Early Access Program is now open. Apply today →

EARLY ACCESS PROGRAM — NOW OPEN
SPENDSHARK — AGENTIC EDITION

The Next Evolution of
SpendShark.

SpendShark has upgraded from a rule-based detection tool into an autonomous, agentic investigation system. We took our proven enterprise data pipeline and injected a coordinated team of AI agents directly into the decision engine.

The same robust enterprise controls you rely on, now powered by agentic intelligence.

Scroll to explore
The Proven Foundation

Built on a 6-Stage Enterprise Pipeline

Before we talk about AI Agents, let's talk about the foundation. SpendShark already processes billions in transactions through a highly robust data pipeline connecting your ERPs to our core Inspectors. This foundation isn't changing. It's getting an upgrade.

1. Feeding The Shark

Continuous data ingestion from ERPs, Connectors, or Manual uploads.

2. Data Pipeline

Qualifying, enriching, and standardizing raw transaction data.

3. Core Inspectors

The analytical engine of SpendShark. This is where Agentic Intelligence is injected. Algorithms, rules, and AI Agents collaborate to generate Confidence Scores and KPIs based on deep investigation.

ML Algorithms AI Agents Metrics & KPIs

4. Insights Database & Dashboard

Every decision, pattern, and confidence score flows into the central SpendShark DB. It becomes a living asset utilized by AP Specialists, Data Scientists, and Chatbots.

5. Continuous Feedback loops

Ground truth from Recovering Specialists continuously trains the models.

6. Enterprise Integrations

Directly connected to Workday, Oracle, Infor, and robust 3rd party APIs.

20+
Specialized Agents
50+
Detection Techniques
<1s
Decision Time
100%
Audit Trail Coverage

Upgrading the Inspector Layer

The foundation above is proven. But instead of relying on static rules that throw false positives, we've wrapped our battle-tested Inspectors with autonomous AI agents. Here is how the shift from rules to agents changes the game.

Previous SpendShark
SpendShark — Agentic Edition
Architecture
Rule-based duplicate checks
Multi-agent investigation framework
Analysis Speed
Batch or near-real-time
Real-time, event-driven — <1 second
Decision Logic
Single fraud or risk score
Context-aware, evidence-based reasoning
False Positives
High alert volume
Aggressively reduced via recurrence & vendor resolution
Human Review
Heavy manual review process
Human-in-the-loop only when confidence is low
Detection Methods
Static detection logic
8 parallel techniques: exact, fuzzy, Benford's, velocity & more
Explainability
Score only, no reasoning
Full reasoning trail, agent contributions, confidence level
Learning
Static — no self-improvement
Continuously learning detective agents, always running

SpendShark's foundation remains — its intelligence has evolved.

From Detection to Investigation

Traditional payment controls rely on static logic and siloed alerts. SpendShark — Agentic Edition wraps every core function in an intelligent agent that actively investigates, collaborates, and resolves.

PREVIOUS SPENDSHARK
  • Rule-based duplicate checks
  • Batch or near-real-time analysis
  • Single fraud or risk score
  • High alert volume & manual review
  • Static detection logic
  • After-the-fact alerts
SPENDSHARK — AGENTIC EDITION
  • Multi-agent investigation framework
  • Real-time, event-driven decisions
  • Context-aware, evidence-based reasoning
  • Aggressively reduced false positives
  • Continuously learning agent network
  • Prevention before money moves

"SpendShark is a digital fraud investigation team — not a fraud score."

AGENTIC AI 101

What exactly is an Agent?

In enterprise software, an "Agent" is an AI system given a specific goal, a set of tools, and the autonomy to figure out how to achieve that goal.

Standard Generative AI

You ask a question. It predicts the next word to generate an answer. It has no tools and acts in a single step.

Agentic AI (Goal-Driven)

You give it a complex goal (e.g., "Verify if this invoice is legitimate"). It:

  • Plans a series of investigative steps.
  • Uses tools (queries databases, cross-checks ERP records).
  • Reasons about the results it found.
  • Collaborates with other specialized agents to reach a final conclusion.
THE MISSION
Investigate Payment Request #49281

Detective Agent

Reasoning Engine

ERP Query Tool
Web Search Tool
Math Tool

The Agentic Pipeline

Every payment passes through a coordinated team of AI agents — each applying a different investigative lens. They don't vote. They argue. Evidence is weighed. Conflicts are surfaced. A decision is reached.

Live Payment Event
AGENT 0 · Real-Time Ingestion

Real-Time Ingestion Agent

Listens to live payment events from ERP, AP systems, or event streams. Normalizes data instantly.

T+0.0s
AGENT 1 · Signal Interpretation

Signal Interpreter Agent

Parses intent, urgency, and key fields. Assigns an initial risk baseline before deeper analysis begins.

T+0.1s
Parallel Investigation — Agents 2 through 5
AGENT 2
Similar Invoice Detection

Fuzzy-matches invoice numbers, amounts, and date windows. Catches duplicates even when IDs are altered.

AGENT 3
Vendor Resolution

Resolves vendor aliases, shared tax IDs, and subsidiary relationships. Eliminates vendor-identity noise.

AGENT 4
Recurrence Pattern

Identifies legitimate recurring payments (monthly rent, subscriptions). Applies variance tolerances.

AGENT 5
Outlier Detection

Spots anomalies within recurring patterns: amount spikes, timing deviations, frequency changes.

T+0.3s — T+0.6s
AGENT 6 · Evidence Synthesis

Risk Aggregation Agent

Synthesizes all agent evidence using weighted reasoning. Business-policy aware. Conflicts are surfaced, not hidden.

T+0.7s
AGENT 7 — CORE ProtoVision Decision Engine

Decision Agent

The final arbiter. Makes the actual decision based on risk score, recurrence confidence, business rules, and historical outcomes. Always explains its reasoning.

Auto-Approve Block Escalate
T+0.9s
AGENT 8 Conditional — low confidence only

Human-in-the-Loop Gate

Only triggers when confidence falls below threshold. Pauses the payment, and requests human approval.

AGENT 9
Outcome Validation

Confirms what actually happened: Was payment reversed? Was vendor contacted? Was dispute raised?

AGENT 10
Learning & Memory

Updates confidence models, vendor aliases, recurrence patterns. Every decision makes the system smarter.

Full pipeline decision:

≤ 1 Second

Before payment is released

EXISTING FEATURE — ELEVATED

Inspectors Power the Agents

The Inspectors are one of SpendShark's strongest proven features. In the Agentic Edition, they don't get replaced — they become the detection engine each agent calls, with their results interpreted in context rather than fired as isolated alerts.

Previous SpendShark

Inspector Alert Human

Every inspector fires independently. Every alert requires manual review.

Agentic Edition

Inspector Agent Decision

Inspectors run inside agents. Humans review only when confidence is low.

Inspector
Owned by Agent

Exact Match

Company, vendor, invoice, date & amount all identical

Agent 2 — Invoice Similarity

Fuzzy Match

Levenshtein distance — catches altered invoice IDs

Agent 2 — Invoice Similarity

Clustering & LLM

Semantic similarity — catches meaning-level duplicates

Agent 2 — Invoice Similarity Detective Agents

Velocity Check

Rapid-fire invoices from the same vendor in a time window

Agent 4 — Recurrence Pattern

Benford's Law

Leading-digit statistics — catches manipulated numbers

Agent 5 — Outlier Detection

Round Amount

Perfectly round numbers — often estimates or fraud signals

Agent 5 — Outlier Detection

Weekend Posting

Saturday/Sunday invoice timing — unusual for B2B

Agent 5 — Outlier Detection

ML Anomaly Detection

Trained on historical data — finds statistical outliers

Agent 5 — Outlier Detection Detective Agents
All agent evidence flows into

AGENT 6

Risk Aggregation

AGENT 7 — ProtoVision Core

Decision Agent

✓ Approve ✗ Block ⚠ Escalate

The key difference

In the previous version, an inspector firing on a recurring monthly payment would generate an alert — and a human would spend time reviewing it, only to clear it. In the Agentic Edition, the Recurrence Pattern Agent sees that same inspector signal, cross-references 14 months of payment history, and suppresses the false positive automatically — with confidence 94% and a full explanation. No human time wasted.

ALWAYS RUNNING IN THE BACKGROUND

The Detective Agents

These agents aren't triggered by a single payment. They continuously study the data — like a senior fraud analyst who never sleeps.

Behavioral Drift Agent

Detects changes in vendor or payer behavior over time — subtle shifts that only emerge across hundreds of transactions.

Temporal Pattern Agent

Studies timing anomalies across the full payment history — payments too soon, too frequent, or on unusual schedules.

Amount Deviation Agent

Spots subtle overpayments hidden within normal ranges — 2%, 5%, or 8% above historical means that rules would never catch.

Pattern Discovery Agent

Continuously discovers new recurring payment patterns, emerging fraud tactics, and vendor behavior shifts — then propagates findings to the live pipeline.

What the Detective Agents Produce

New hypotheses

Updated pattern libraries

Improved agent heuristics

Reduced false positives

"ProtoVision discovered a new quarterly billing pattern for this vendor."

Meet the Agents

Each agent is independent, opinionated, and specialized. They don't average results — they present evidence and let the Decision Agent reason through conflicts.

Real-Time Ingestion Agent

Connects to your payment systems and streams live transactions into the analysis pipeline. Event-driven, zero batch delay.

Agent 0 · Entry Point

Signal Interpreter Agent

Classifies, normalizes, and scores each payment event before it reaches the specialist agents. Sets the analytical baseline.

Agent 1 · Intake

Invoice Similarity Agents

Exact and fuzzy matching across invoice IDs, amounts, and date windows. Catches duplicates even when identifiers are deliberately altered.

Agent 2 · Exact + Fuzzy Match

Vendor Resolution Agent

Detects duplicate vendors, legal vs trade name aliases, shared bank accounts, and subsidiary relationships. Massive reduction in false positives.

Agent 3 · Identity Resolution

Recurrence Pattern Agent

Distinguishes legitimate recurring payments (subscriptions, rent, payroll) from risky repetitions. Applies variance tolerances and cadence models.

Agent 4 · Context Analysis

Outlier Detection Agent

Spots anomalies within recurring patterns: amount spikes, timing shifts, frequency changes. Because recurring does not mean safe.

Agent 5 · Anomaly Detection

Risk Aggregation Agent

Synthesizes evidence from all agents using weighted, context-aware reasoning. Not a simple average — conflicts are surfaced explicitly.

Agent 6 · Evidence Synthesis

Decision Agent — Core

The ProtoVision decision engine. Auto-approves, blocks, or escalates based on aggregated evidence, business rules, and learned history. Always explains its reasoning.

Agent 7 · ProtoVision Core

Human-in-the-Loop Gate

Automatically pauses payments when confidence is low. Presents reasoning clearly. Human approval is logged and used to train future decisions.

Agent 8 · Trust & Compliance

Outcome Validation Agent

Closes the loop by confirming the real-world result of every decision — reversal, vendor contact, dispute, or clean payment.

Agent 9 · Validation

Learning & Memory Agent

Updates vendor aliases, recurrence models, and confidence weights after every outcome. The system improves without constant reconfiguration.

Agent 10 · Continuous Learning

Detective Learning Agents

Always-on background agents that study historical data to discover new fraud tactics, emerging patterns, and behavioral changes — before they become a problem.

Background · Always Active

The Directive

Tools & Capabilities

Live Reasoning Trace

T-ANOMALY DETECTED
ALREADY BUILT & RUNNING

The Platform in Action

The agents aren't theoretical. They're already wired into the SpendShark dashboard. Explore the platform to see how each capability works in practice.

More to Explore in the Dashboard

The platform includes Analytics, Predictions, ROI & Value, Vendor Risk, AP Performance, Credits, and Integration Setup.

REAL-TIME BY DESIGN

Decisions at Payment Speed.

SpendShark connects directly to your payment systems and analyzes transactions in real time — using agentic AI to stop fraud before money moves.

Event-driven ingestion

No batch delays. Every payment is analyzed the moment it enters the system.

Parallel agent execution

Agents 2 through 5 run simultaneously, delivering depth without sacrificing speed.

Configurable decision modes

Payment is temporarily paused while agents analyze. Released automatically if approved — blocked if not. Recommended default. Best balance of safety and speed.

Decision must complete before the payment is authorized. Highest protection — ideal for high-value or first-time vendors. May add latency for edge cases.

Payment is released immediately; agents flag anomalies for audit and recovery. Best for organizations prioritizing payment continuity, with recovery workflows in place.

Deterministic fail-safe behavior

Safe under latency pressure. No decision is worse than a cautious fallback.

spendshark — live analysis feed
LIVE

Waiting for payment event...

Explainable. Auditable. Trusted.

Every decision SpendShark makes includes the full reasoning trail — so your finance team, auditors, and executives can always understand why.

Full Reasoning Trail

Every decision includes which agents contributed, what evidence was considered, and why a specific action was recommended.

Confidence Scoring

Every recommendation includes a confidence level. Low confidence automatically triggers human review — autonomy is earned, not assumed.

Human-First Design

Automation is gated by confidence. Humans remain in control. Every approval or override is captured and improves future decisions.

Sample Decision Output
Decision AUTO-APPROVED
Confidence 91%
Primary Reason Matches known recurring vendor pattern observed 14 times over 12 months. No historical disputes. Amount within normal variance (±3%).
Contributing Agents
Invoice Similarity ✓ Vendor Resolution ✓ Recurrence Pattern ✓ Outlier Agent ⚠ mild Detective Agent ✓ learned

Why This Matters

Built on years of real-world experience in payment systems and enterprise data platforms. This reflects lessons from production — not experiments.

Fewer False Positives

Finance teams see dramatically fewer alerts without losing detection power.

Faster Decisions

Sub-second to seconds-level payment decisions — before money leaves the account.

Less Manual Review

AP teams trust the automation. Human review reserved for genuinely uncertain cases.

Continuous Improvement

The system gets smarter over time without constant reconfiguration or retraining cycles.

Finance Teams

"Fewer alerts"

Auditors

"Explainable decisions"

Engineers

"Composable, extensible"

Executives

"Gets smarter over time"

BEYOND DETECTION

The Complete Solution

SpendShark goes beyond just stopping bad payments. Every decision feeds an enterprise Insights Database, empowering our Recover Specialists to act.

The Insights Database

Every transaction analyzed, every hidden relationship discovered, and every confidence score generated isn't just a point-in-time decision—it becomes a permanent, queryable asset.

  • Vendor Risk Profiling: Know exactly which vendors have the highest error rates.
  • Macro Trends: Identify systemic process failures across business units.
  • AI-Ready Data: Directly queryable by Chatbots and BI tools for instant reporting.

Recover Specialists

Detecting the error is only half the battle. Our dedicated team of Recover Specialists uses the Insights Database to take action, returning lost capital directly to your bottom line.

  • Evidence-Backed Claims: Specialists approach vendors armed with the AI's exact reasoning trail.
  • Vendor Etiquette: We handle disputes professionally, carefully preserving your supplier relationships.
  • Closing the Loop: Resolution data flows back into SpendShark to retrain the core models.

Frequently Asked Questions

Common questions from Enterprise IT and Finance leaders.

EARLY ACCESS PROGRAM

Work Directly With Our Team.

SpendShark — Agentic Edition is available to select organizations through our Early Access Program. Early Access partners work directly with the iTK team on implementation, validation, and configuration — shaping the platform as it scales.

Dedicated iTK Team Access

Custom Configuration

Influence the Roadmap

Enterprise-Grade Support

Built on years of real-world experience in payment systems and enterprise data platforms. iTK Technologies has analyzed over $2.2 trillion in transactions across healthcare, retail, and financial services.

Accepting a limited number of Early Access partners. Contact us to discuss fit and timeline.