AI & Data Processing

The first 3,000 characters of each uploaded PDF are sent to a small, fast model that answers one question: what kind of document is this? The answer routes the file to the right reader below. It is recorded alongside the document type you picked and never overwrites it. Emailed attachments are the one exception - see below - because nobody chose a type for them.

Up to the first 12,000 characters, labelled by page. The model returns the parties, contract name and type, start and end dates, auto-renewal, notice and cancellation notice periods, annual value, the breach notification window, contact, definition, remediation and the breach clause as written, and the eight NCUA-relevant provisions: audit rights, data security, breach notification, business continuity, termination, subcontractor oversight, performance standards and dispute resolution.

Every field comes back with the page number it was read from, the sentence it was read from, and a confidence score of 0-100. Fields that score below 55, and material fields that score below 80, are re-read by a second, stronger model before you see them.

Read in two passes: about 15,000 characters from the front of the report for the period covered, the auditor who signed it, the opinion type, the scope, and whether a bridge letter is attached and through what date; then up to 45,000 characters of the pages most likely to carry findings, for subservice organizations, complementary user entity controls (CUECs) and exceptions with the auditor's own management response.

Up to 25 exceptions are then rated for significance to your institution rather than in the abstract, and the ones rated HIGH or MEDIUM - plus any the first pass could not rate - are read again by a stronger model. A HIGH-significance exception, or a bridge letter gap, raises an item in your Action Center for a person to review. On Continuous, where a prior year's report is on file, this year's reading is also compared against it - from the two stored readings, not by re-opening the PDFs.

VendorLockbox uses page classification to identify the table of contents, executive summary, scope, opinion, and findings sections, then analyzes those targeted sections. The report is not processed page by page - relevant sections are identified first, then sent to the model. This means a finding buried in an appendix section not classified as relevant may not be surfaced. Reviewers should treat the analysis as covering the key report sections, not a guarantee of exhaustive page-by-page review.

Up to 24,000 characters, page-labelled. The model returns the insurer, the policy number, effective and expiration dates, each coverage type and its limits, and whether your institution is named as an additional insured. Whether a certificate has lapsed or is about to is then computed from those dates, not decided by the model.

Attachments forwarded or sent to your organization's inbound address are processed identically to documents you upload by hand: classified first, then routed to the SOC or insurance reader above. The classification is also allowed to set the document type here, which it is not allowed to do anywhere else, because no person picked one. Below Complete, or with AI off, the attachment is still stored and still raises an action item - it simply arrives unclassified for a person to type.

The sender address and subject line are used to file the message - to match it to a vendor by domain, and to name the document something more useful than scan_0001.pdf. They are not sent to the AI provider. They are stored on the document record and on the action item the arrival raises, so that a filing decision can be audited later. See Data retention.

AI writes a question set for the vendor in front of you, from the vendor's type and risk tier, the evidence already on file and how close it is to lapsing, and the contract provisions somebody confirmed present or absent - so a question a document already answers is asked about that document instead of from scratch. The standard questionnaire is used when AI is off, and can be chosen per review.

When responses come back, AI reads the answers - up to 1,200 characters of each of the seven questions that imply a specific document - to decide what the vendor actually claimed. This cross-check runs on any plan with AI switched on, and falls back to reading the answers as plain text when it is off. Whether the matching document exists and is current is then a database check, not the model's opinion. Discrepancies become an item for a person to look at; nothing is written to the vendor's record or its review status.

Nightly, for CRITICAL and HIGH risk vendors only, up to 20 vendors per organization per run, taken least-recently-checked first. The vendor's name goes to a search service; the resulting headlines, the vendor's name, website domain and category go to a fast model, which classifies the story as BREACH, ACQUISITION, REGULATORY, FINANCIAL, LEADERSHIP or IRRELEVANT with a severity.

Anything classified IRRELEVANT, or not judged relevant to a financial institution, is discarded and never stored. The model is told to answer IRRELEVANT when unsure - a same-named company is the common failure here, and silence is the cheaper mistake. What survives is recorded as an event and, when severe enough, raises an action item. Acting on it is your decision; a stored event is closed only by a person dismissing it.

The action items themselves are computed - an expiring document, a renewal deadline, a missing SOC report - by rules over your records, with no model involved. AI writes the plain-English explanation attached to an item: why it matters and what to do about it. If the model is unavailable the item still exists, without the paragraph.

On Complete and above, missing-document items get the same treatment plus a short request email drafted to the vendor contact. It is a draft. It is stored on the item and opens in your mail client for you to edit and send. VendorLockbox never sends it.

Three to four paragraphs of prose over the figures the board report already computed - vendor counts, tier distribution, the readiness score and what moved it, open items. The model is given those facts and nothing else, so it has no number of its own to reach for.

It arrives unapproved, and the generated PDF omits the narrative entirely until a person approves it. Editing is expected; regenerating withdraws any approval already given.

Within 90 days of the next examination date you set, AI reads your own exam prep checklist and readiness score - not a document, and no external benchmark or peer data - and returns likely examiner focus areas with the procedures each came from, a remediation list with priorities and effort estimates, and question-and-answer talking points. Briefings are cached for seven days, and a person marks one as read. The mock examination feature works the same way, pairing examiner-style questions with the evidence in your own file that answers them.

Effort estimates and predicted focus areas are the model's judgement. They are not a commitment, and they are not a prediction any examiner has endorsed.

After a risk assessment is saved, AI writes the reasoning paragraph an examiner asks for - inherent versus residual risk, what evidence is on file, what is still open - from the tiers your assessor chose and the records behind them. The assessment is recorded whether or not the paragraph is written.

On a contract, AI writes a plain-English summary of what was signed and when something has to happen. It is written from the fields on the record, never by re-reading the PDF, so every sentence in it is traceable to a value you can see and correct on the same page.

On Continuous, the monthly readiness digest email includes three AI-written paragraphs over that month's figures. The email goes to your own alert address on a schedule; it is never sent to a vendor or any outside party, and it still goes out with its figures if the model is unavailable.

For a reportable cyber incident - incident tracking is a Continuous feature - AI improves the four prose fields of the 12 CFR §748.1(c) notification draft on top of a deterministic prefill from the incident record. The prefill is the floor: a field the record cannot support is left blank for a human, and the contact block is overwritten from your organization's data after the model runs. VendorLockbox does not file anything with the NCUA.

The Action Center search box turns a typed sentence into a filter. The model sees your sentence and a list of your vendor names - never an action item, a document or a contract - and answers with the filter, which you can see and clear.