Bring your own key
Your LLM provider, model, and credentials are scoped to your tenant — Anthropic, an OpenAI-compatible endpoint, or your own. We never put our keys on your traffic.
Governed autonomy for enterprise IT
AI BUILDING BLOCKS
FOR ENTERPRISE IT —
THAT ACT SAFELY.
Velgent turns messy tickets, approvals and documents into structured, policy-checked, confidence-scored signals — the trust layer your agents need before they take action. Platform-agnostic by design, starting with ServiceNow.
See what's shippedaiengine.velgent.com — POST /api/v1/summarise
● awaiting response…summarise · intent · sentiment
Built for production
The numbers behind the building blocks.
5
Structured signals
Triage — intent, urgency, due-date, sensitivity, sentiment
<50ms
Deterministic verdict
Policy Engine validate mode — no LLM hop
0–1
Per-field confidence
Data Extractor — every value source-anchored
0
Bytes persisted
File data in memory only, cleared on response
Tenant-isolatedAudit-logged every callNo persistent storagePHI-aware recognizersPer-tenant LLM routing · BYOK
Live in production
Three building blocks. Already in production.
Triage
AvailableOne-line pinpoint + structured intent for tickets, approvals and knowledge articles. Drop in any ticket body, support thread or approval note — get back the one-line summary your agent needs, plus action_needed, due_date, urgency_score and sensitivity as structured fields.
- Pinpoint summary + structured intent fields in a single call
- Tenant-isolated, audit-logged on every call
- PHI-aware recognizers — tenant-flippable Healthcare profile
POST /api/v1/summarise { "summary": "VPN drops every ~10 min since the 4.2 client update.", "intent": { "action_needed": true, "urgency_score": 0.82, "due_date": "2026-06-05", "sensitivity": "internal" }, "sentiment": { "label": "frustrated", "score": 0.74 } }
Policy Engine
AvailablePolicy as code, authored in plain English. Each policy is English text plus a declared input schema, usable in five modes — validate, generate, decide, score, classify. Compose into chains or graphs (DAGs) with conditional branching. One call returns per-step results plus a weighted aggregate score, bucketed pass / review / block.
- Validate mode: pure DSL, deterministic verdicts in < 50 ms (no LLM hop)
- Plain-English authoring — paste, preview, publish. No DSL, no code review, no compiler step in your CI
- Chain or DAG of N policies in a single API call — weighted aggregate back, no glue code
POST /api/v1/policies/chain { "aggregate": "review", "steps": [ { "policy": "pii_guard", "verdict": "pass" }, { "policy": "approval_tier", "verdict": "review", "score": 0.6, "reason": "spend over $5k" } ] }
Data Extractor
AvailableSchema-driven JSON, HTML fields, or full semantic HTML from text, image, or PDF. Every value comes with a 0–1 confidence score and an anchor pointing at the exact source phrase the model read — so your workflow can auto-approve the easy cases and escalate the uncertain ones without re-parsing the LLM's prose.
- Three output modes — JSON for automation, HTML fields for ticket comments, full document HTML for re-flowing scans
- No persistent storage — file bytes live in memory only, cleared before the response returns
- Anchors mandatory — hallucinations drop to 0.2 confidence; auditors can cite the exact source phrase
- PII visibility — Presidio on every text input, pii_detected events with per-category counts in your audit feed
POST /api/v1/extract { "fields": { "invoice_total": { "value": "1,240.00", "confidence": 0.97, "anchor": "Total Due $1,240.00" } } }
Enterprise grade
Built for the security review.
The questions an enterprise architect asks first — answered in the architecture, not an FAQ.
Zero data retention
Request payloads and file bytes live in memory for the lifetime of the call, then are cleared before the response returns. Nothing written to disk, queue, or database.
Tenant-isolated
Every credential, prompt, and policy is tenant-scoped. A request from one tenant can never reach another tenant’s configuration or data.
Residency-aware routing
The router honours your residency policy — a region-locked tenant is never routed to an out-of-region provider, BYOK or otherwise. TLS in transit; admin traffic HMAC-signed.
Coming next
Two surfaces building on what already ships.
The next things we ship slot into the same engine the three products above already use — same per-tenant LLM routing, same audit feed, same residency + isolation guarantees. Built once, consumed many times.
In progress
Knowledge Hub
The trust layer between your scattered knowledge and your agents. First it makes data ready — messy tickets, docs and PDFs become typed, confidence-scored, source-anchored objects, not random text chunks. Then it keeps that knowledge current — a freshness signal and verified-at timestamp on every answer, a verify-against-the-source step before an agent acts, and when two sources disagree you get the authoritative answer and the reason, not both.
- Agent-ready structuring — raw documents become typed, source-anchored objects with per-field confidence. Live today in Data Extractor
- Currency-scored retrieval — every answer carries freshness + confidence + verified-at, so agents can decide whether to trust it
- Verify-before-act — a dedicated path confirms against the source system before an agentic action runs
In development
MCP Gateway
One governed doorway between AI agents and your platforms. Point Velgent at your systems and get back a scoped MCP endpoint any agent host can call — Claude, Copilot, Cursor, or your own. Every tool call is policy-checked before it runs, redacted before it returns, and audited after it lands. One set of rules across every system.
- Compiled from your schema — typed tools from each platform’s data dictionary, custom fields included; agents never compose raw queries
- Fields that never leave — allowlists enforced at the source API, PII redaction on free text, untrusted content wrapped
- One neutral audit trail — who asked, as whom, what was returned, what was filtered — outside the platform being audited