Decision infrastructure for the AI era Discover Arbitrium

The decision authority layer for AI agents

Arbitrium — the threshold where automation stops and human authority takes over. HumanLayer routes high-stakes requests to a qualified Sentinel, then returns an auditable, documented decision artifact.

Are you a licensed expert? Become a Sentinel

Automation continues everywhere else, without breaking compliance. Only high-stakes requests cross the Arbitrium threshold.
Architecture 3 layers

Orchestration · Decision · Evidence

Target SLA From 30 min

Depending on the decision level

Audit trail Complete

Verdict, reasoning and timestamp

Built for sectors where accountability matters

Credit & Risk Legal & M&A Health & Diagnostics HR & Hiring Industrial compliance Insurance & Claims Spend & Procurement Cybersecurity & Access
GDPR Law 25 (Quebec) EU AI Act KYC / AML FDA / ISO traceability Zero-trust Exportable audit log Documented non-discrimination

01The problem

AI agents fail at the same point: the accountability moment.

Today, AI can automate 80–95% of a workflow… But as soon as you need to:

  • 01validate compliance
  • 02rule on a gray area
  • 03approve a spend
  • 04sign an act
  • 05issue an enforceable decision

…everything goes back “out of system”, manually, without SLA or traceability.

The brake on enterprise adoption isn’t execution. It’s the lack of authority.

Data collectionSTEP 1
✓ Automated
AI analysisSTEP 2
✓ Automated
Case preparationSTEP 3
✓ Automated
?
Decision & AccountabilitySTEP 4
Out of system
Manual No SLA No traceability
Share of a workflow AI can automate 80–95%
AI execution The accountability moment

Agentic AI is scaling

The context

1.3B AI agents in use (proj. 2028)
$307B AI software revenue (proj. 2027)
$143B GenAI spending (proj. 2027)
$423B Total AI spending (proj. 2027)

Projections 2027–2028. Amounts in USD. Sources: IDC (AI software revenue, GenAI/AI spending); Microsoft citing IDC (AI agents).

02Arbitrium

The threshold where automation stops.

Arbitrium is not a person, but a system threshold. Only high-stakes requests cross it: automation continues everywhere else, without breaking compliance. Pick a scenario and play with the parameters.

Agent request

“€42 expense report — client lunch, compliant with internal policy. Reimburse?”

8
96%
20
10

Automatic execution

Below the threshold: the agent continues its workflow. The action is logged.

Level
—
Sentinel
—
Target SLA
—
Deliverable
—
{ "arbitrium": false, "action": "auto_execute" }

Illustrative simulation. In production, each organization defines its own risk thresholds, automatic rejection rules and mandatory escalations.

03The solution

On-demand human decisions, API-native.

HumanLayer orchestrates human decisioning exactly where it’s indispensable. A three-layer system, from the AI-prepared case to the audit-ready decision.

  1. LAYER 01

    Orchestration

    Detect Arbitrium and route to a qualified Sentinel. The AI agent submits a complete case with context and supporting documents; HumanLayer routes it to a qualified Sentinel based on the required criteria.

    • Evaluates the level
    • Routes the request
    • Required criteria
  2. LAYER 02

    Decision

    A qualified Sentinel issues an explicit verdict: a structured decision, with comments and justification. HumanLayer does not automate the decision; it orchestrates the intervention of an accountable human.

    • Explicit verdict
    • Justification
    • Second opinion
  3. LAYER 03

    Evidence

    A timestamped, reasoned, audit-ready decision artifact. Back to the agent: the decision is logged, auditable and ready to execute.

    • Timestamp
    • Exportable log
    • Enforceable

Result: human authority triggered only at the right threshold.

Architecture flow

From AI case to enforceable decision

AI AgentPrepares the case
HumanLayer APIRoutes the request
Arbitrium thresholdEvaluates the level
SentinelDecides and justifies
DecisionLogged and enforceable

Result: AI can complete the workflow end-to-end without breaking compliance.

Designed with Legal, Risk and Compliance teams

04Who decides?

HumanLayer does not automate the decision.

It orchestrates the intervention of an identified, qualified, accountable Sentinel. Every decision is:

  • Assignedto a real person
  • Timestampeddate and time of the decision
  • Justifiedcomments and reasoning
  • Auditedexportable log, chain of evidence
What HumanLayer prevents

Automating decisions without an identified human accountable for them.

Decision record dec_01J9ZQ4M7K
Level 3 · Authority
Sentinel
S-0142 · Certified credit analyst
Timestamp
2026-03-12 · 14:32:07 UTC
Reasoning
Stable repayment history over 36 months; debt ratio within policy.
Audit
AI context, decision, timestamp, identity — exportable
Case prepared by AIBack to the agent · ready to execute

Illustrative example, fictitious data.

05Use cases

Where accountability matters.

HumanLayer integrates into every sector where a decision commits someone. AI prepares, the Sentinel decides.

Credit & Risk

“AI analyzed 200 credit files. For this customer, the score is in a gray zone (480). Approve the €50,000 loan?”

The AI

Prepares the complete file with history and recommendation.

The Sentinel

A certified credit analyst decides within 2 hours (custom SLA).

Enforceable decision + audit trail

06Developers

Decision-as-an-API.

HumanLayer integrates like Stripe or Twilio — but for decisions. Arbitrium is exposed as an escalation state in the system.

POST/api/v1/decisionsSubmit a case or a second opinion
GET/api/v1/decisions?id=Status, SLA and signed decision
GET/api/v1/decisions?agent_id=Agent history
GET/api/v1/decision-keyVerification public key

Webhooks

decision.completed decision.rejected decision.escalated sla.breached

Integration

REST / JSON Signed webhooks (HMAC) Signed decisions (Ed25519) Any agent framework
View the API docs
import requests

API = "https://humanlayer.systems/api/v1/decisions"
HEADERS = {"x-api-key": "hl_live_..."}

# Arbitrium threshold crossed: submit the case
created = requests.post(API, headers=HEADERS, json={
    "agent_id": "agent-credit",
    "request_type": "compliance",  # level 1, 30-min SLA
    "domain": "finance",
    "summary": "KYC: partial match. Validate?",
    "context_json": {"customer_id": "c_4521"},
    "callback_url": "https://your-app.example/hl",
}).json()

# Track it (or wait for the signed webhook)
res = requests.get(API, headers=HEADERS,
                   params={"id": created["request_id"]})
req = res.json()["request"]

if req["verdict"] == "approved":
    continue_workflow(req["decision"])  # Ed25519-signed
Ready · POST /api/v1/decisions

07Differentiation

Not a task marketplace. An accountability infrastructure.

What you buy is not human time: it’s verified authority, plugged into your agents, with the evidence to back it.

Verified Authority Network

A network of verified Sentinels

Qualified Sentinels, traceable accreditations. Qualifications and authorizations verified before any activation.

Decision-as-an-API

Plugged into your agents

Endpoints, formats, agent integrations.

Audit by Design

Evidence, by default

Evidence, logs, context, versioning.

Guaranteed SLAs

Minutes, not days

Guaranteed timelines, escalations, fallback.

Trust & compliance

Built for Legal, Risk and Compliance teams.

Verified identity

Qualifications and authorizations verified.

Audit log

Exportable, signed decisions (Ed25519).

Automatic escalations

SLA breaches, second opinions, reassignment.

Enhanced security

Mandatory MFA for Sentinels, encrypted secrets (AES-256).

Data minimization

Collection limited to what is strictly necessary.

Compliance

SOC 2 Type II certified hosting; processes aligned with GDPR and Law 25.

08Pricing

Priced per decision. Not per hour.

Pay for value, not time. Each level maps to a type of responsibility and expertise; up to 60% of every decision goes to the Sentinel who renders it.

Level 1 · Compliance validation

Compliance

Validation of rules and policies.

€20/ decision
  • Structured verification
  • 30-minute SLA
  • Full audit trail
  • Verified Sentinels
Examples
KYC/AMLGDPRSpend approvalInternal policies

Value: speed + accountability + audit

Get started
Level 3 · Decision authority

Authorized authority

Final act by a qualified person.

€250/ decision
  • Sign-off by an authorized professional
  • 4-hour or custom SLA
  • Assumed accountability
  • Signed, exportable proof
Examples
Legal sign-offCredit decisionMedical validationRegulatory approval

Value: assumed accountability + strong barrier to entry

Get started

Enterprise options: premium SLAs, dedicated pools, governance, integrations, reporting.

Get pricing

09Join the network

Are you an expert? Become a Sentinel.

Lawyer, physician, analyst, CISO… Monetize your decision authority by validating AI agents’ requests. Full flexibility, paid per decision.

LawyersPhysiciansAnalystsCISOsNotariesEngineers
Apply as a Sentinel

Diligent review of your application.

SSentinel portal · Legal
SLA 01:59:59
New request · Level 2dec_8Pq2

Clause §4.2.b — change of control

Out of 500 contracts analyzed, 12 include this clause. The AI flagged the critical points and attached the full context.

Legal & M&ACase prepared by AI
Decision recorded: timestamped, justified, archived.

Illustrative preview of the Sentinel portal. Click to decide.

Deploy AI agents… without losing control.

HumanLayer makes autonomy auditable, accountable and enforceable.

As AI advances, HumanLayer defines when not to automate.