Process map › C5 Alert review › Sanctions alert review

Sanctions alert review

What AI can do at each step of sanctions alert review, what ScalePoint Research's tests found, and how that compares with what the 22 tools on sale claim.

Sanctions alert review: on this page

Our verdict
How the work is done today
How an AI agent would work an alert
What our AI research has found
Vendor claims against evidence
The sanctions alerts report

1Our verdict · what AI can do, and what you can buy

AI can do every step of the review. The tools on sale stop at scoring and suggesting, and ScalePoint Research found no independent test of a whole-alert agent.

Step	How AI does it	Can you buy it?
1 Screen and raise the alert	Code compares, rules set the threshold, machine learning cuts weak hits	Standard
2 Filter the obvious	Machine learning	Widely sold
3 Gather the facts	Code pulls the records, language model compares them	Copilots
4 Clear or escalate	Language model (22 of 22 right in S3)	Copilots; five vendors claim an agent
5 Record the reason	Language model	Copilots
6 Second review	Person decides, language model prepares the case	Some
7 Act on a true match	Code blocks and reports, person signs off	Standard
8 Tune and check	Code tracks results, machine learning helps tune thresholds	Standard
The whole alert	Agent, no independent test found	Two on sale, three more described

Yes Partly

Where people must stay. Alerts with conflicting records, and every possible true match. In S3, a cheaper model cleared some conflicting alerts wrongly, and was as confident when wrong as when right.

Model choice matters. A general frontier model got every call right. The cheaper model cost less to run, but one wrongly cleared alert is a sanctions breach, so the saving isn't worth it.

"How AI does it" comes from ScalePoint Research's analysis of the process and its model test, S3. "Can you buy it?" comes from a scan of 22 vendor tools on 9 November 2026. Where a step uses more than one technology, "Can you buy it?" covers them working together. Both will be reviewed as new evidence comes in.

Now shows how most firms do each step today. What AI could take on is in the verdict above.

Code: Fixed software that moves, looks up and matches data the same way every time.
Rules: If-then settings that people write and tune, such as match thresholds or a list of known false matches.
ML: Machine learning: models trained on past decisions that score or rank new cases.
LLM: Language models that read documents and records, make a judgement and write it up.
Person: An analyst or reviewer.
Emerging: In use at some firms, not yet the norm.

Screen and raise the alert
Screening software checks customers when they join and whenever the lists change, and checks payments as they pass through, often in real time. Code runs the comparison, and rules set how close a name must be, allowing for spelling, word order and transliteration, before an alert is raised with a match score. Some screening tools now use machine learning in the matching itself, to cut weak hits before they become alerts.
Now:CodeRulesML (emerging)
Stays with a person: Sets the matching thresholds and which lists apply.
Filter the obvious
Rules close repeat hits an analyst has already cleared, using a list of known false matches, such as a regular customer whose name resembles a listed person. Machine learning goes further: trained on past decisions, it scores every new alert and closes or deprioritises the weakest. Both run cheaply and instantly on every alert, and the call rests on structured data such as match scores and identifiers. A language model would add cost and delay here for little gain, so it is better kept for the alerts that are left, where reading and judgement matter.
Now:RulesML
Stays with a person: Approves what can be closed automatically.
Gather the facts
Pull the customer or payment details, such as date of birth, nationality, address and ID numbers, and set them against the list entry.
Now:CodePersonLLM (emerging)
Stays with a person: Researches ownership and control for complex companies.
Hard part: Payment messages often carry little more than a name. For a company, the call can turn on who owns or controls it, and US guidance treats 50% ownership as a minimum test, not the whole answer¹^,2.
Clear or escalate
Decide whether the hit is a different person or entity, or a possible match. A different date of birth or nationality usually clears it.
Now:PersonML (emerging)LLM (emerging)
Stays with a person: Takes every possible true match, and any alert with conflicting documents or doubtful identifiers.
Hard part: Common names, transliteration and conflicting records make many hits genuinely unclear. These were the alerts the cheaper model cleared wrongly in study S3.
Record the reason
Write down which list entry was checked and which detail decided the case, so the decision can be audited later.
Now:PersonLLM (emerging)
Stays with a person: Spot-checks the written reasons.
Second review
Escalated alerts go to a senior analyst or the money laundering reporting officer. Many firms also check a sample of cleared alerts.
Now:Person
Stays with a person: Decides on escalated alerts. A model can prepare the case.
Act on a true match
Block or reject the payment, freeze the funds and report to the regulator. In the US, reports of blocked property and rejected payments are due within 10 business days³^,4. In the UK, firms report to OFSI as soon as practicable⁵.
Now:CodePerson
Stays with a person: Decides to block or freeze, and signs off the report to the regulator.
Tune and check
Track alert volumes and wrong decisions, adjust thresholds and false-match lists, and validate any scoring model.
Now:CodePerson
Stays with a person: Approves changes to thresholds and models.
Hard part: A large round of new designations means rescreening every customer, and payments held for review wait while the queue clears.

3How an AI agent would work an alert

The agent's own actions, lettered A to F so they aren't confused with the eight steps in section 2. Together they cover steps 2 to 5.

AAlert arrives
BPulls the customer record
CPulls the full list entry
DFor a company, looks up owners and controllers
ECompares the detailsLoop
FWrites the reason and closes, or escalatesSign-off

Loop at EIf a detail is missing or conflicting, the agent asks for more information or routes the alert to a person.

Sign-off after FA person reviews every alert the agent escalates, which includes every possible true match and any alert with conflicting records.

What makes it an agent. A model in a single step answers the one question it is given. An agent decides its own next step, calls tools such as the customer database or a company registry, and works across several systems until the alert is done.

What changes about the risk. A mistake early on can carry through to the decision. The agent needs access to customer data. Every action has to be logged so it can be audited, and the agent validated like any other model a regulator would expect a firm to check.

4What our AI research has found so far

The decision step works. In study S3, Claude Opus 5.5 made the right clear-or-escalate call on all 22 alerts at an EMI, in both runs, at low and high effort: 44 of 44 answers.
The cheaper model cleared alerts it shouldn't have. Mistral Medium 3.5 got 40 of 44 right with reasoning on and 39 with it off. Its misses were alerts with conflicting documents or a doubtful date of birth, which the rules send to an analyst.
Confidence didn't help. Mistral was about as confident in its wrong answers as its right ones, so its confidence couldn't be used to send risky alerts to a person.
Messy records are the weak spot. Alerts with conflicting records should go to a person, whichever model is used. Read study S3 →

What S3 covers. The single clear-or-escalate decision, with the facts already gathered. A whole agent flow, as described in section 3, hasn't been tested. The reference agent in the report would do that.

What our tests add. ScalePoint Research tests how well general AI models do the work, and what each case costs. That sets a baseline: a tool that charges for alert review should do clearly better than a well-prompted general model running at a known cost per case.

Confidence: medium. Based on one ScalePoint Research test on 22 synthetic cases, public regulator data and vendors' own claims. No independent before-and-after results were found.

5Vendor claims against evidence

On 9 November 2026, ScalePoint Research read each vendor's product pages, documentation and press coverage, and recorded what each says its AI does when an alert is reviewed. The table compares those claims with the only independent evidence there is so far: ScalePoint Research's own model test, S3, in section 3. No vendor product has been tested.

AML and financial crime suites (9). FICO, Hawk, Napier AI, NICE Actimize, Oracle, SAS, SymphonyAI, Temenos, Tookitaki
Screening platforms (7). Bottomline, ComplyAdvantage, Eastnets, Finastra, LexisNexis Risk Solutions, Ripjar (two products)
List data providers (3). Dow Jones, LSEG, Moody's
Case management tools (2). Lucinity, Pega
AI-first alert review (1). Silent Eight

Step	What vendors claim	What the evidence shows
Filter the obvious	Machine learning that scores alerts and closes or deprioritises likely false positives, sold by LexisNexis Risk Solutions, LSEG, Moody's and Napier AI among others.⁶^,7^,8^,9	The most established AI in this process, and in wide use. ScalePoint Research hasn't tested it.
Gather the facts and clear or escalate	Copilots that summarise an alert and recommend a decision, from NICE Actimize, Oracle, SymphonyAI and Lucinity.¹⁰^,11^,12^,13 Five vendors say an agent makes the call, set out below.	In study S3, a general frontier model made the right call on all 22 test alerts. A cheaper model cleared alerts with conflicting records that should have gone to an analyst.
Record the reason	Silent Eight, ComplyAdvantage, Ripjar and Hawk say their agents record a reason or an audit trail for each decision.¹⁷^,18^,16^,15	S3 asked the models to name the list entry and the detail that decided each alert. A written reason is within reach of a general model.
The whole alert	Five vendors say an agent works whole alerts, set out below. Two say the product is on general sale, and none names a customer for its current agent.	ScalePoint Research found no independent test of an agent working whole alerts, and hasn't run one itself.

The gap. The evidence so far suggests a general AI model can already make the clear-or-escalate call on well-documented alerts, while most tools on sale stop at scoring or suggesting. The open questions are alerts with conflicting records, and whether an agent holds up on a firm's own data.

2 agent in production4 copilot5 ML scoring10 unclear1 no AI

Who claims an agent, and the evidence of live use

Five vendors say an AI agent decides sanctions alerts. None of these claims has been tested independently, and none names a customer for its current agent.

Vendor	What it claims	Evidence of live use
Silent Eight	Decides every alert: a clearance, an escalation or a policy exception¹⁷	On general sale. Its named bank customers, such as HSBC, date from earlier versions of the product.
ComplyAdvantage	Clears false positives it is confident about and leaves the rest for an analyst¹⁸	Generally available, by its own account. No customer is publicly tied to the screening agent.
Ripjar	Auto-closes low-risk alerts and escalates the rest with evidence and an audit trail¹⁶	Quotes a 77% cut in human effort, with no baseline, and describes the assistant as an emerging capability. No customer named.
Hawk	A payment screening agent that triages hits with reasoning for each decision¹⁵	Named on its product page. No launch date, customer or results.
SymphonyAI	Agents that work sanctions alerts and cut the workload¹⁴	Results from a proof of concept at an unnamed US bank.

Based on each vendor's public material, checked on 9 and 10 November 2026. Ripjar's wider figures, such as up to 91% fewer false positives, are for its screening platform, not the agent.

Vendor claims come from the vendors' own published material, summarised, with a numbered source for each. The group list and the AI ratings come from the same scan. "Unclear" means the vendor names AI but doesn't say what it does with an alert. All vendors in this cluster →

6The sanctions alerts report

The report is built around a reference agent. ScalePoint Research builds a working AI agent for sanctions alert review from models any firm can buy, tests it on a set of realistic alerts, and uses the results to judge build or buy, impact, implementation and the vendors on sale. Each report is written once five or more firms have registered interest.

No obligation to buy. Once five firms have registered, we will email you an offer.

Process mapC5 Alert review VendorsWho sells software for this cluster Study S3Includes a test on this process Transaction monitoring and SAR draftingThe other alert review work in financial crime teams

Sources

Vendor counts and AI ratings come from a ScalePoint Research scan of vendor sites, documentation and trade press, checked on 9 November 2026. A description of how the process typically runs, with ScalePoint Research's view on AI. Practice differs by firm, regulator and list. Status as of November 2026.

Get an email when new research is out

Sanctions alert review

2How the work is done today

Screen and raise the alert

Filter the obvious

Gather the facts

Clear or escalate

Record the reason

Second review

Act on a true match

Tune and check