How Clivanta's AI pre-checks documents — without ever auto-deciding

A client uploads the quarterly bank statement against the wrong Request Item. It’s the right bank, the right account, the right kind of document — just the statement for the previous quarter, not this one. In an email inbox, that file looks exactly like a correct one. Nobody is checking the date on every attachment, so it sits there looking done until a reviewer opens it three weeks later and notices the period is off.

That small class of error — present but wrong — is where a lot of quiet audit risk lives. It’s also exactly the kind of thing a machine is good at spotting and a busy human is likely to miss. So Clivanta runs a set of AI pre-checks on every upload. What matters as much as what they do is what they deliberately don’t: they never decide anything.

What the pre-checks actually do

When a Client uploads a Document, the pre-checks run asynchronously, in the background, so the client isn’t left waiting on a spinner and the upload never blocks. By the time a Firm User opens the item to review, the analysis is usually already sitting there. Concretely, the pre-checks:

• Classify the document type with a confidence level — is this a bank statement, an invoice listing, a passport, a set of financial statements.
• Extract the document date and the entity name from the file itself, so the reviewer can see what the document says it is, not just what it was filed as.
• Detect duplicates by file hash within a Pack, so the same file uploaded twice against two items is flagged rather than silently accepted in both places.
• Flag a likely wrong period or wrong entity by comparing the extracted date and name against the Engagement and Period the item belongs to.
• Suggest matches for bulk-uploaded files that arrived unassigned, proposing which Request Item each one probably answers.

Each of these is a piece of information laid next to the document, framed as a suggestion. “This looks like a bank statement for FY2025, but this Period is FY2026.” “This file has the same hash as a Document already uploaded on item 7.” Useful, specific, and — crucially — inert until a person acts on it.

The principle: advisory only, always

Here is the line Clivanta will not cross, and it’s worth stating plainly. The AI never accepts, rejects, moves, or concludes anything. It does not change a Request Item’s status. It does not promote a Document to Evidence. It does not reassign a file from one item to another. It does not close an item or mark it satisfied.

Every one of those is a decision, and every decision is made by a Firm User whose name and timestamp go on the record. The AI’s job ends at surfacing what it noticed. A person’s judgement begins there and doesn’t hand back control at any point.

This is a design choice, not a limitation we’re apologising for. A document collection tool that quietly auto-accepted files because a model was “confident enough” would be faster and worse. Confidence is not the same as correctness, and the whole point of the evidence record is that a human stood behind each item. Automating the decision away would hollow out exactly the thing the firm is accountable for.

Why advisory-only matters for trust and defensibility

Think about what a defensible record has to show: who asked, who sent, who reviewed, who accepted, and when. If an AI could accept a Document to Evidence on its own, that chain breaks at the most important link. “The system accepted it” is not an answer you want to give a reviewer or a regulator six months later.

By keeping the AI strictly advisory, the audit trail stays clean. Every acceptance is a decision with a person attached. The AI’s flags can appear in the item’s activity — “pre-check flagged possible wrong period” — as context for why a reviewer looked twice, but the flag never becomes an action on its own. The record reads as it should: a machine noticed something, a qualified person decided what to do about it.

That separation also protects the client relationship. Clients extend trust to your firm, not to an algorithm. A file rejected by a model with no explanation feels arbitrary. A revision requested by your reviewer, citing a specific reason, is a normal professional exchange. Keeping the human on the decisions keeps the relationship human.

There’s a security dimension too. Advisory-only means the AI has no authority to alter the record — it can read and annotate, not conclude — which is one fewer path by which the system state can change without a person and a log entry behind it. We go into the broader posture in more detail on the security page.

When the pre-checks can’t read a file

Not every upload is clean. A photo of a document taken at an angle, a corrupt PDF, a scan too faint to parse — these happen, especially from clients uploading on a phone under time pressure.

When a pre-check can’t process a file, it says so. The result surfaces as “unable to process” rather than a wrong guess, and — this is the important part — it doesn’t block review. The Document is still there, still uploaded against the right item, still fully available for a Firm User to open and assess by eye. A failed pre-check degrades gracefully to the situation you’d have had anyway: a person looking at a file. It never holds an item hostage because the model couldn’t read something.

That’s the right failure mode for an advisory system. The floor of the experience is “a human reviews the document,” which is exactly where firms operate today. The pre-checks only ever add signal on top; they never subtract capability when they fall short.

The division of labour

The clean way to think about it: the AI does the tedious, high-volume noticing, and your people do the deciding. It reads dates and entity names off hundreds of files so a reviewer doesn’t have to squint at each one. It remembers every hash so a duplicate can’t slip through in two places. It proposes where a pile of bulk-uploaded files probably belong, so sorting them is a confirm rather than a hunt.

And then it stops. The accept, the reject, the reassignment, the sign-off — those stay with the firm, because those are the parts that carry judgement and accountability. That balance is deliberate throughout Clivanta: let the machine catch the wrong-period bank statement, and let a person decide what to do about it.

The AI is a good pair of eyes. It was never meant to be the one who signs.