Transaction monitoring · UK banks and payment firms

Every alert
shows its working.

Adept scores financial crime risk with rules your model risk team can read, and puts an AI analyst on top that cannot state a figure it has not fetched. Built for the moment when someone asks why the system flagged this account.

Why UK firms are re-tendering now

Receiving banks now carry half the cost of APP fraud. Liability is shared with the firm holding the money mule's account. PSR

Failure to prevent fraud carries an unlimited fine. The defence is reasonable procedures, which includes ongoing monitoring and review. ECCTA s199

Every figure on this page carries its source. That is the product's discipline, so it is the website's too. Our method

ADEPT · ALERT QUEUE 27 open
EntityScenarioValuePtsSeverity
Bello Freight Ltd
CUS-0021
GEORISK Elevated-risk corridor£248,41120.0Critical
Bran Marchetti
CUS-0007
MULERING Shared identifier ring£61,20426.0Critical
Leyla Demir
CUS-0008
PASSTHRU Rapid pass-through£54,88018.4High
Priya Raman
CUS-0003
STRUCT Cash structuring£78,93022.0Critical
Owen Mensah
CUS-0011
CRYPTO Fan-in then off-ramp£47,32018.0High
Hale Trading Ltd
CUS-0024
VELSPIKE Velocity against baseline£112,00012.0Medium

Synthetic data. Every row opens to the transactions that produced it, and every point of the score to the scenario that raised it.

The UK position

Detection moved from a compliance cost to a balance sheet item.

Two changes in UK law did that. Neither is about having a monitoring system. Both are about being able to show that the one you have works.

Payment Systems Regulator

50 / 50

You pay for the mule accounts you bank

Since 7 October 2024, reimbursing an APP fraud victim on Faster Payments is mandatory, capped at £85,000 per claim, due within five business days. The cost splits evenly between the sending firm and the receiving firm. The PSR is explicit that the receiving side is the firm banking the money mule. PSR PS24/7

Between October 2024 and June 2025, receiving firms reimbursed over £74m to sending firms. Only 67% paid inside the five day limit. PSR, Oct 2025

Mule ring detection is no longer a hygiene control. It is the line item.

Economic Crime and Corporate Transparency Act

Unlimited

Failure to prevent fraud, in force since 1 September 2025

A large organisation is criminally liable where an associated person commits a specified fraud offence intended to benefit it, unless the organisation proves it had reasonable fraud prevention procedures in place. Conviction carries an unlimited fine. ECCTA s199

The Home Office guidance names six principles for that defence. One is ongoing monitoring and review. It also warns that being regulated does not by itself make your existing compliance processes reasonable procedures. Home Office guidance

You will be asked to evidence that your monitoring is reviewed and tuned. In writing.

Note on scope: money laundering is not a base offence under the failure to prevent fraud regime. We cite it because it drives fraud prevention procedure, not because it extends to AML obligations, which sit under the Money Laundering Regulations 2017 and JMLSG guidance.

The architecture

Two layers that never blur into each other.

Most vendors are bolting a language model onto a black box scoring engine and calling the result explainable. We separated the two, and the separation is enforced in the code, in the interface, and in what each layer is permitted to do.

Deterministic

The engine decides

Eleven configurable typology scenarios run over the ledger. Each contributes named points to an additive score, multiplied by the customer's KYC risk rating. The same data always produces the same score, and any point can be traced to the transactions that caused it.

  • Reproducible. No sampling, no temperature, no drift
  • Every point attributable to one named scenario
  • Thresholds visible and versioned, not buried in weights
  • Replays over historic data for back-testing
Assistive

The AI analyst explains

A tool-using agent reads the case through eight logged functions. It cannot see the ledger any other way, cannot assign or revise a score, and cannot state a number it has not fetched. Every answer shows the lookups behind it.

  • Structurally barred from scoring or deciding disclosure
  • Grounded: no figure without a retrieved source
  • Fixed answer schema, so output is scannable not chatty
  • Full tool trace attached to every response

Why this matters commercially: when validation asks how the score was produced, you point at the rule. When an investigator asks what it means, they ask the analyst. Neither question contaminates the other, which is what makes the model reviewable.

Inside the console

Four things you will not find in a standard queue.

01 · Deterministic layer

A score you can take apart

Click any segment and you land on the transactions that produced those points. The arithmetic is shown, not implied: rule points, multiplied by the KYC risk rating, equals the score on the file.

This is the panel model risk functions ask for and rarely get. It turns the validation conversation from "explain the model" into "read the line".

ENTITY 360 · CUS-0007 · SCORE 79
Score decomposition 68.4 pts × 1.15 KYC = 79
MULERINGShared identifier mule ring+26.0
PASSTHRURapid pass-through of funds+18.4
WATCHCounterparty on internal watchlist+13.7
GEORISKElevated-risk jurisdiction exposure+10.3
4 scenarios fired · 41 evidence transactionsBand: HIGH
AI ANALYSTgrounded

Three accounts share one device and forward funds to a single collector within hours of receipt.

High
Evidence
01CUS-0007, CUS-0008 and CUS-0009 log in from the same device and IP DEV-SHARED-4417
02Seven credits forwarded within 20 hours retaining under 12% of value TX-001284TX-001291
03Onward payments route to a watchlisted counterparty in the UAE CUS-0007
Reading

Consistent with a controlled mule network where two accounts feed a collector. Retention rate and timing are the strongest indicators.

If it is not suspicious Family members sharing a household device, pooling funds for a shared expense. Would be supported by a consistent counterparty and no watchlist exposure.
3 data lookups get_entity → get_network → get_transactions

02 · Assistive layer

Answers with a fixed shape and a visible trace

The analyst returns a structure, not a paragraph: the finding, the evidence with references back to real transactions, the reading, the innocent explanation, and what to do next. Investigators learn the shape once and then read it in seconds.

Expand the trace and you see every lookup the answer was built from. If a figure is not in the trace, the analyst was not permitted to say it.

03 · The unusual one

An assistant that argues against the alert

Every alert and every entity carries a button that asks the analyst to make the strongest case that this is a false positive. It uses the same data and the same grounding rules, pointed the other way.

Assistants built only to confirm the alert launder confirmation bias into something that reads like analysis. Given that a closed alert is a decision a regulator can review, an investigator needs the counter-argument on the file, not just the case for suspicion.

ALT-GEOVEL-0017 · CHALLENGE

The geo-velocity firing is most likely a card-not-present transaction misread as travel.

Medium
Counter-evidence
01Both presentments carry merchant category codes typical of online retail TX-001702
02No card-present indicator is held in the ledger for either leg CUS-0017
03Customer has prior legitimate spend in both regions CUS-0017
What would settle it Pull the card-present flag and the 3DS outcome for both legs. If either is card-not-present, the scenario should be suppressed for this class of transaction rather than dispositioned case by case.
TUNING · R01.threshold SWEEP
ThresholdAlertsEntitiesPrecisionRecallF1
£5,00026988.9%88.9%0.89
£8,00026988.9%88.9%0.89
£10,000271090.0%100.0%0.95
£15,00026988.9%88.9%0.89
Figures from the shipped synthetic benchmark, not a client deployment.

04 · Deterministic layer

Threshold changes come with evidence attached

Move one parameter and the console reports what it does to alert volume, precision and recall across the whole book. That output is the paper a tuning committee needs before it approves a change, produced in the tool rather than reconstructed afterwards in a spreadsheet.

In deployment the labels come from investigator dispositions and confirmed SAR outcomes, which is the only defensible basis for a production threshold.

Model your own numbers

What the receiving side is costing you

Set your figures. Nothing is sent anywhere, and nothing here is a claim about what Adept would deliver. The reduction slider is yours to set so you can decide what a change would need to be worth.

Celent puts the industry figure above 95%. We default lower so the output stays conservative. Celent

Analyst time on alerts closed with no concern—
Full-time equivalents absorbed—
Annual triage cost, no-concern alerts—
Annual receiving-side reimbursement exposure—
Value of the reduction you set—

Reimbursement exposure applies the PSR's 50% receiving-side share to your claim volume, capping each claim at £85,000. PSR PS24/7 It excludes investigation cost, funds repatriated, claims declined for gross negligence, and the £100 excess. Treat it as an order of magnitude for a conversation, not a forecast.

For validation and audit

Built to be reviewed, not just to work.

The hardest meeting in a monitoring procurement is not with the fraud team. It is with model risk. Adept is designed around that meeting.

Reproducibility

The engine is deterministic. Re-run any historic period and you get the same alerts and the same scores. Scenario definitions and thresholds are versioned configuration, not embedded weights.

Attribution

No composite score exists without a decomposition. Every point maps to one named scenario and the transaction references that triggered it. There is no residual term.

Contained AI risk

The assistive layer has no write path to scores, dispositions or thresholds. Removing it entirely changes no alert and no score. That containment is what lets you adopt AI here without re-validating the detection model.

Tuning evidence

Threshold changes are proposed with before-and-after performance across the book, generated in the tool. That output is the artefact your committee minutes can reference and the ongoing monitoring and review your ECCTA defence needs.

What we do not claim

Performance figures shown anywhere on this site come from a labelled synthetic benchmark that ships with the product. We have no client deployments to cite yet, and we will not borrow someone else's numbers. Run it against your own data during evaluation and use those figures instead.

Questions we get early

Straight answers

Is this a rules engine or a machine learning model?

Rules today, deliberately. The detection layer is eleven deterministic typology scenarios with an additive, decomposable score. The roadmap adds an unsupervised anomaly layer over peer-group features and a supervised model trained on investigator dispositions, with the rule set retained underneath as an explainable floor. We would rather ship something a validation team can approve than something that demos better.

Where does the AI actually sit?

On top, reading only. It queries case data through eight logged functions and cannot see the ledger any other way. It has no write path to scores, thresholds or dispositions. Turn it off and every alert in your queue is identical.

What about data residency and customer data leaving our estate?

The detection engine runs in your environment and needs no external calls. The assistive layer calls a model API, so it is scoped to the case data required for the question and is the part to bring to your DPIA. UK deployments generally want that layer pinned to a UK or EU region, or run against a self-hosted model. Both are supported. This is the first thing to work through in evaluation, not the last.

How does it fit alongside our existing monitoring system?

Most firms start it in parallel rather than as a replacement. The engine consumes a transaction and customer feed and produces scored alerts with evidence, so it can run alongside an incumbent for a comparison period on the same data. That gives you a like-for-like precision and recall comparison rather than a vendor's benchmark.

Can the AI write our SARs?

It drafts, from evidence it has fetched, alongside a deterministic template so you can compare the two. It cannot decide a disclosure and it is not submittable output. An investigator verifies every figure, adds the relationship narrative, and the MLRO takes the decision. Any vendor telling you otherwise is selling you a problem.

What stage is the product at?

Early. The detection engine, console and assistive layer are built and working against a synthetic benchmark. We are looking for a small number of UK design partners to run it against real data under an evaluation agreement. If you need a system of record with a decade of deployments behind it, we are not that yet, and we will tell you so on the first call.

Design partner programme

See it run against your typologies.

A 45 minute technical walkthrough with your fraud, AML and model risk people in the room. We will run the engine live, take the console apart, and show you the parts that are not finished.

No sales sequence. One reply from the person who built it.

Prefer email? info@adeptfcd.co.uk