By Pierre Gérard, CEO and co-founder, Scorechain
When we founded Scorechain in Luxembourg in 2015, “blockchain analytics” was not yet a category. We spent the first few years explaining to banks and regulators why the transparency of a public ledger was an opportunity rather than a threat. A decade later, I am watching the same misunderstanding attach itself to artificial intelligence (AI), and it is costing the industry time it does not have.
Two stories dominate the conversation. The first says AI will soon replace compliance teams altogether. The second says AI is too unpredictable to let anywhere near regulated financial activity. I do not believe either one, and I say that as someone whose company has risk-assessed more than 2,800 virtual asset service providers (VASPs) since 2015, and has spent the past two years adding AI where it genuinely helps, as a separate layer rather than something baked into the compliance tools our clients depend on.
Start with the problem that every compliance officer I speak to raises within the first five minutes: noise. A sanctions screening system tuned the way a nervous bank tunes it can throw off false positives on the order of 95%. Transaction monitoring is not far behind. So a trained analyst, someone who understands typologies and can read a fund flow, spends the bulk of the working day clearing alerts that were never risks: dismissing name matches on a common surname, reading five adverse media hits that turn out to describe a different person entirely. Each of those adverse media checks takes an analyst 10 to 20 minutes. That is the actual texture of compliance work today, and it is why good people burn out of the profession.
This is where automation earns its place, and it is a narrower place than the hype suggests. I am not neutral about it. Scorechain AI exists to hand an analyst a single report: a wallet’s risk score, the entity types it has interacted with, and the named services and counterparties it has been exposed to. That is work that used to mean hours of manual tracing across a ledger. But notice what the report does and does not do. It does not decide anything. It compresses the evidence so that the compliance officer, the person who has to sign off on that decision and defend it to a regulator later, can read it in minutes and then make the call. That is the whole game. Good automation does not shrink the compliance function; it moves it off the treadmill of triage and back toward judgment.
The distinction matters, because the alternative is dangerous. In a regulated setting, a model cannot answer to a supervisor. The Sixth Anti-Money Laundering Directive (AMLD6) and the Markets in Crypto-Assets Regulation (MiCA) both require an institution to explain and stand behind its decisions. “The algorithm flagged it” is not a defence at an inspection, and “the algorithm cleared it” is worse. So the only responsible design is AI as a support layer sitting on top of trustworthy data, with a named compliance officer retaining the decision and the accountability that comes with it. Human oversight is not a training-wheel we remove once the model matures. It is the architecture.
And a model is only ever as good as what sits beneath it. This is the part outsiders miss. On its own, an AI reading a blockchain sees only anonymous strings of characters moving value to other anonymous strings. It cannot tell that the wallet three hops upstream is a sanctioned exchange, or that the counterparty receiving the funds is a mixer rather than a payroll provider. Supplying that missing context is the whole job of blockchain analytics: attaching identity and risk to raw on-chain activity, tracing indirect exposure across multiple hops rather than just checking the address in front of you, and scoring it against the more than a billion data points and over a million crypto entities we have labelled since 2015. Take a concrete case: a wallet looks clean at first glance, but tracing its flows shows that most of its balance arrived, two hops back, from an address tied to a ransomware operator. That is the finding a model would never reach on raw chain data alone, and the one a compliance officer has to act on. Feed a model that context and it can reason on solid ground. Feed it thin data and it produces confident nonsense, which in compliance is more dangerous than an honest gap, because it clears things it should not.
Here is what I find genuinely new. AI is no longer only a tool that compliance teams use; it is becoming a participant in the market they monitor. Autonomous agents that initiate payments under preset limits have moved from demo to deployment, pushed along by real infrastructure: Coinbase’s x402 standard for machine-to-machine payments, Visa’s Trusted Agent Protocol, the PayPal and OpenAI checkout integration. Software is starting to transact with other software, settling in crypto assets, at a volume no treasury team could match by hand.
It raises a question the industry has not answered cleanly yet: how do you apply Know Your Transaction principles to a counterparty that is a piece of software? The direction, at least, is clear. When agents transact on their own, controls cannot live only at onboarding. They move to the transaction layer itself: real-time monitoring, velocity limits, provenance, and the ability to intervene while money is still in flight. The transparency we spent years defending to sceptics turns out to be the one thing that makes autonomous on-chain activity auditable at all.
This is the future we decided to build for rather than wait on. We recently launched Scorechain MCP, which exposes our risk scoring and entity intelligence through the Model Context Protocol, the emerging standard that lets AI agents call external tools directly. The intelligence lives in our platform, and the AI stays outside it, calling in for answers rather than being embedded in the compliance tool itself. The premise is simple: an agent should never transact blind. Before it moves funds or approves a counterparty, it can ask Scorechain in the same breath whether that address is a sanctioned entity, a mixer, a known scam, or a clean private wallet, and receive a risk score in return. This is not a hypothetical throughput. We already run more than 1.5 million AML checks a day, and a screening call returns in roughly 235 milliseconds, quick enough to sit inside a live transaction rather than slow it down. We put it where those agents and workflows actually live, as an app inside ChatGPT and Claude, and as an integration on automation platforms such as n8n and Zapier. A compliance check that sits inside the flow, at the moment the decision is made, is worth far more than one bolted on after the money has already moved. The most basic check of all, whether an address appears on a sanctions list, should not sit behind a paywall for anyone. That is why we offer it as a free sanctions screening API that any developer, agent, or workflow can call. Screening for sanctions exposure is not where a compliance provider should be extracting value; it is the floor the whole market should be standing on.
There is a second-order shift here that token issuers and asset managers are only starting to reckon with. When value moves into stablecoins and tokenised assets at machine speed, the risk that matters is no longer only the individual transaction but the asset itself: who holds it, how concentrated that ownership is, and how much of the supply sits with sanctioned or otherwise high-risk entities. That is a different question from transaction monitoring, and it is the one our Digital Asset Intelligence is built to answer, giving an issuer or an asset manager an asset-level view of holders and exposure before they mint, list, or allocate.
Europe is readier for this than it is given credit for. MiCA and AMLD6 already assume continuous monitoring and clear accountability rather than a one-time check at the door, and a regime that assumes activity must be explainable is exactly what you want when software starts moving money. So yes, AI is coming for crypto compliance. It will remove a great deal of tedious work, and I welcome that. What it will not remove is the need for judgment, accountability, and verifiable data. It raises the bar on all three. The teams that treat AI as a faster analyst, grounded in reliable on-chain intelligence and kept firmly under human control, are the ones who will still be standing when the machines start transacting. That is closer than most people think.
