AI Safety & Security
Feedzai Farol's 20% Claim Has a $9.05 Labor Benchmark
Feedzai says Farol cuts alert handling time 20%. An external wage benchmark values the reduction at $9.05 per baseline hour, before review costs.
Existing Feedzai customers should evaluate Farol as review assistance, not autonomous approval, after its September 24 launch claimed a 20% reduction in alert-handling time. Applying that reduction to the national financial-examiner wage gives $9.05 of wage-equivalent time per baseline handling hour, a screening benchmark rather than demonstrated bank savings.
A faster draft leaves a decision behind
Farol is embedded in RiskOps Studio and launches with support for risk strategy, investigations, product knowledge, and suspicious-activity-report drafting. Feedzai’s argument is that an agent working inside the existing data and analyst environment avoids the friction of a disconnected assistant. That is a credible reason for installed customers to investigate the product before building another integration. It is not evidence that every bank should replace its current fraud platform.
The economic translation uses an explicitly external input. The Bureau of Labor Statistics reports a May 2025 financial-examiner median wage of $45.27 an hour. Multiply that by the launch release’s 20% time reduction: $45.27 × 0.20 = $9.054, rounded to $9.05. The denominator is one hour of work in the baseline process, not one alert and not one hour after deployment.
Financial examiners are an adjacent occupational benchmark, not a measurement of Feedzai customers’ fraud-investigator pay. The calculation excludes benefits, implementation, subscription charges, and any extra verification effort. It also assumes the claimed reduction applies to active labor in the evaluated task. Replace both the wage and the reduction with local measurements before using the result in a purchasing decision. An external benchmark can discipline a conversation without becoming a forecast.
The wording itself deserves clarification. The launch release says handling time is reduced by 20%, while the product page describes reviewing or closing alerts 20% faster. A 20% increase in processing rate is not mathematically identical to a 20% decrease in elapsed time. This calculation follows the release’s explicit time-reduction wording. Ask which definition the vendor’s evaluation used and keep that definition in the acceptance criteria.
Other impressive figures concern other tasks. The launch release claims SAR drafting can be up to 12x faster, while the product page advertises faster rule-strategy analysis. Neither is a multiplier for fraud losses prevented or the team’s entire working day. A draft, an alert review, and a rule-analysis exercise have different starting points, outputs, and review obligations. Combining their headline improvements would invent a result.
The product’s own description preserves a useful human boundary: agents surface insights and recommendations, while analysts verify outcomes and act. Make that the pilot’s operating contract. A concise case summary can reduce assembly work while still requiring an investigator to inspect the underlying evidence. Measure the work after the summary appears, not merely the time until text becomes visible on screen.
Keep the release separate from the roadmap
A bank already using the platform should choose an existing review queue and compare matched work under the established process and Farol-assisted process. Include corrections, reopened cases, escalations, and the time spent checking generated material. These are recommended evaluation categories, not alleged product failures. The objective is a defensible decision reached with less burden, not a larger number of summaries produced.
Start with assistance that does not change production rules automatically. A recommendation can be evaluated against the same evidence and approval process used today. An action that changes detection behavior needs its own authorization, testing, and recovery path. Embedding both in the same interface does not make them the same risk. Keep the authority to approve a change separate from the convenience of asking for one.
The Farol product FAQ calls Bring Your Own Agent an upcoming capability still under active development. That matters for architecture planning. A bank can assess the currently described embedded assistant without assuming its own agent fleet can already connect through the promised path. Require availability in the actual account and a supported integration specification before retiring existing connections or purchasing around that dependency.
Price is another open input. Neither retrieved product material nor the release supplies a public per-alert tariff that can be compared directly with the $9.05 benchmark. Obtain the charging unit, implementation scope, usage limits, and support terms. Distinguish the incremental cost for an installed customer from the cost of adopting the broader platform. The same interface improvement can have very different economics for those two buyers.
Our Anthropic evaluation analysis distinguished investment in assurance from evidence about a deployed system. Farol requires that same separation at workflow scale. A transparent explanation is useful, but reviewers still need to trace important statements to case records and identify omissions. A plausible narrative should not become an independent authority merely because it appears inside trusted enterprise software.
The strongest case for adoption is integration. If investigators already spend time collecting evidence across the platform, an embedded assistant may remove friction that an external model cannot. The strongest counterpoint is verification cost: a faster first draft may require enough checking or correction to consume the gain. The retrieved launch material does not publish a full, independently audited distribution of those outcomes, so the bank’s trial must supply the missing evidence.
Today’s Ando lead asks whether shared context removes a handoff without erasing a boundary. Farol earns expansion on the same terms. Proceed when total review effort falls while decision quality, traceability, and human authority remain intact. Delay when the advertised speed depends on an unclear clock or an unavailable integration. A $9.05 wage-equivalent opportunity is worth testing; it is not permission to let the assistant approve its own work.