Build a searchable evidence register with AI
The bid says the supplier is ISO 14001 certified, and there's a certificate attached. Finding is not supporting: a source can be the right document and still fail to back the headline.
What you'll make
an evidence register: a table that puts each claim next to its source, the exact passage and a note on whether that passage really supports it.
What you'll need
the document making the claims and the sources behind them, as files you’re allowed to use with AI; ChatGPT, Claude, Copilot or Gemini in your work account; and about 20 minutes.
Example result.
| Claim | Source and excerpt | Support |
|---|---|---|
| C1 ISO 14001 certified across all sites | S1 certificate, p.1: “Scope: Leeds depot” | Partial: one site only |
| C3 Can start on 1 December | S3 cover letter: “subject to contract signature by 6 November” | Conditional |
1. Collect the claims and their sources
Attach the document that makes the claims and every source it relies on, in one chat.
| Tool | Add the files |
|---|---|
| ChatGPT | + in the message box |
| Claude | + › Add files or photos |
| Copilot | The paperclip in the Copilot Chat box, or a OneDrive or SharePoint file |
| Gemini | Add files › Upload files, or Add from Drive |
Pick the claims that matter for the decision. Ten important claims make a better register than fifty trivial ones.
Why this works: the register can only check what it can read. A source left out turns a supported claim into an unsupported one.
What you should see: the files attached and a short list of claims, each with an ID such as C1, C2.
2. Ask for the register as a table
Build an evidence register as a table with six columns: claim ID, claim,
source ID, location (page or section), exact excerpt, support note.
Use only the attached files. Copy excerpts word for word. In the support
note, say whether the excerpt fully supports, partly supports, makes
conditional or does not support the claim, and why. Keep unsupported
claims in the table. Do not invent quotes, page numbers or sources.
Return it as an .xlsx file I can sort and filter.If your tool can’t create a file, ask for the table in the chat and paste it into Excel or Google Sheets. In Claude, file creation needs Settings › Capabilities › Code execution and file creation turned on.
Why this works: separating “is the quote exact?” from “does it support the claim?” stops a correct link being marked as proof.
What you should see: one row per claim, with an excerpt you can find in the source and a support note in plain words.
3. Check the support notes, not just the links
Filter the table to rows marked “fully supports” and open each source at the stated location. Is the excerpt word for word? Does it say as much as the claim, or less?
For each row marked as fully supported, compare the claim with what the
excerpt establishes. Flag any changed scope (fewer sites, fewer items,
a narrower date), missing condition or stronger conclusion. Suggest a
narrower claim the excerpt does support, and keep the original claim
in the row.Why this works: the model is good at finding the right document and less good at noticing the claim says more than the document does. A second, narrow pass catches that.
What you should see: some rows downgraded to partial or conditional, each with a narrower wording.
Check the result
- Every excerpt appears word for word at the stated location.
- Each support note matches what the excerpt says, not what the claim hopes.
- Unsupported claims are still in the table, not filtered out.
- Sort by claim ID: every claim from the document has a row.
The usual failure: finding is not supporting. The model links the right certificate or report and marks the claim supported, although the passage covers one site, some items or a conditional date. Spot it by reading the claim and the excerpt aloud together.
Row [claim ID] is marked supported, but the excerpt says [what it says].
Change the support note to partial, conditional or not supported, give
the reason in one line and suggest a narrower claim the excerpt does
support. Keep the exact excerpt and location unchanged.[claim ID] is the row; [what it says] is the limit you noticed in the excerpt.
Worked example: a cleaning contract bid
Fictional example. Farah is comparing bids for an office cleaning contract. One bid makes three claims and attaches three documents:
C1 "ISO 14001 certified across all sites."
C2 "We use 100% eco-certified products."
C3 "We can start on 1 December."
S1 ISO 14001:2015 certificate. Scope: Leeds depot. Valid to 30 June 2027.
S2 Product list: 8 of 11 products carry the EU Ecolabel.
S3 Cover letter: "Start date of 1 December subject to contract signature by 6 November."The first register marked C1 “Supported — S1 ISO 14001 certificate”.
- The failure. S1 is the right certificate, but its scope is one depot. The bid says “all sites”. That’s finding, not supporting.
- The repair. The repair prompt, with C1 and “Scope: Leeds depot”, changed the note to “Partial: certificate covers the Leeds depot only” and suggested “ISO 14001 certified at the Leeds depot”.
- The other rows. C2 was correctly marked not supported (8 of 11 is not 100%), and C3 conditional on signature by 6 November. Farah now has three specific questions for the supplier.
Task kit for this job: Download (ZIP) · Reusable instructions
Go deeper
Make it searchable when the register gets large
A sortable table is enough for a few dozen claims. For hundreds, ask for a small local HTML page instead: “Turn this register into one self-contained HTML file with a search box that matches claim text, claim IDs and source IDs, and a full-record view for each row. No external scripts or network calls.” Then test it: search for a claim ID and a source ID and check both return the right record.
Keep source versions stable
Record the file name and version, or the date you downloaded it. When a source changes, review the rows that cite it. An old excerpt may no longer represent the current document.
Carry the trail into the final summary
When the register feeds a summary, keep each short claim linked to its qualification. The executive-summary guide shows how to compress without losing the trail.