Governed Agentic Operations

AI agents that do your
operational work

Matching, investigation, and exception handling — governed, auditable, and embeddable via API.

workflow.json

{
  "name": "invoice_reconciliation",
  "actions": [
    { "type": "matcher", "matchOn": ["invoice_id"], "tolerance": 0.02 },
    { "type": "loop", "mode": "react", "objective": "Investigate exceptions" },
    { "type": "PbotApproval", "comment": "Review recommendation" }
  ]
}

Core Architecture

Graduated exception handling

One pipeline. Three layers. Each handles the cases it's best suited for.

80%

Deterministic Matching

Exact keys, numeric tolerance, date windows, fuzzy text. No AI. No cost.

matcher primitive

15%

AI Investigation

Bounded ReAct agents investigate exceptions with declared tools, iteration caps, and stuck detection.

react loop

Human Judgment

Edge cases escalate with full context. The human decides. Their decision enters the audit trail.

PbotApproval

AI as Compiler

Describe it. Compile it. Run it.

AI generates the spec upfront. Execution is deterministic. Enterprises audit specs, not vibes.

Define in plain language

Describe the workflow. The AI compiler generates a deterministic JSON specification — the branches, conditions, escalation paths.

Execute deterministically

The compiled spec runs exactly as written. No improvisation. No hallucinated steps. Inspectable and versionable before any data is touched.

Reason within boundaries

When exceptions need investigation, bounded agents reason with declared tools and iteration caps. Every thought and action captured in the trace.

Governed Orchestration

Three deployment patterns

The agent decides what to do. The spec defines what it's allowed to do.

Pattern A

Agent as Workflow Step

One reasoning step inside a deterministic pipeline. Match 10,000 invoices, hand 47 exceptions to an agent.

Step 1 matcher → auto-match
Step 2 react → investigate
Step 3 approval → human reviews

Pattern B

Agent as Trigger

Receives unstructured input, reasons about it, dispatches the right workflow. Intelligent routing without rules.

Input "incoming document"
Agent → identifies as tax form
Runs tax_processing_workflow

Pattern C

Agent as Orchestrator

Coordinates multiple workflows. Decides sequence, synthesizes results, escalates when confidence drops.

Agent → run(kyc_verification)
Agent → run(sanctions_screening)
Agent → conflict → pause_for_human

Developer Experience

One API. Full pipeline.

Define workflows in JSON or generate from natural language. Register actions once, use everywhere.

Multi-criteria matching with tolerance, date windows, fuzzy text
ReAct agents with structural permissioning and reasoning traces
Human-in-the-loop as a first-class primitive
Multi-tenant isolation per organization by default
OAuth integrations for Gmail, Slack, Outlook

agent.json

{
  "objective": "Investigate unmatched invoices",
  "tools": [
    "lookup_erp",
    "check_payment_status",
    "gmail_send",
    "__pause_for_human__",
    "__complete__"
  ],
  "max_iterations": 10,
  "on_stuck": { "action": "escalate" }
}

Differentiation

What nobody else ships

Built-in data matching

Multi-criteria matching engine as a first-class primitive. Exact keys, numeric tolerance, date windows, fuzzy text — structured exceptions that feed directly into investigation.

vs. every other platform → "build it yourself"

Governed agent orchestration

AI agents that reason within structural constraints. Tool permissioning, iteration bounds, stuck detection, checkpoint/resume, human escalation — one runtime.

vs. LangGraph → no governance · vs. Temporal → no reasoning

AI as compiler, not runtime

AI generates the workflow spec upfront. Execution is deterministic. Auditors inspect the spec, trace every decision, reconstruct reasoning after the fact.

vs. runtime agents → powerful but unauditable at scale

Embeddable infrastructure

API-first. Multi-tenant by default. SaaS platforms embed Hyphen into their products. Your brand, your customers, our engine.

vs. agent platforms → they own the customer relationship

Use Cases

Where operations meet intelligence

Financial Reconciliation

Invoice-to-payment matching, bank reconciliation, intercompany settlement. Auto-reconcile, AI-investigate, human-review.

Insurance Claims

Claims-to-policy matching, duplicate detection, coverage verification. Graduated from auto-adjudicate to adjuster review.

Healthcare Revenue Cycle

Remittance-to-claim matching, denial management, underpayment detection. AI agents that analyze denial codes and recommend appeals.

Supply Chain

PO-to-invoice verification, goods receipt reconciliation, vendor compliance. Flag quantity and pricing discrepancies automatically.

KYC & Compliance

Identity verification, sanctions screening, adverse media. Auto-clear low-risk, AI-investigate medium, human-review high.

SaaS Embedding

Vertical SaaS platforms embed Hyphen to add governed agentic operations to their own products via API.

AI agents that do youroperational work