with fixed columns: citation, question, method, sample, finding, limitation, and the page you took the finding from. Fixed columns are what make the matrix checkable.
Ask AI to fill rows only from papers you provide. For each row, it should quote or point to the sentence it used. If it cannot point to a page, the row stays empty and you decide whether to keep the paper at all.
Sort by claim, not by paper
Then reorganize the same rows by claim. This is the step that turns a reading list into a review: for each claim, list the papers that support it, the papers that qualify it, and the papers that contradict it.
Ask AI to mark claims with only one supporting source and claims where the methods differ too much to compare. Both marks save reviewers from overstating what the literature says.
Draft from the matrix
Write each section from one claim and its rows. The draft should name the evidence, not just the conclusion: how many studies, what kind, and how consistent. When a claim has thin evidence, say so in the sentence rather than in a footnote.
If you ask AI to draft, give it the claim and the rows for that claim, not the whole pile. Smaller inputs keep each sentence traceable to a row you already checked.
Check citations before you share
Run three checks: every in-text citation exists in the matrix, every matrix row has a page or quote, and every claim without support is either removed or labeled as open. These checks catch the failures that make a review unusable.
Do not let the tool invent a citation to fill a gap. An empty cell is a finding; a fabricated reference is a defect.
How to run it on Sigla
Keep the matrix, the PDFs, and the draft in one workspace. Open papers that need a sign-in in the built-in browser, and ask AI to add rows only after you confirm the page. Then draft section by section from the matrix.
For the research process behind it, read how to do AI deep research. If the review is part of a client deliverable, the AI research workflow for consultants shows how to carry it into the handoff.
FAQ
How is a literature review workflow different from a summary?
A summary compresses one paper. A review compares papers by claim, shows disagreements, and states what is still unresolved. The matrix is what makes the comparison possible.
Can AI read the PDFs for me?
It can extract findings and page references from the files you provide. It should not guess when a page or a number is unreadable; leave the cell empty and check it yourself.
How many papers can one review hold?
As many as your matrix can stay readable. Split by claim or by sub-question when rows stop fitting on one screen, and keep the split visible in the draft.
Next step
Create the matrix columns before you add the first paper. Then add five papers and check whether each row can point to a page.