How Phyvant Works

One platform. All five context layers. Executable output. No consultants.

Enterprise AI deployment stalls on documentation. Process mining tells you what to automate; it doesn't produce the machine-executable specs agents need. Phyvant closes that gap with a continuous loop.

01/055 layers

Observe — mine history, watch the living workflow.

Current tools are single-layer specialists. Celonis sees ERP event logs. Mimica records desktop activity. Scribe captures docs. None see the full picture. Phyvant captures all five in-house:

Process context— ERP connectors normalize event streams into a unified work-session timeline.
Task context— Browser extension and desktop agent record off-system clicks and spreadsheet work.
Document context— Policies, SOPs, and tribal knowledge ingested as typed artifacts.
Operational context— Every tool call logged, giving a real-time picture of system state.
User & situational context— Expert corrections and judgment calls, captured inline during normal work.
01Process Context

End-to-end workflows from SAP, Oracle, NetSuite event logs

SAPOracleNetSuite
02Task Context

Desktop clicks, copy-paste, spreadsheet work process mining misses

BrowserDesktop
03Document Context

Policies, SOPs, tribal knowledge ingested as typed artifacts

ConfluenceSharePointNotion
04Operational Context

Every API call and tool action logged for real-time system state

APIsWebhooks
05User & Situational

Expert corrections, judgment calls, and edge-case decisions inline

ChatInline
the layerfoundation

Knowledge graphs — the substrate the loop runs on.

What we observe doesn't sit as raw events. It resolves into a typed graph: customers, vendors, SKUs, contracts, invoices, and the relationships between them. The inferencer reads from the graph. Agents reason over it. Corrections write back to it.

Inside Phyvant's graph

EntitiesCustomers, vendors, contracts, orders, SKUs, invoices, employees, regions — resolved to one canonical node across SAP, Salesforce, NetSuite.
RelationshipsTyped edges (has_contract, supplies, billed_as, contains) carry meaning a vector embedding can't.
ProvenanceEvery node and edge carries the source system, timestamp, and the observation that produced it.
ConflictsDuplicate vendors, mismatched SKUs, currency drift surface as graph anomalies before agents act on them.

Why a graph, not a vector store

One truth across silos— The same customer in Salesforce, SAP, and NetSuite resolves to one node. No more "is this the same vendor?"
Reasoning beats retrieval— Agents traverse relationships ("invoices for EMEA orders under contract A-14") instead of guessing from text similarity.
Conflicts caught before action— Dangling references and duplicates show up as graph anomalies, not silent agent failures downstream.
Audit comes for free— Every decision an agent makes traces back to the entities and edges it touched. Provenance is the data model, not a log.

Sample subgraph

ACME Corp · EMEA

resolvedconflict
has_contractplacedlocated_inreferencescontainsbilled_assuppliesconflicts_withCustomer · ACMEEMEAContract A-14Order ORD-889Vendor · NovaPartsSKU-7741SKU-7741-AInvoice INV-2234
8 entities · 8 edges1 unresolved conflict
02/05infer

Infer — spec inferencer writes AOPs, not SOPs.

The orchestrator assembles work sessions from all five layers and hands them to the spec inferencer. It detects recurring patterns and outputs machine-readable JSON that agents consume directly.

Decision rulesAgent-operating proceduresEntity mappingsProcess specs
// agent-operating procedure
AOP-12 tenant-scope billing
when invoice.currency ≠ tenant.currency
require R-441 · R-119
then decide(confidence ≥ 0.92)
else escalate → inbox
// output: JSON · exportable to n8n, LangChain, CrewAI
03/05verify

Verify — test against ground truth before promotion.

Candidate specs are tested against correction history and gold-standard references. Passing specs go live. Failing ones return to the inferencer with a clean diff.

spec
tests
coverage
status
AOP-41
2,061 / 2,070
99.6%
active
S-84
1,522 / 1,530
99.5%
active
AOP-14
891 / 912
97.7%
review
R-227
444 / 445
99.8%
active
04/05execute

Run — agents execute against the spec, inside your VPC.

The platform owns execution end-to-end. The LLM never holds tool definitions, eliminating prompt-injection risk. Every action — every API call, every escalation, every decision — has provenance back to the spec it ran against and the observation that produced the spec.

In-flight trace
AOP-12 · tenant-scope billing
14:32:01
AOP-12 invoked
14:32:01
R-443 evaluated (rules engine)
14:32:02
confidence 0.94 → decided
14:32:02
action posted → SAP (FI-AR)
05/05sharpen

Sharpen — every correction is an edge, not a ticket.

Experts don't document. They correct. When a domain specialist fixes a classification or overrides a mapping during normal work, the correction writes back to the live spec at 100% confidence. Agents downstream pick it up on their next run.

C-1Invoice currency rule: use spot rate when document currency does not match functional currency
C-2Entity mapping: two supplier records merged to one canonical vendor
C-3Tax classification: line reclassified to the correct regime after AP review
100%90%80%70%60%64%78%89%94%97%wk 1wk 6wk 12wk 18wk 24

See it working on your data.

We'll connect to your systems and show you what Phyvant infers from your data in the first week — no embedded engineers, no multi-month implementation, no retraining cycle.