LIGHTS OUT FINANCE
Lights Out Finance · In this paper: Risk & Compliance Operations

The alert factory

Financial-crime operations manufacture documentation: 95% false positives, triaged in arrival order, judgment buried under assembly. Agents that investigate the funnel — so humans can judge the neck — invert the economics and the risk.

AB
Adil Bahir
Founder & Editor, Lights Out Finance · Two decades in finance transformation, quantitative finance, and enterprise AI
Interactive white paper · July 2026 · lightsoutfinance.net · 9-min read · Print / PDF
In the thesisLayer 3, pointed at the alert queue.
In brief
Detection improved for two decades; disposition never did. The result is a factory that manufactures documentation, triaged by arrival date rather than risk.
Alert investigation is a definitional machine match: evidence assembly at volume, policy-enveloped disposition of the clear majority, escalation with the work attached.
The regulatory objection runs backwards. What supervisors sanction is the human status quo — backlogs, inconsistency, thin narratives. Census-grade disposition is what the rules were asking for.

Walk into any financial-crime operations floor and you will find the purest expression of the broken back-office bargain anywhere in the enterprise. Screening systems, tuned defensively because the cost of missing is existential, pour out alerts of which the overwhelming majority — industry folklore says north of ninety-five percent — are false positives. Against that flood: rows of analysts, working the queue in arrival order, spending most of each investigation not deciding anything but assembling — pulling the customer file, the transaction history, the prior dispositions, the adverse-media hits — before the thirty seconds of actual judgment. It is an alert factory. It manufactures documentation.

The perversity is structural, not managerial. Detection improved for two decades; disposition never did. Every tuning improvement that raised recall raised volume; every volume increase was answered with headcount; and the headcount, drowning, triages by age of alert rather than by risk — meaning the genuinely dangerous case waits behind two hundred harmless ones because it arrived on a Tuesday.

Exhibit 1
The funnel and who works it
Alerts generated by screening · ~100% Genuinely ambiguous after investigation · a few % True escalations Today: humans triage the whole funnel, in arrival order Target: agents investigate the funnel; humans judge the neck
The funnel narrows from all alerts to a sliver of true escalations. Today humans work the whole funnel in arrival order; the target state has agents investigate the funnel and humans judge the neck.

Investigation is the sweet spot for machine work

Look at what alert disposition actually is: gather evidence from a dozen systems, resolve the entity, reconstruct the transaction context, compare against typologies and prior cases, write a defensible narrative, recommend a disposition. Repetitive in shape, unique in content, evidence-hungry, and rule-governed at the edges — it is almost a definitional match for the investigation pattern this publication has traced through reconciliations (The Desk After Dark) and exception queues (The System of Record Learns to Act). An agent works every alert the same hour it fires, assembles the full evidence file, disposes the clear majority within a written policy envelope, and escalates the ambiguous minority with the investigation attached. The human analyst — now properly a judge — opens a case and finds the work done, the question framed, the evidence cited. Risk-ranked, not date-ranked.

Exhibit 2 · Interactive
The alert-factory economics
Financial-crime operations by the numbers: what the queue costs today, and what changes when agents do the investigation and humans do the judgment.
Analyst hours per month, today
Investigator seats implied, today
Analyst hours after — judgment only
Attention multiplier on real cases
Post-agent state assumes escalated alerts arrive with the investigation done — entity resolution, transaction context, prior dispositions — cutting human minutes per case to a review of evidence rather than a hunt for it (modeled at 40% of manual minutes on the escalated share). Seats at 140 productive hours/month. Illustrative; the funnel shape is what matters.
The dangerous case waits behind two hundred harmless ones because it arrived on a Tuesday. That is not a staffing problem. It is a design problem.

The compliance objection, inverted

The reflexive objection — the regulator will never accept machine dispositions — has the evidence exactly backwards. What supervisors sanction, year after year, is the human status quo: backlogs aged past thresholds, inconsistent dispositions of identical fact patterns, narratives that cite evidence nobody actually pulled, quality-assurance samples failing at embarrassing rates. Against that baseline, the autonomous design offers what The Auditor Will See You Now called census assurance: every alert worked, identically, against a versioned policy, with a complete evidence file — and a human judgment concentrated precisely where the regulation actually demands it, on the suspicious minority. Consistency, timeliness, and documentation are not what autonomy threatens. They are what it is for. The tell-tale early sign to demand from any supervised-autonomy pilot: quality-assurance scores should rise on the agent-investigated population first — and if they do not, the policy envelope, not the model, is usually what needs work.

Burning down the backlog

Backlogs occupy a strange place in compliance reporting: measured to the alert, aged to the day, and treated as geological. The burn-down model exposes the small arithmetic underneath the large anxiety — a backlog is nothing but the integral of inflow minus capacity, and it clears on a computable date or never. What the autonomous layer changes is not the equation but the elasticity of the capacity term: human disposition capacity moves in hiring quarters; machine capacity moves in configuration changes. A function that can surge throughput for six weeks — without a recruitment cycle, without the quality collapse that always accompanied backlog “task forces” of borrowed analysts — converts the permanent condition into a project with an end date. Regulators, who have consent orders full of backlog commitments that slipped, notice the difference between a promise staffed with requisitions and one staffed with headroom.

Exhibit 3 · Interactive
The backlog burn-down
Every financial-crime function carries a backlog it reports carefully and reduces rarely. Three rates decide whether yours is a project or a permanent condition.
Net burn-down per day
Working days to clear
Capacity headroom
Verdict
Headroom is the share of daily capacity left after absorbing inflow — the resource that actually clears backlogs and absorbs volume spikes when a tuning change or a news event doubles the alert rate overnight. Sub-10% headroom means the next spike rebuilds the backlog; the autonomous layer’s elasticity (Paying for the Machines) is what makes headroom cheap for the first time.

The same inversion is arriving, on a slower clock, in the other half of the financial-crime estate: KYC refresh. Periodic review — the one-, three-, five-year cycle — is the alert factory’s quieter sibling: a calendar-driven ritual in which analysts re-collect documents that have not changed for the theoretical benefit of noticing the ones that have. Perpetual KYC flips the trigger from the calendar to the event: a directorship change in a registry, an ownership move, an adverse-media hit, a transaction pattern shift — each detected as it occurs, materiality-assessed within policy, and either auto-refreshed with evidence or escalated as a genuine review. The economics mirror alert disposition almost exactly — the overwhelming majority of periodic reviews conclude “no material change,” which is to say the current model spends its scarcest resource confirming the absence of events a feed could have confirmed for free. The functions that industrialize disposition first will find perpetual KYC is the same architecture pointed at a different queue: census monitoring, policy envelopes, evidence as exhaust, judgment on the residual.

Tuning without fear

The dirtiest secret of detection tuning is that it runs backwards: thresholds are set not where the risk is, but where the ops floor can survive. Every model-risk committee has sat through the euphemisms — “operational feasibility,” “alert budget” — that mean we detune the scenario because we cannot staff its output. Break the capacity constraint and tuning is finally free to serve detection: scenarios set to the risk appetite, thresholds where the typology lives, new scenarios piloted without a headcount business case, and the false-positive tax paid in cheap agent-minutes instead of scarce analyst-years. This is the deepest sense in which the alert factory’s successor is not an efficiency story. Detection quality was always hostage to disposition capacity; machine disposition is the ransom paid — and the model-risk committee’s minutes get shorter and more honest in the same quarter.

The narrative is the product

One artifact decides how a financial-crime function is judged from outside: the suspicious-activity filing and the narrative inside it. Regulators read them; law enforcement acts on them; quality-assurance samples them; and every weakness in the funnel upstream — the rushed investigation, the evidence nobody pulled, the template prose — lands in that document with a timestamp. This is where the autonomous design earns its keep beyond economics: a filing drafted from a complete, machine-assembled evidence file, in a controlled narrative structure, reviewed and signed by a human who had time to actually read it, is simply a better instrument of the regime’s purpose than the fortieth narrative an exhausted analyst wrote that Friday. The measure of the alert factory was volume processed. The measure of its successor is the quality of what gets escalated — and quality, unlike volume, is what the entire apparatus exists to produce.

Sanctions screening deserves one clarifying sentence, because it is where the “machines deciding” anxiety runs hottest: nothing in this design lets an agent wave through a potential match. The envelope for sanctions is asymmetric by construction — agents assemble the evidence, resolve the obvious mismatches in name-only hits against incompatible dates of birth or geographies where policy explicitly permits, and route everything else, investigation attached, to a human whose decision the regime actually requires. Autonomy’s contribution to the highest-stakes queue is not the verdict. It is that the verdict finally gets the file it deserved.

Where the freed capacity goes

And here financial crime differs from every other back office in this series: the freed capacity has an obvious, high-value destination. Investigators released from triage do not need redeployment workshops — they move up the funnel to the work the queue never left time for: typology development, network analysis, the proactive investigation of what the alerts mean in aggregate. The alert factory’s tragedy was never that it employed too many people. It is that it employed them at the wrong end of the funnel. The Last Org Chart’s pyramid-to-diamond argument, with the sharpest risk payoff in the enterprise.

The maturity question — where your disposition process actually sits between manual triage and governed autonomy — takes two minutes below, and your anonymous answer builds the cross-industry benchmark.

What leaders should do
Re-baseline the function in hours, not headcount.

Run the alert-factory model with your volumes and minutes; the attention multiplier on true cases is the risk argument, not the cost one.

Put agents on investigation, keep judgment human by design.

Evidence assembly at census scale, disposition within a written envelope, sanctions asymmetrically escalated — the regulator’s objection runs the other way.

Free the tuning.

Once disposition capacity is elastic, set thresholds to the risk appetite rather than the ops floor’s survival — and document the change for the model-risk committee.

Where does your operation sit?

The Lights Out Maturity Index: six questions, two minutes, no scales to interpret. Your anonymous result joins the inaugural Lights Out Finance Survey — the benchmark this publication reports on.

Take the Autonomy Readiness CheckTake the Maturity Index Browse all papers
Notes & references
Interactive models in this paper are the author’s analysis. Default values are illustrative; every input is exposed so you can calibrate with your own figures.
About the author
AB
Adil Bahir

Founder & Editor of Lights Out Finance. Big 4 partner in CFO Advisory & Finance Transformation with two decades across the Americas, EMEA, and APAC; DEng in AI (George Washington), MBA in Finance (Cornell), Master in Financial Engineering (Queen’s Smith); US CPA, CGMA, FRM, CQF, CTP, CDAA. Full profile →

Related papers