Bank loan docs. AI underwriting. 90%+ accuracy .
VZU's financial services playbook. Data processing, AI underwriting, voice agents, and decision intelligence. Powered by Sentinel (audit), Vega (voice), and Quill (reconcile) on the VZU OS runtime. Real deployments. Real metrics. The brief is the contract. The work is the work. The audit trail is the regulator's form.
Industry
Financial Services
90%+
Extraction accuracy
Direct answer
What does VZU do for financial services?
VZU's financial services playbook is data processing, AI underwriting, and document AI on the VZU OS runtime. Sentinel audits every extraction. Vega runs the inbound voice. Quill reconciles the ledger. Real deployments: F500 bank loan document processing at 90%+ accuracy, Global Bank 40% lower storage cost, Bancassurance 50% faster policy issuance.
Use cases
Five deployments. Real numbers.
-
Loan document processing
F500 bank, 2.3M docs/year across 14 jurisdictions. 90%+ extraction accuracy. 40% lower storage cost. The Sentinel agent audits per-page. The Quill agent reconciles the loan ledger.
-
AI underwriting
Bancassurance provider. 50% faster policy issuance. The Oracle agent writes the underwriting prompt. The Sentinel agent audits the decision. The Quill agent writes the policy to the core system.
-
Legacy archive digitization
F500 bank legacy archive. The Hunter agent pulled 1.4M documents. The Quill agent extracted the data. The Sentinel agent audited the extraction. The audit trail is per-page.
-
KYC + account opening
Multi-jurisdiction KYC packs. The Hunter agent pulled the documents. The Sentinel agent ran the sanctions check. The Quill agent wrote the KYC record. The audit trail is per-record.
-
Inbound voice for retail banking
Vega agent runs the call. Sub-200ms p50. 14 languages. CRM writeback via MCP. The audit trail is on from the first character. The Vega agent escalates to a human when needed.
A case study · F500 Bank
2.3M loan documents. 90%+ accuracy. 40% lower storage cost.
Chapter 01 · The brief
A F500 bank, buried under 2.3M loan documents a year.
A Fortune 500 bank ran 14 jurisdictions. The loan operations team processed 2.3M document images a year. Each document was a 90-step manual flow. The first step was scanning. The second was data entry. The third was reconciliation. The fourth was a quality check. The fifth was a hand-off to the loan officer. The sixth was a re-key into the loan ledger. The seventh was a re-key into the regulatory report. The eighth was a re-key into the audit trail.
The brief from the bank was specific. Build an extraction pipeline that hits 90%+ accuracy on the first pass. Build a reconciliation layer against the existing loan ledger. Build a per-page audit trail. Build it on-prem. Build it SOC 2 + ISO 27001 aligned. Build it in 12 weeks. The brief is the contract. The work is the work.
2.3M
Documents a year
Chapter 02 · The architecture
Sentinel, Quill, and Vega on the VZU runtime.
The Sentinel agent runs the extraction. The model is the bank's choice. The prompt is the Oracle agent's. The eval suite scores every extraction. The audit trail records the model, the prompt version, the eval result, and the operator's review.
The Quill agent reconciles the extraction against the loan ledger. Row-level diff. Reason per row. The reconciliation rate is 99.7%. The reconciliation is on per-row. The audit trail is the regulator's form.
The Vega agent runs the inbound voice. Sub-200ms p50. 14 languages. CRM writeback via MCP. The Vega agent escalates to a human when the borrower asks for a loan officer. The audit trail records the call. The work is the work.
“90%+ extraction accuracy. 40% lower storage cost. Per-page audit trail. The brief is the contract. The work is the work. The audit trail is the regulator's form.”
Chapter 03 · The result
40% lower storage cost. 90%+ accuracy. 6-week payback.
The pipeline shipped in 12 weeks. The first-pass accuracy was 91.4% across the bank's full document set. The reconciliation rate was 99.7%. The storage cost dropped 40% because the documents are now AI-native, not scanned-image-native. The regulatory report is generated from the audit trail, not re-keyed by a junior analyst.
The payback was 6 weeks. The audit posture is the runtime. The model is the bank's. The data residency is the bank's. The compliance is the runtime. The brief is the contract. The work is the work.
40%
Lower storage cost
The Sentinel agent · A real loan document
Watch the Sentinel agent audit a loan document.
The Sentinel agent runs the extraction. Audits the model output. Produces a per-page audit trail. The brief is the contract. The audit trail is the regulator's form.
- 01sentinel › read brief
read brief
- 02sentinel › thinking: postmta security audit. owasp top 10. 2 weeks.
thinking: postmta security audit. owasp top 10. 2 weeks. : postmta security audit. owasp top 10. 2 weeks.
- 03sentinel › mcp.code.read
47,000 lines, hono + prisma + workers
- 04sentinel › mcp.semgrep.run
23 findings, 3 high, 8 medium, 12 low
- 05sentinel › mcp.owasp.scan
auth: 4 issues, rbac: 2, input: 5, secrets: 1 : 4 issues, rbac: 2, input: 5, secrets: 1
- 06sentinel › mcp.deps.audit
4 cves, 2 high, 2 medium
- 07sentinel › mcp.report.write
32-page report with exploit chains + remediation
- 08sentinel › ship. 3 weeks. 23 findings. 4 cves. 32-page report. done
ship. 3 weeks. 23 findings. 4 cves. 32-page report.
Audit trail · 6 entries
- T+00:00:00 brief.received 01
- T+00:00:30 scope.drafted 02
- T+00:08:00 scanners.run 03
- T+00:14:00 findings.drafted 04
- T+01:00:00 report.completed 05
- T+03:00:00 brief.completed 06
- KPI · 01 now
3 weeks
Median audit duration
52% - KPI · 02 now
23
Avg findings per audit
27% - KPI · 03 now
4
Avg CVEs found
69% - KPI · 04 now
100%
Re-audit pass rate
35%
The trust strip
What the runtime has shipped for financial services.
-
F500 Bank · 90%+ accuracy
-
Global Bank · 40% lower storage
-
Bancassurance · 50% faster issuance
-
F500 Bank · 1.4M legacy archive
-
Multi-jurisdiction KYC
A case study · F500 Bank
Case study
91.4%. First-pass accuracy.
The brief
A Fortune 500 bank was processing 2.3M loan documents a year across 14 jurisdictions. Each document was an 8-step manual flow — scan, data entry, reconciliation, QA, handoff to the loan officer, then re-key into the loan ledger, the regulatory report, and the audit trail. Three systems to keep in sync, every time.
The analysis
Eight re-keys per document. No end-to-end audit trail. The regulator wanted per-page provenance. The bank wanted 90%+ first-pass accuracy and a 12-week ship date, on-prem, SOC 2 + ISO 27001 aligned.
The build
Sentinel ran the extraction. Quill reconciled the extraction against the loan ledger — 99.7% reconciliation, row-level diff, reason per row. Vega ran the inbound voice, sub-200ms p50, 14 languages. The audit trail was on from the first call. The data stayed on-prem.
The result
Shipped in 12 weeks. First-pass accuracy: 91.4% across the bank's full document set. Storage cost down 40% because the documents are now AI-native, not scanned-image-native. The regulatory report generates from the audit trail. Payback: 6 weeks.
The runtime
Every VZU brand ships on the NetWit Agentic OS runtime. Cloud, hybrid, on-prem. SOC 2 + HIPAA + ISO 27001.