ROI

—

Net Annual Benefit

—

Payback Period

—

Solution Cost

—

Analyst Capacity Freed

—

— hrs/yr

IRR (directional)

—

3-Year NPV

—

Live

Fraud Detection AI · Business Case

Your Numbers.
Your Real ROI.

Model fraud loss avoided, chargeback savings, false positive recovery, labor savings, and regulatory exposure. Solution cost shown in full — no hidden denominator.

Start Here — Institution Profile

Pre-filled with industry benchmarks. Change any number and results update instantly above.

Deployment scope

Implementation per additional typology ($) Tuning · thresholds · HITL playbook · validation

Scope moves both sides of the ratio. Additional typologies add benefit and add implementation cost. A model where only the numerator grows will produce a number no procurement team believes. The default is the core ten, which reproduces the published baseline exactly.

AI inference cost

Simple case — tokens 30k in + 3k out

Standard case — tokens 80k in + 8k out

Narrative case — tokens 200k in + 20k out

Blended rate — $ per million default $4.10/M · claude-sonnet-4-6 · yours to change

—

AI run cost · per year

—

of total solution cost

Tokens per case are a property of the workload. The rate is a property of your commercial agreement — change it and everything recomputes. Inference is disclosed because it should be, not because it drives the case.

Benefit composition

Hard benefits only — exclude fraud loss avoided

—

Hard · —

Labour no longer worked, chargeback and return fees not incurred, false-positive revenue retained, penalties not assessed. Realised.

—

Counterfactual · —

Fraud loss avoided — loss that would have occurred, not cash that appears. Rests on baseline detection and prevention factor, both of which are institution facts rather than vendor claims.

—

Institution Asset Size Drives platform tier

Determines base platform fee

Institution Type Auto-fills case volumes

Sets scenario volume defaults

Fraud Analyst Hourly Rate ($) Benchmark: $85

Fully-loaded including benefits & overhead

Senior Analyst / BSA Officer ($) Benchmark: $110

HITL review, SAR sign-off, complex cases

Fraud Analyst FTE Count Your team size

Caps labor savings at team capacity

Annual Transaction Volume (M) For FP recovery

Total monthly transactions × 12 ÷ 1M

Monthly Case Volumes — 10 Fraud Types

Pre-filled from institution type · change any volume · leave unused types at 0

—▼

Fraud type	Type	Vol/mo ✦	Manual hrs	AI hrs	Avg loss/case ($)	Detection lift (%)
ATO Wire Fraud	SAR	Benchmark: 12	1.5 hrs	9 min HITL	$47,500	30%
BEC Wire Fraud	SAR	Benchmark: 4	2.5 hrs	12 min HITL	$185,000	35%
Synthetic Identity	HITL	Benchmark: 8	1.5 hrs	11 min HITL	$18,200	28%
Mule Network	SAR	Benchmark: 5	4.0 hrs	21 min HITL	$47,200	32%
Pig Butchering NEW	SAR	Benchmark: 2	3.0 hrs	18 min HITL	$340,000	40% (speed)
CNP / Card Fraud NEW	CB	Benchmark: 180	0.75 hrs	5 min HITL	$280	35%
ACH Origination NEW	SAR	Benchmark: 6	2.0 hrs	12 min HITL	$62,000	40%
Elder Exploitation NEW	SAR	Benchmark: 3	3.0 hrs	17 min HITL	$92,000	30%
New Account Fraud NEW	HITL	Benchmark: 10	1.5 hrs	11 min HITL	$14,500	28%
Check Fraud NEW	HITL	Benchmark: 8	2.0 hrs	13 min HITL	$6,800	32%
Zelle — Unauthorised	REG E	Benchmark: 22	0.9 hrs	7 min HITL	$3,200	25%
Zelle — Authorised Scam	SAR	Benchmark: 31	0.7 hrs	6 min HITL	$3,450	12%
RTP / FedNow	SAR	Benchmark: 4	1.1 hrs	8 min HITL	$18,400	30%
First-Party Dispute	DISPUTE	Benchmark: 38	0.8 hrs	7 min HITL	$2,180	18%
Card-Present / ATM	REG E	Benchmark: 4	1.2 hrs	9 min HITL	$31,600	22%
Wire Recall (GPI)	GPI	Benchmark: 1	1.0 hrs	8 min HITL	$264,000	10%
Deepfake / Voice Clone	SAR	Benchmark: 1	1.4 hrs	11 min HITL	$410,000	35%
Insider Fraud	SoD	Benchmark: 1	2.1 hrs	18 min HITL	$88,400	28%
Crypto Off-Ramp	§5324	Benchmark: 2	1.6 hrs	12 min HITL	$127,000	26%
Merchant Bust-Out	RESERVE	Benchmark: 1	1.3 hrs	10 min HITL	$340,000	24%
Check Kiting	REG CC	Benchmark: 2	1.5 hrs	11 min HITL	$71,000	20%

✦ Detection lift = % improvement in cases caught. For pig butchering this is an intervention speed factor (% of victim-initiated wires intercepted before clearing), not a standard detection rate — default 40% is conservative.

Fraud Loss Avoided

The largest ROI lever for most fraud buyers — previously absent from this calculator

—▼

Loss avoided is computed per scenario from the case mix table above (vol × avg_loss × detection_lift). The inputs below let you tune the pig butchering intervention speed factor separately — it works differently from standard detection.

Pig Butchering — Intervention Speed Factor Victim-initiated: different formula

% of wires intercepted before clearing 40%

Wire interception requires real-time detection before Fedwire settlement (typically 2–4 hour window). 40% is achievable with the agentic pipeline; 60%+ requires proactive customer outreach protocols beyond AI detection alone.

Gross Fraud Exposure Rate (% of txn volume) Benchmark: 0.15%

Gross attempted fraud as % of total transaction volume — higher than net loss rate

Annual Transaction Volume ($M) Total payment volume

Total annual transaction volume processed (not fee revenue)

Loss Preventability Factor (%) Default: 75%

Of incremental fraud detected, what % results in actual loss prevention? Accounts for partial recoveries, detection latency, and false-positive fraud cases. Industry range: 60–85%.

Labor Realization Factor (%) Default: 75%

Of freed analyst hours, what % translates to measurable savings? Covers redeployment lag, partial reallocation, and roles not directly reducible. Conservative orgs: 50–65%; progressive: 75–85%.

Current Baseline Detection Rate (%) Critical input

Your fraud team's current detection rate before CAIBots. Detection lift inputs above represent incremental improvement over this baseline. Typical range: 55–75%. Higher baseline → smaller incremental improvement → lower L2.

Implied annual fraud losses

—

Calculator L2 loss avoided

—

L2 as % of implied losses

—

Sanity check: L2 loss avoided should be 10–60% of implied gross fraud exposure. The baseline detection rate above moderates this — at 65% baseline, detection lifts apply to the remaining 35% residual fraud pool. Values above 80% suggest inputs are too optimistic. Values below 5% suggest baseline is set too high or detection lifts are conservative.

Chargeback & Return-Fee Savings — CNP & ACH NEW LEVER

Distinct from fraud loss — chargebacks add processing + dispute cost on top of the transaction loss

—▼

Chargeback cost is separate from fraud loss. A CNP fraud generates both the transaction loss AND a chargeback processing fee. Both are real costs; there is no double-count between L2 and L3.

Monthly CNP Transaction Volume Total card-not-present

Total CNP transactions/mo across all card products

CNP Chargeback Rate — Before (%) Benchmark: 1.2%

CNP Chargeback Rate — After (%) Benchmark: 0.5%

All-in Cost per Chargeback ($) Benchmark: $75

Processing fee + dispute handling + card brand penalties

NACHA Return Fee / ACH Fraud Item ($) Benchmark: $25

ODFI return processing cost per fraudulent ACH item

Avg ACH Fraud Items per Case Benchmark: 8

Items per phantom payroll or ACH fraud incident

False Positive Recovery

Revenue recovered from unblocking good transactions — often the largest single lever for card issuers

—▼

False positive recovery counts margin on recovered transactions only. Do not also count customer lifetime value from the same population — that is a different metric and would double-count.

Monthly Fraud Alerts (all types) Benchmark: 2,000

False Positive Rate — Before (%) Benchmark: 12%

False Positive Rate — After (%) Benchmark: 5%

Avg Blocked Transaction Value ($) Benchmark: $185

Contribution Margin (%) Benchmark: 18%

Gross profit on recovered transaction revenue

FP Resolution Time (hrs) Benchmark: 0.5

Analyst time per false positive cleared

Regulatory Exposure — Probability-Weighted EV

Expected value of fine avoidance — supporting context for CFO, not hard ROI

—▼

Regulatory EV carries inherent uncertainty. Present these as probability-weighted ranges, not guarantees. Apply your own judgment based on your examination history and regulatory standing.

P(SAR Filing Failure) — Before (%) Benchmark: 5%

P(SAR Filing Failure) — After (%) Benchmark: 1.5%

Avg SAR Failure Penalty ($) Benchmark: $2.5M

P(Reg E Late Credit) — Before (%) Benchmark: 8%

P(Reg E Late Credit) — After (%) Benchmark: 1%

Avg Reg E Violation Penalty ($) Benchmark: $500K

P(NACHA ODFI Violation) — Before (%) Benchmark: 6%

P(NACHA ODFI Violation) — After (%)Benchmark: 1%

Avg NACHA Penalty ($) Benchmark: $250K

P(Elder Mandatory Report Miss) — Before (%)Benchmark: 10%

P(Elder Report Miss) — After (%)Benchmark: 2%

Avg State Elder Penalty ($) Benchmark: $150K

Solution Cost — Full Stack

Platform fee + AI inference + integration. The denominator most ROI tools hide.

—▼

Legal Entities / Business Lines

Jurisdiction

Data Residency

Contract Term

SLA Tier

Integration & Setup (one-time)

Regulatory Standing

Analysis Results Live

Annual Benefits

Analyst time savings — 10 fraud workflowsHigh confidence

—

Fraud loss avoidedVerify inputs

—

Chargeback & return-fee savings — CNP + ACHNEW

—

False positive revenue recoveryMed confidence

—

Regulatory exposure EVLow-med confidence

—

Total Annual Benefits

—

Annual Solution Cost

Platform fee (asset tier × entity multiplier)

—

Jurisdiction + data residency

—

Regulatory standing adjustment

—

SLA tier uplift

—

Term discount applied

—

AI inference cost (10 fraud workflows modeled; the demo’s scenario 11 is a governance control at ~$0)

—

Integration (amortized)

—

Total Annual Solution Cost

—

Cost per case — Before—

Cost per case — After—

Annual analyst hours freed—

Labor Saving by Fraud Type

Benefit Composition

Sensitivity Analysis

What carries this number. Two inputs account for most of it, and neither is a product capability. Baseline detection is how much fraud the institution currently misses — the lower it is, the more there is to recover, so this model is partly paid by an existing weakness. Prevention factor is how much detected fraud is actually stopped in time. Both are institution facts, not vendor claims, and both should be replaced with the institution's own figures before this goes near a board.

Two stress tests worth running here. Tick hard benefits only above. At the core-ten scope the model returns 70% with a 26-month payback on realised benefits alone — labour, fees, retained revenue and penalties, with the counterfactual layer removed entirely. Set baseline detection to 95%, meaning the institution is already very good, and the full model returns roughly 260%. Neither reading is the headline, and both are worth showing before someone else finds them.

Three things this model does not claim. AI inference cost is roughly 0.03% of total cost, so that column is disclosure rather than a lever — changing it cannot move the outcome. The analyst-capacity cap is real and binding in principle, but at benchmark volumes the team is about a quarter utilised, so it does not bind here. And L2, fraud loss avoided, is a counterfactual: it is loss that would have occurred, not cash that appears. It is the largest layer and the least verifiable, which is why the hard-benefit layers are shown separately above.

Conservative

—

−20% volume · 50% benchmark performance · +15% cost

Base Case ← Your Numbers

—

As entered · Demo Brief benchmarks

Optimistic

—

+20% volume · Full benchmark performance · base cost

⚠ Important Limitations — Read Before Presenting

This calculator models efficiency gains, fraud loss reduction, and partial revenue/risk uplift from the 5-agent CAIBots Fraud Detection pipeline. The following components are not calculated and must be assessed separately:

Model degradation cost of the current system over 24–36 months without upgrade (cost of inaction)
Adversarial adaptation value — speed of containing novel attack waves before losses compound
Customer retention impact from false positive reduction (CLV uplift — avoid double-counting with Section 04)
SR 11-7 independent model validation: $50K–$200K — not included in solution cost
Time-to-benefit: shadow mode and pilot validation mean full benefits may not be realized until Month 6–9
When ROI exceeds 300%, the driver is typically L2 (fraud loss avoided) — validate case volumes and detection lifts with your fraud team before presenting to the CFO

Illustrative. Actual results vary by institution, case mix, regulatory environment, and operating model. Benchmark figures derived from FinCEN examination statistics, NACHA annual reports, and industry publications. Not a performance guarantee. Not legal or financial advice.

Key Default Assumptions — Fraud Detection

Assumption	Default Value	Source / Basis
Baseline fraud detection rate	65%	Industry median (Aite-Novarica 2024); range 55–80% by fraud type
Detection lift (AI incremental)	Applies to residual undetected pool only	Conservative: AI can only detect within the currently-missed 35% — not re-detect already-caught fraud
Prevention / recovery factor	75%	Not all detected fraud is prevented in time; adjust per your ops model
Labor realization factor	75%	Slider in Section 01; industry standard workforce automation discount
Year 1 benefit ramp	70% (hard-coded)	Reflects pilot + shadow-mode period; add ramp slider for client customization
FP revenue recovery margin	User-defined (default 2%)	Net margin on unblocked good transactions; institution-specific
Regulatory EV	Probability × penalty (user-defined)	Expected value; verify penalty ranges with your compliance team
Revenue growth assumption	5% per year	Applied to 5-year IRR projection; conservative; adjust to your forecast

See This Applied to Your Institution →

30-minute architecture session · We map this pipeline to your fraud platform, data infrastructure, and analyst workflow

Legal Disclaimer & Material Assumptions

For informational and illustrative purposes only. This calculator generates forward-looking financial estimates based solely on the inputs you provide. It does not constitute financial advice, investment advice, legal advice, or a binding commercial commitment. CAIBots makes no representation or warranty, express or implied, as to the accuracy, completeness, or fitness for any particular purpose of outputs generated by this tool.

Actual results will vary. Projected savings, ROI, payback periods, IRR, and NPV are estimates derived from user-supplied inputs and publicly available industry benchmarks. They are not guarantees of future performance. Realized benefits depend on actual transaction volumes, staffing levels, integration complexity, regulatory environment, model validation timelines, and organizational factors not fully captured by any calculator.

Financial methodology. ROI = (Net Annual Benefit − Total Annual Cost) ÷ Total Annual Cost. Payback via cumulative monthly cash-flow simulation; Year 1 benefit ramp default 70% (adjustable). IRR uses Newton-Raphson iteration on a 5-year cash-flow series: Year 0 = one-time implementation cost only; recurring platform and API fees deducted from each future year. IRR figures for SaaS models are directional — a small one-time Y0 capex relative to large recurring savings produces high percentages. Compare to your internal hurdle rate; do not interpret absolute value. 3-Year NPV discounted at 8% WACC (adjustable). Nominal USD; no inflation adjustment.

Sensitivity scenarios. Conservative: −20% volume · 70% of benchmark savings · +15% cost · 60% Year 1 ramp. Optimistic: +20% volume · 100% benchmark savings · base cost · 80% Year 1 ramp. These parameters are identical across all CAIBots ROI calculators to enable consistent cross-product comparison.

Benchmark sources. Default inputs derived from: FFIEC examination statistics, FinCEN SAR/CTR annual reports, NACHA ACH network data, Celent/Aite-Novarica/Oliver Wyman industry surveys, and aggregated anonymized data from CAIBots institutional evaluations. CAIBots strongly recommends replacing defaults with your institution’s own volume, cost, and rate data before presenting results to executive leadership, boards, or procurement committees.

Compliance Scope SR 11-7 GDPR SOC 2 Privacy Terms

Why this is priced in tokens. A dollar figure per case asserts a rate. Tokens state the workload and let you apply your own. A case is a multi-agent run — each agent receives a system prompt, the case record and its retrieved context and returns findings, then a synthesis step writes the narrative. Summed across agents that is roughly 33k tokens for a simple disposition, 88k for a standard investigation and 220k where a long written narrative is produced. No prompt-caching discount is assumed, so these are the pessimistic end.

And why it is not in the per-case table. It used to sit beside the manual cost per case, a 350× gap in adjacent cells. That reads as a rigged comparison even when the arithmetic is right, and a reader who doubts it starts doubting the manual figure too — which is the number that actually carries the case. Inference cost belongs in the cost stack next to platform fee and implementation, where it is under half a percent of the total. It is disclosed because it should be, not because it drives anything.

Your Numbers.Your Real ROI.

Your Numbers.
Your Real ROI.