Anthropic’s compliance agents have landed in banking. Your marketing copy is next on the audit trail.

Anthropic spent the first week of May 2026 making one thing clear: it wants Claude to be the operating layer for regulated financial work. On 4 May, the company announced a partnership with FIS — the back-office system of record for transactions, payments, deposits, and customer activity across roughly 12% of the global economy — to build a Financial Crimes AI Agent embedded directly inside FIS’s existing infrastructure. One day later, on 5 May, Anthropic released 10 ready-to-run agent templates for financial services, with a Know-Your-Customer (KYC) Screener as the headline product. BMO and Amalgamated Bank are the first development partners; broader availability is planned for the second half of 2026.

For traditional banks, the pitch is straightforward: cut the cost of compliance. For crypto exchanges, the implications are sharper — and they land squarely in the marketing function, not just compliance.

What the KYC Screener actually does.

The KYC Screener ingests an onboarding record, assembles the entity file, reviews source documents, applies the firm’s own KYC/AML rules, assigns a risk rating, and packages escalations for human review. Output is structured JSON — risk rating, document checks, rule-by-rule citations — designed to be ingested by downstream corporate systems. The other nine templates cover pitch building, meeting preparation, earnings review, model building, market research, valuation review, general ledger reconciliation, month-end closing, and statement auditing.

The KYC Screener is generic. Firms tune it to their own rule set. That matters because the rule set — the firm’s authorisation scope, its custody architecture, its AML thresholds, the jurisdictions it serves and excludes — is exactly what a CASP’s marketing pages currently claim, in human language, on every product page, every onboarding screen, every blog post and KOL contract. For the first decade of crypto-exchange marketing, those claims lived in a “we say what we say, compliance reviews when they have bandwidth” zone. That zone is closing.

The structural change: claim defensibility moves from human review to machine verification.

Here is the architectural fact every crypto-marketing leader should internalise: the same Claude reasoning layer that, by H2 2026, will be inferring risk ratings and SAR-narrative drafts from a bank’s core systems can also be pointed at the bank’s marketing copy and asked to reconcile claims with rule sets. Anthropic doesn’t ship that as a product. But the FIS Financial Crimes Agent’s architecture — client data stays inside the firm’s controlled infrastructure, Claude reasons one layer back, every decision is traceable and auditable — is the same pattern any sufficiently funded regulator, journalist, or competitor can replicate against a CASP’s public marketing surface in an afternoon.

In other words: the marketing claims your exchange makes are about to be subject to the same evidentiary pipeline as your AML investigations. Not by your choice. By the structural availability of the tooling.

The MiCA Article-60 problem.

For European CASPs, this isn’t theoretical. MiCA Article 60 places clear and specific requirements on the content of marketing communications — they must be fair, clear, not misleading, and consistent with the disclosures in the white paper or the authorised scope of the CASP. ESMA’s guidance through 2025 and the operational reality of 1 July 2026 enforcement mean that any inconsistency between a marketing claim and the firm’s actual authorised scope is enforcement-eligible from day one.

Three categories of claim are most exposed:

The first is authorisation-scope claims — “fully MiCA-compliant,” “EU-authorised,” “regulated under MiCA” — written months before the firm’s actual CASP authorisation status is settled. A Claude-driven reconciliation between marketing copy and the firm’s actual filings will surface these in minutes.

The second is custody-architecture claims — “your assets are segregated,” “qualified custodian,” “cold-storage by default” — where the marketing-page abstraction is more flattering than the operational reality. The KYC Screener pattern, pointed at the custody policy document and the marketing page, produces a citation-grade diff in JSON.

The third is AML-and-jurisdiction claims — “we don’t serve [restricted country],” “all users are KYC-verified to FATF travel-rule standard,” “we screen against [sanctions list]” — where the marketing claim is a near-verbatim restatement of what the AML policy should say, but the operational reality may have drifted. The Financial Crimes Agent pattern surfaces drift automatically.

None of this requires the regulator to act. It requires only that one journalist, one short-seller, one competitor’s compliance team, or one disgruntled ex-employee runs the prompt.

What marketing leaders at MiCA-jurisdiction CASPs should be doing this quarter.

Three actions, in priority order, before 1 July 2026:

Run a defensibility audit of every public marketing claim against the firm’s authorised scope. Not the marketing-team’s understanding of the scope. The actual filings, the actual policy documents, the actual rule sets the compliance team would hand to a KYC Screener if asked to tune it tomorrow. Every claim that can’t survive a citation-grade reconciliation gets rewritten or removed before the enforcement window opens.

Align marketing taxonomy with compliance taxonomy. If the compliance team calls it “transaction monitoring threshold of €1,000,” the marketing page should not call it “industry-leading AML standards.” If the authorisation covers “exchange of crypto-assets for fiat currency,” the marketing copy should not imply custody services unless the custody licence is also in hand. The vocabularies converge or the gap becomes enforcement-eligible.

Build the interpretation layer that Anthropic explicitly leaves to humans. Claude Opus 4.7 scored 64.37% on Vals AI’s Finance Agent benchmark — Anthropic itself describes this as a workflow where users “stay firmly in the loop, reviewing, iterating on, and approving Claude’s work before it goes to a client, gets filed, or is acted on.” That interpretation layer — the named-expert sign-off on what a JSON risk rating means for a specific marketing claim, on a specific product page, in a specific MiCA jurisdiction, on a specific date — is the work that does not get automated. It also happens to be the highest-margin work in the marketing-compliance interface.

Why NorthPoint exists for exactly this slot.

There is a structural gap between the agent-shipped reasoning layer (Anthropic + FIS) and the operator-grade interpretation layer (what to actually do about a flagged marketing claim under MiCA Article 60, against a specific authorisation scope, with a specific go-to-market timeline). Big Four corporate-compliance practices will sell a €300k+ engagement to fill that gap with generic frameworks. White-shoe law firms will sell a €100k+ memo. Neither has shipped a single exchange marketing page in production.

NorthPoint sits in the operator’s seat. Exchange-grade crypto marketing for the post-MiCA, post-agent-compliance era — where every marketing claim has a defensibility audit trail, every marketing taxonomy aligns with the compliance taxonomy, and every campaign has an operator-grade interpretation layer between the JSON risk rating and the published landing page. Run a free MiCA-marketing check at northpoint.fi/check/mica, or read the Finnish Sovereign Bitcoin Reserve Whitepaper Annex D for the broader operator-grade framework on the agent-economy regulatory landscape.

The compliance agents have landed in banking. The marketing audit trail is next. The exchanges that act this quarter will be the ones whose marketing copy survives the first wave of automated reconciliation. The exchanges that wait until 1 July will be the ones whose marketing copy becomes the enforcement case study.

— Jukka Blomberg, Helsinki, 25 May 2026