Coco
Use cases

KYB onboarding

How Coco governs an agent onboarding business customers against the organisation's own AML/CTF programme, and what we have not yet measured.

Decide each action an onboarding agent takes against your AML/CTF programme, and record why.

An onboarding agent gathers register extracts, ownership papers, and screening results for a business customer. Coco decides each action the agent takes with them, and records why.

Keep in mind: No KYB case has been run, so this page describes the mechanism, not a result.


The problem in KYB work

Before a reporting entity starts a designated service, s28 of the AML/CTF Act 2006 asks it to establish, for a customer who is not an individual, "the identity of any beneficial owners of the customer" (s28(2)(d)). Section 28(2)(e) asks whether the customer or any beneficial owner is a politically exposed person or designated for targeted financial sanctions. In the United States, 31 CFR 1010.230(d) names each individual owning 25 per cent or more of a legal entity customer, plus one individual who controls it.

An agent falls short in two ways:

  • Registration is not identity: A company registered last month under a name close to a real one passes a register check.
  • Collected is not verified: Section 28(3)(d) asks for verification "using reliable and independent data". An agent that confirms a business from its own website has collected, not verified.

When the regulator asks what the agent did, the answer today is a log the agent wrote about itself.

How Coco governs an onboarding agent

Coco decides each proposed action before it runs, in code the model doesn't control.

Keep in mind: The rules below are drafts. A model annotated them from the public texts and a build step compiled them into contracts. No person has reviewed either step, and the extraction pipeline hasn't yet run on the Act's s28 or on 31 CFR 1010.230. An organisation's own programme produces its own.

Control

Rules decide first, and an optional classifier only tightens a verdict. Each rule reads facts from the organisation's systems of record, never what the model says. The draft contract for approve_customer carries s28(2)(d) as one check, shown here trimmed:

- id: kyc_aml.au_amlctf_act_2006.approve_customer
  status: draft
  action: approve_customer
  checks:
  - id: cdd_beneficial_owners_of_non_individual
    check: '"customer" in state and "type" in state.customer and (not (state.customer.type != "individual") or ("cdd" in state and "beneficial_owners_established" in state.cdd and (state.cdd.beneficial_owners_established == true)))'
    on_fail: DENY
    document: au_amlctf_act_s28_s30

It denies approval of a customer who isn't an individual until the beneficial owners are established.

From 31 CFR 1010.230(d)(1), ownership_prong_25_percent denies open_account unless every individual owning 25 per cent or more appears among the identified beneficial owners.

On verify_customer, a classifier check from s28(3)(d) asks whether verification leaned on data the customer authored. It escalates from a score of 0.5 and denies from 0.9. When a fact a check reads is missing, the check fails at its own verdict.

Intervention

Every action ends in one of four verdicts: AUTO_EXECUTE (run), OBSERVE (run and flag), ESCALATE (pause for a person), or DENY (deny with the reason).

  • Pause for a person: Under AML/CTF Rules 2025 s5-5, when the customer or a beneficial owner is a foreign politically exposed person, approve_customer waits for a senior manager (senior_manager_approval_foreign_pep, ESCALATE).
  • Deny with the reason: Under s5-3, transfer_funds to a payee designated for targeted financial sanctions is denied (no_assets_to_designated_person).
  • Run and flag: An account opened without the owners' identifying information on record runs, and the receipt flags it (bo_record_identification, OBSERVE).

Evidence

Every decision leaves a hash-chained receipt, and every check names its source document and clause. An auditor replays any rule decision and gets the same verdict. Receipts are chained, not signed.

Policy from the organisation's own documents

Coco compiles the organisation's AML/CTF programme into contracts (behaviour contracts), one per action, each check citing its clause. Anything missing or conflicting goes to a person. A contract runs first in testing, where Coco records each verdict and enforces none, and goes live when the organisation promotes it.

How Coco fits KYB onboarding

The agent calls verify_customer and screen_customer for the entity and each beneficial owner, approve_customer or open_account, and transfer_funds or release_funds. Coco decides each call. It doesn't run the register, identity, or screening checks. It reads their results as facts supplied with each call.

The drafts load in the Coco SDK engine. Once an organisation reviews one and promotes it to testing or live, it runs as middleware on coco-framework, LangChain, or LangGraph, through the HTTP gate, or in the Claude Code hook. No KYB agent runs on any of them today.

What the guardrails stop

There's no KYB measurement: no KYB case set exists, and no draft contract has decided a case. The figures below measure the same engine on payments and injection benchmarks.

  • Payments benchmark: On a corpus of 73,356 payment transfer calls built from 4,371 human-written attacks, rules let 0 of 15,550 refused transfers run and held 0 of 57,806 permitted ones. A rule decision took 0.27 ms at the median.
  • Injection benchmark, AgentDojo v1.2.2: Of 588 planted attacks, rules alone prevented 467 (run 20260929-185238). Rules plus classifier prevented 582 and held 18 of 97 legitimate tasks for a person (run 20260929-184422).
  • Policy extraction, not enforcement: On two AML/CTF Rules 2025 extracts, the pipeline found 27 of 31 gold rules, and 27 of its 34 rules matched. A model annotated this gold, and no person has reviewed it.

To reproduce from an SDK checkout, run the commands below. Classifier figures move by a few calls between runs, and extraction moves with the model. The payment corpus downloads on first run from its public host and needs an account that accepted its terms. The AgentDojo cases come from bench/agentdojo/export.py.

coco eval vault --arms rules
coco eval agentdojo --arms rules,hybrid --question v2 --classifier <classifier-url>
coco eval policy --run --max-pages 3 --docs 2 --sets kyc_aml/au_amlctf_rules_2025 --out bench/policy/runs/kyb
coco eval policy --predictions bench/policy/runs/kyb --by-set

What Coco can't promise yet

  • No connector to a business register, a sanctions list, or a screening provider exists.
  • Coco doesn't trace ownership through layered companies and trusts. The 25 per cent rule compares the list of owners it's given.
  • The extraction pipeline hasn't run on the Act's s28 or on 31 CFR 1010.230.
  • The classifier hasn't been measured on KYB checks.
  • The corpus leaves out AML/CTF Rules 2025 Part 6 Division 1 and the Act's s5 definition of beneficial owner, and the US drafts don't reflect FinCEN's exceptive relief order of 13 February 2026. A KYB number also needs a case set with legitimate twins, reviewed contracts, and a replay.

From our team

"A control the agent satisfies by itself is not a control. Onboarding a business turns on facts the agent does not own, such as who stands behind the company. Coco takes those facts from the organisation's records, decides each step against its own programme, and keeps a receipt an auditor replays." Khoa Pham, co-founder and CTO, CocoAI

Working with us

If you run agents in business onboarding or AML/CTF work, contact us at trustcoco.ai.


On this page