Arcana Vertical Deep-Dive #1
AI Agents in AML
2025 Market Intelligence Report
How financial institutions are deploying AI agents for transaction monitoring, investigation automation, and SAR generation. Vendor displacement data, adoption signals, and practitioner intelligence.
214+
Respondents
Banking, Insurance, Fintech
Scope
March 2025
Published
by Arcana × Partner
01 — Market Brief
The Bottom Line
68% of financial institutions now use or pilot AI in AML — up from 41% in 2023. But only 22% have reached production scale. The market is stuck in pilot purgatory: high trial, low production.
False positive reduction remains the #1 objective (74%). Speed to SAR is emerging as the #2 driver as real-time payment mandates accelerate. Yet average institutions still run 95%+ false positive rates on TM alerts.
The market is consolidating around 3 tiers: legacy incumbents (NICE Actimize, SAS), modern platforms (Feedzai, Unit21), and AI-native upstarts (Sardine, Lucinity). Switching intent among legacy users is now 30%.
01 — Market Brief
The Bottom Line
68% of financial institutions now use or pilot AI in AML — up from 41% in 2023. But only 22% have reached production scale. The market is stuck in pilot purgatory: high trial, low production.
False positive reduction remains the #1 objective (74%). Speed to SAR is emerging as the #2 driver as real-time payment mandates accelerate. Yet average institutions still run 95%+ false positive rates on TM alerts.
The market is consolidating around 3 tiers: legacy incumbents (NICE Actimize, SAS), modern platforms (Feedzai, Unit21), and AI-native upstarts (Sardine, Lucinity). Switching intent among legacy users is now 30%.
68%
Use or pilot AI in AML
+27pp vs 2023
74%
Cite false positive reduction as #1 goal
$1.2M
Avg. annual AML AI spend (mid-market)
+45% YoY
38%
Actively evaluating new vendors
81%
Expect AML AI spend to increase in 12mo
The Arcana View
The Centaur model — agent-assisted, human-approved — is the pragmatic path.
While every other enterprise function races to deploy autonomous AI agents, AML teams are stuck in the validation gap: regulators haven't blessed agentic workflows, model risk teams can't explain them, and compliance officers won't sign off. The vendors who crack the Centaur workflow — AI agents that triage, draft, and recommend while humans approve and file — will own the next decade of financial crime tech. Fully autonomous agents will hit AML last. That's the opportunity: build the explainability and audit infrastructure now. We're watching Unit21 and Lucinity most closely here.
02 — Adoption Signal
Where the Market Stands
AI in AML: Current Status
n=214 // Financial services practitioners
68% have AI at some stage — but only 22% at broad production. Pilot purgatory.
Primary Objectives for AI in AML
"Why are you investing in AML AI?" (select all)
Implementation Model
42%
Primarily vendor solution(s)
31%
Mix of vendor + internal
18%
Primarily in-house
9%
SI-led (Accenture, TCS, etc.)
03 — The Displacement Cycle
Buy / Sell / Hold / Short
Peer-sourced performance data for the 6 most-cited AML AI vendors. Adoption rates, satisfaction scores, and switching intent from practitioners who actually use these tools.
Click a node to jump to its card →
NICE Actimize
Legacy Incumbent
34%
Adoption
62
Satisfaction
28%
Switching
Strengths
Weaknesses
Feedzai
Modern Platform
22%
Adoption
78
Satisfaction
8%
Switching
Strengths
Weaknesses
Unit21
Modern Platform
16%
Adoption
82
Satisfaction
5%
Switching
Strengths
Weaknesses
SAS
Legacy Incumbent
28%
Adoption
58
Satisfaction
32%
Switching
Strengths
Weaknesses
Sardine
AI-Native Upstart
8%
Adoption
85
Satisfaction
3%
Switching
Strengths
Weaknesses
Lucinity
AI-Native Upstart
6%
Adoption
88
Satisfaction
2%
Switching
Strengths
Weaknesses
NICE Actimize
Legacy Incumbent
34%
Adoption
62
Satisfaction
28%
Switching
Strengths
Weaknesses
Feedzai
Modern Platform
22%
Adoption
78
Satisfaction
8%
Switching
Strengths
Weaknesses
Unit21
Modern Platform
16%
Adoption
82
Satisfaction
5%
Switching
Strengths
Weaknesses
SAS
Legacy Incumbent
28%
Adoption
58
Satisfaction
32%
Switching
Strengths
Weaknesses
Sardine
AI-Native Upstart
8%
Adoption
85
Satisfaction
3%
Switching
Strengths
Weaknesses
Lucinity
AI-Native Upstart
6%
Adoption
88
Satisfaction
2%
Switching
Strengths
Weaknesses
04 — Capital Flow
Where the Money Goes
Annual AML AI Spend Distribution
Median: $1.2M/yr (mid-market financial institutions)
12-Month Spend Outlook
"How do you expect AML AI spend to change?"
81%expect spend to increase in 12mo
05 — Heat Map
Technology Heat Index
Proprietary heat scores based on adoption velocity, investment intent, and practitioner excitement. Higher = faster adoption trajectory.
92
AI-powered alert triage
Mainstream86
SAR narrative generation (LLM)
Rapid adoption78
Network / entity resolution
Growing72
Autonomous agent investigation
Early pilot68
Real-time transaction scoring
Established64
Sanctions screening AI
Growing58
Synthetic identity detection
Emerging42
Behavioral biometrics for AML
Niche06 — Risk Register
What’s Holding Teams Back
WARNING: Regulatory uncertainty / model risk
Cited by 64% of respondents as a primary blocker
WARNING: Data quality & integration
Cited by 58% of respondents as a primary blocker
WARNING: Explainability / auditability
Cited by 52% of respondents as a primary blocker
Biggest Blockers to Deeper AI Adoption in AML
"What prevents you from doing more with AI in AML?"
07 — The Dark Pool
Practitioner Intelligence
Anonymized field intelligence from survey respondents and Research tier members. What practitioners say when the vendors aren’t in the room.
"We built our own LLM pipeline for SAR narratives. Saved 40% of analyst time. But model risk won't let us auto-file — every narrative still gets human review."
— Head of Financial Crime, T█p 10 █S B█nk
"Our legacy vendor wants $3M/yr for 'AI features' that are basically logistic regression with a new UI. We're piloting Unit21 on the side."
— VP of Compliance Technology, R█g██n█l B█nk
"The real bottleneck isn't the AI — it's getting clean data out of our core banking system from 1997. No model can fix garbage in, garbage out."
— Director of AML Operations, █ns█r█nc█ C█mp█ny
"Lucinity's copilot is the first AML tool my investigators actually like using. That matters more than any benchmark score."
— BSA Officer, D█g█t█l B█nk
$ cat /intel/field-report-001.log
"We built our own LLM pipeline for SAR narratives. Saved 40% of analyst time. But model risk won't let us auto-file — every narrative still gets human review."
— Head of Financial Crime, T█p 10 █S B█nk
SIG:343$ cat /intel/field-report-002.log
"Our legacy vendor wants $3M/yr for 'AI features' that are basically logistic regression with a new UI. We're piloting Unit21 on the side."
— VP of Compliance Technology, R█g██n█l B█nk
SIG:534$ cat /intel/field-report-003.log
"The real bottleneck isn't the AI — it's getting clean data out of our core banking system from 1997. No model can fix garbage in, garbage out."
— Director of AML Operations, █ns█r█nc█ C█mp█ny
SIG:906$ cat /intel/field-report-004.log
"Lucinity's copilot is the first AML tool my investigators actually like using. That matters more than any benchmark score."
— BSA Officer, D█g█t█l B█nk
SIG:27008 — Trade Ideas
Long / Short / Hedge
Unit21 / Lucinity
High satisfaction, low switching intent, strong upward trend
SAR narrative automation
Fastest ROI, clearest regulatory path, 86 heat score
Legacy on-prem implementations
High cost, low agility, 30%+ switching intent
Fully autonomous agent workflows
Regulatory uncertainty too high — pilot purgatory for 2+ years
Keep SAS / Actimize for regulatory reporting
Maintain compliance continuity while running modern platform pilots
API-wrappers for detection layer
Use modern ML via API overlay without ripping out core infrastructure
Analyst Notes
Start with alert triage, not autonomous agents
Alert triage AI has the highest heat score (92) and the clearest regulatory path. Reduce false positives by 50-70% before attempting agentic workflows.
Pilot SAR narrative generation now
LLM-generated SAR narratives are the fastest ROI in AML AI. But keep human-in-the-loop — regulators aren't ready for auto-filing. The Centaur model wins here.
Evaluate modern platforms alongside incumbents
Switching intent among legacy vendor users is 30%. If you're in a renewal cycle with NICE Actimize or SAS, run a parallel evaluation with Unit21 or Feedzai.
Budget for 45%+ spend increase
81% of respondents expect AML AI spend to increase. The median mid-market institution spends $1.2M/yr — plan for $1.7M+ by 2026.
Invest in explainability infrastructure
The #1 blocker to deeper AI adoption is regulatory uncertainty around model explainability. Build audit trails and model documentation before you need them.
Methodology
This report is based on Arcana’s Q1 2025 Buyer Survey, with 214 respondents from banking, insurance, and fintech organizations. Respondents are VP+ level practitioners directly involved in AML technology decisions.
Vendor data is peer-sourced: adoption rates, satisfaction scores, and switching intent reflect what practitioners report, not vendor claims. Heat index scores are proprietary, calculated from adoption velocity, investment intent, and practitioner sentiment.
Supplemented with 12 structured interviews with AML practitioners at Research tier member organizations.
About This Report
This is a co-branded publication by Arcana and [Partner]. Arcana maintains full editorial independence. Partner branding does not influence methodology, data, or recommendations.
Arcana Research
arcana-advisors.com
© 2025 Arcana Advisors. All rights reserved. This report contains proprietary data and analysis. Do not distribute without permission.