Platform · tools
Seven agentic specialists, a knowledge hub that compounds with every closed case, and a crypto rail that sits on top of your chain-analytics provider. This is the shared machinery beneath every Corrobo product: agents gather and corroborate, the deterministic engine still makes the call.
The agentic bench · seven specialists
Every agentic investigation follows a fixed playbook: plan, gather, match, write, verify. The agents work only at designated steps, on an audit trail that cannot be edited after the fact.
← See the eight products these tools powerAn onboarding hit: 'Md Abdul Rahman b. Yusuf' against a watchlist entry reading 'Mohammed Abdulrahman bin Yousef'. Transliteration drift, honorific, initials: resolved as the same name at 0.93, with every step of the resolution shown. A lesser screen clears him; a cruder one drowns the analyst in false hits.
REQUEST THE FULL-CASE WALKTHROUGHSix personal accounts, all receiving just under the reporting threshold, all forwarding to one beneficiary within days. No shared surname, no shared address, but shared registration metadata and one registry officer in common. The graph draws the ring from evidence, and the case is scored at the ring, not the account.
REQUEST THE FULL-CASE WALKTHROUGHA counterparty three weeks old, registered by an agent who filed two hundred companies that quarter, directors stacked across a dozen entities, no trade history. Every signal from public record, each with its extract attached: a shell profile in minutes, not an afternoon of registry tabs.
REQUEST THE FULL-CASE WALKTHROUGHA deep sweep before an exposure decision: adverse-media pass with the name matched properly (not a string search), plus licensed number intelligence. SIM tenure two months, a disposable-line pattern on the contact number. Each finding filed as cited evidence, none of it asserted beyond its source.
REQUEST THE FULL-CASE WALKTHROUGHA share-purchase agreement, scanned, in Arabic. The agent reads it, extracts the sale value, parties and date (each field carrying its page and region) and the case file cites 'p.4' the way a memo would. One extracted figure later fails corroboration, and that is exactly the point: it turns yellow instead of becoming a fact.
REQUEST THE FULL-CASE WALKTHROUGHTwo past cases with the same shape: structured inflows, one fresh beneficiary. One was cleared in 2025: documentation arrived late but complete. One was confirmed. Both surface beside today's alert with their outcomes and reasons, so the analyst inherits the institution's judgment, not just its data.
REQUEST THE FULL-CASE WALKTHROUGHEvery closed case becomes searchable precedent.
Every analyst decision becomes a training label for your models.
Every confirmed link becomes an edge in your graph.
Patterns your rulebook misses surface as proposed rules, drafted in the same readable grammar, evidence attached. Your MLRO signs every one.
Eighteen months in, Corrobo doesn't just run at your institution. It remembers it. And it never leaves the building.
The draft narrative claims the customer's business sale is 'evidenced at AED 28M.' The verifier walks every sentence back to its cited record, and one claim has no record behind it. The sentence is flagged yellow and a document request goes out. Nothing unverifiable is ever shown as fact.
REQUEST THE FULL-CASE WALKTHROUGHThat is not a limitation. It is the feature your regulator will ask for.
The rules · nine packs
Structuring, rapid pass-through, dormancy-activation, mule fan-in and fan-out, high-risk exposure, large cash, new-account velocity, and two packs for virtual-asset firms. Every rule readable by an MLRO; every change backtested and signed before it activates. Alongside them, the baselines: each customer's own normal, learned across eight time windows. The bar is 'unusual for this customer,' not 'unusual in general.'
Money mules · two lenses
A mule account rarely fails one rule loudly. It fails a shape — money from senders with no reason to know each other, out faster than it came, through an account whose history says none of this is normal. Corrobo runs both lenses over the same customer and shows you each one’s working.
Fan-in and fan-out, dormancy reactivation, velocity, structuring — readable by an MLRO, versioned, backtested and signed before anything activates.
A calibrated score on the customer, over five feature groups — transaction patterns, demographics, network, temporal, device and channel — across 7/30/90/180-day windows. Credit-to-debit imbalance, a short window pulling away from the long one, dormant accounts waking, fresh accounts fanning in. A quiet account can score harder than a busy one.
A score is a candidate, not a verdict. The model abstains when the evidence is thin; every score decomposes into feature-level evidence an analyst reads and an MLRO signs. Nothing here blocks a payment.
Every alert carries its attributions — which feature groups moved the score, in which direction, against which window. The narrative is written from those attributions and nothing else, and the Verifier checks every sentence against the record it cites. No black box, and no figure an analyst cannot open.
False-positive reduction says what your analysts stopped reading. Yield — the share of reviewed alerts that turn out to be true positives — says what the reading was worth. A customer-level lens can raise alert volume and raise yield at the same time: the extra alerts are mules the rules never fired on.
Peer-reviewed production results published by a major Singapore bank report mule-alert yield of roughly 60% under rules alone against roughly 90% with a customer-level model — published evidence for the approach class, not a Corrobo measurement. Our figures come only from pilots on your own book.
RBIH’s MuleHunter.AI is rolling out across banks, and the Ministry of Home Affairs has directed financial institutions to integrate with the platform by December 2026. Corrobo is complementary and integration-ready: a MuleHunter flag arrives as one receipted evidence source, and Corrobo supplies what a flag alone does not carry — investigation, narrative, disposition and the audit trail a regulator asks for afterwards.
Corrobo is not affiliated with, endorsed by, or a partner of the Reserve Bank Innovation Hub or the MuleHunter programme.
Digital assets · crypto screening
The same agentic investigation, extended to virtual assets. Corrobo is the sovereign decision layer on top of your blockchain-analytics provider (Chainalysis, TRM, Elliptic), never a replacement for it. Agents pull wallet and transaction risk as evidence; the deterministic engine fuses it with the fiat side of the same customer; your analyst holds the pen. We make every chain-analytics licence in the Gulf more valuable.
A fiat leg settles; an on-chain transfer is instructed. The agent reads the connected provider's risk signal on the destination wallet (indirect mixer exposure, three hops) and files it as cited evidence with its receipt. The engine fuses it with the fiat side of the same customer; the analyst reads one case, not two systems.
REQUEST THE FULL-CASE WALKTHROUGHSovereignty holds: only public on-chain identifiers (wallet addresses and transaction hashes) ever leave the perimeter. Customer identity and fiat data never do. Built for the ADGM-licensed VASPs and the Gulf banks that service them.
These tools power eight products across fiat and crypto. See how they come together on one governed flow.