AI document review
Check whether an evidence's documents meet the requirement before the audit, with every finding backed by a passage from the file.
What it is for
The AI review reads the documents attached to an evidence and compares them with the requirement and with the evidence the framework expects. It returns a short conclusion, saying whether the documents meet the requirement, and the findings that support it.
Your team still makes the call. The AI points out gaps, and whoever owns the evidence applies or dismisses each finding.
Findings with quotes
Every finding quotes the passage it relies on and, when available, the page. You can check it against the file before acting.
A second opinion
A model from another provider tries to refute each finding. Blocking findings go to a panel of models.
History across versions
When you review an updated document again, you see what is new, what is still open and what was resolved.
Enabled per organization
AI is off by default, because a review sends the content of your documents to an external model provider. If the Assistant item is missing from your menu, ask the Clav team to enable the feature for your organization.
How it works
Open the evidence and click Review with AI
The button sits in the evidence’s action row, after the other actions. It only shows when the evidence has at least one document the server can read. Any member of the organization can start a review.
The review runs in the background
It takes a few minutes, and you can keep working in the meantime. The button changes to Open AI review, which takes you to the review conversation. Progress shows up there in steps: Reading the document, Evaluating against the requirement and Verifying the findings.
Findings are checked
Before a finding reaches you, it is checked in two ways: its quote is searched for in the document’s text, and a second model tries to refute it. At the end, the conversation says how many findings were kept and how many were discarded.
You read the conclusion and decide
The conclusion comes last, in two to four sentences, written from the findings that survived the checks. On each finding, click Apply or Dismiss.
What the AI receives
The review looks at every readable document on the evidence together: the main file, the file on each checklist item (labelled with the item’s title) and the complementary documents.
Along with the files, the AI gets the framework context: the framework, the control, the requirement with its legal basis, the expected evidence and the items it should contain. On a re-run, it also gets the findings that are still open.
Your team’s annotations and notes on the evidence are not sent.
| Limit | Value |
|---|---|
| Documents per review | Up to 10 |
| Total size per review | Up to 20 MB |
Files beyond these limits are left out of the review.
Files with client-side encryption are not reviewed
When a file is stored with client-side encryption, the server only holds the ciphertext and can’t read it. Those files are skipped. If every document on the evidence is encrypted, the Review with AI button doesn’t show. The upload architecture guide explains both storage modes.
Reading a finding
Severity
| Severity | Meaning |
|---|---|
| Blocking | As written, the document does not meet the requirement. |
| Warning | A gap that weakens the evidence. |
| Observation | A point worth noting, with no direct effect on meeting the requirement. |
Quote
Every finding has to be anchored to a real passage in the document, even when it points out something missing. The quote sits below the finding’s title, with the page when the model provides one.
In PDFs, Clav searches for the quote in the text extracted from the file. If the passage isn’t in the document, the finding is discarded before it reaches you. When the quote can’t be checked (in files that aren’t PDFs, for example), the finding carries an Unverified quote badge.
Second opinion
Warnings and observations go through one verifier model. Blocking findings go through a panel of two models, and a single disagreement is enough for the finding to be disputed.
A disputed finding is not removed. It shows a Disputed by a second model badge, with the reason in the tooltip. Read those more carefully before deciding.
Status
| Status | Meaning |
|---|---|
| No badge | Waiting for your team to decide. |
| Applied | Someone applied the finding. It was recorded in the evidence notes. |
| Dismissed | Someone dismissed the finding. It stays visible in the history. |
| Resolved in a newer document | A re-run on a changed document found the gap fixed. |
| Replaced by a later review | The gap was still there on the re-run and got a new version. |
| Contested | The AI changed its mind while the document stayed the same. |
Applying a finding adds a line to the evidence notes, starting with
[AI review], with the finding’s title and detail. The platform records who
applied or dismissed it and when.
Reviewing again after a fix
After updating the document, open the review and click Re-evaluate. The new run happens in the same conversation, and each member has one conversation per evidence.
At the end, the conversation compares the two runs and shows how many findings are new, how many are still open and how many were resolved. A gap that persists gets a new version, and the previous one is marked as replaced.
Two rules keep results from swinging between runs:
- A finding is only marked resolved if the content of the files changed. If the AI says the gap is gone but the document is the same, the finding becomes contested.
- On the same document, a finding’s severity can’t go down. The higher severity stays and the finding becomes contested.
Text written to fool the review
If a document contains instructions aimed at the AI, such as a request to approve the file, the review automatically adds a blocking finding that points to that passage.
Assistant
The same feature enables the Assistant item in the menu. It is a page with your conversations on the left, where you ask about your organization’s data in plain language, for example:
- Which evidences are overdue?
- What changed this week?
- How far along is the anti-money laundering framework?
- Who is responsible for the security policy evidence?
The assistant only reads. It searches items by name, lists frameworks and evidences, counts evidences by status, summarizes recent activity and opens an evidence’s details. It doesn’t change any data.
Inside a review conversation, you can also ask about the findings, for example to have a blocking finding explained. The assistant receives the findings with their status and knows which quotes weren’t verified.
Conversations are private: each member sees only the ones they started. Answers come in the language you use the platform in.
Data and privacy
Documents and questions are sent to language models through OpenRouter. Calls require providers that don’t train on the data or retain it.
Each organization has a daily AI usage limit. Once it is reached, new reviews and questions are blocked until the next day (counted in UTC).
Error messages
| Message | What to do |
|---|---|
| The AI assistant is not enabled for this organization | Ask the Clav team to enable the feature. |
| This evidence has no document to review | Attach a document to the evidence before reviewing. |
| AI review is unavailable while client-side encryption is enabled, because the server cannot read the document | Every file on the evidence is encrypted on the client. See the warning above. |
| The document could not be read | The file couldn’t be read. Check that it opens normally and upload it again. |
| This organization reached its daily AI usage limit | Wait until the next day or contact the Clav team. |
| The evaluation failed. Try again. | Click Re-evaluate. |
Frequently asked questions
Does the AI approve or reject the evidence?
No. The review doesn’t change the evidence’s status. It gives your team findings to decide on, and approval follows the usual review flow.
Can I review spreadsheets or images?
Files are sent in their original format, but quote checking only works on PDFs. In other formats, findings show as Unverified quote. For the best results, use PDFs with selectable text.
Can other members see my review?
No. The review conversation belongs to whoever started it. Applied findings, on the other hand, go to the evidence notes, which the whole team sees.
What happens if I delete the conversation?
The conversation leaves your list and, if a review is running, it is cancelled.
Compliance frameworks
How Banco Central rules arrive on the platform broken down into controls, requirements and evidence, and how your team works each one through to completion.
Wallet verification
Prove your customer controls a self-custodied wallet before releasing a withdrawal, without sending personal data to Clav.