Your staff are already asking AI. Nobody is writing it down.

Why a safeguarded question never returns to a plain answer

The classification is governed by an invariant: a question may be routed to a more constrained outcome, never to a less constrained one. Where more than one assessment applies, the most constrained of them governs.

The reason is that the failure modes are not symmetrical. A question wrongly narrowed costs the practitioner an additional exchange. A question wrongly answered in full, where it should have been constrained, cannot be withdrawn - the response has been produced and read.

Rephrasing a question does not obtain a different route. The wording changes; the assessment runs on the substance, and the more constrained outcome continues to govern. This is deliberate. A control that can be talked around by a user under time pressure is not a control.

Within those constraints the response remains useful. A narrowed route still sets out what the rule requires and what the answer turns on, and states plainly what it has not answered.

Section 333A: why some questions cannot be answered as asked

Section 333A of the Proceeds of Crime Act 2002 makes it an offence, in the regulated sector, to disclose information likely to prejudice an investigation where a protected disclosure under Part 7 has been or may be made. The offence attaches to the disclosure itself.

This is why a compliance assistant cannot simply be helpful. A question that names a party, describes a suspicion and asks what to say to that party is not a request for research, and a drafted response to it is not a neutral output. A general-purpose assistant has no concept of the boundary and will compose the message.

Where the assessment indicates a question sits in this territory, the route is to the person inside the firm who holds the nominated officer role, inside the firm's own procedure, where that conversation is permitted. Companion identifies that the route is required. It does not perform the role, and it does not form a view on whether a disclosure should be made.

These are also the fastest responses in the product, because a routed outcome needs little or no generation. A practitioner who has disclosed a suspicion gets the quickest screen Companion produces. That is the right way round, and it is deliberate.

The exemptions and permitted disclosures at sections 333B to 333D are a matter for the firm and, where appropriate, its own advisers. Nothing here is a view on their application to any set of facts.

What a redacted restatement keeps, and what it drops

The classification runs on Go 2 Glory infrastructure, so it may see the question as the practitioner wrote it - nothing has left the boundary at that point. It produces two outputs rather than one: the classification, and a restatement of the question with the identifying particulars removed.

A practitioner asks: “My client Acme Ltd is a Romanian construction company and the director lives in Dubai - do I need enhanced due diligence?” The restatement reads: “Does a corporate client incorporated in an EU Member State, in construction, with a non-resident director in a third country, require enhanced due diligence?”

What is kept is what the answer turns on: the sector, the corporate form, the fact of non-resident directorship, the regulatory character of the jurisdictions involved. What is dropped is what identifies. This is not a trade-off between privacy and usefulness, because the general question is the one that can properly be answered anyway. The redaction and the routing want the same thing.

Where naming a jurisdiction alongside a sector, a corporate form and a role would narrow to a small enough population to identify, the restatement keeps the jurisdiction's regulatory character rather than its name.

The restatement is held to its own quality standard. A restatement that dropped something material would produce a well-formed answer to the wrong question, and nothing downstream of it could detect that.

The compliance log: what was asked, and what was relied on

Regulation 40 of the Money Laundering Regulations 2017 requires a firm to keep records of the customer due diligence it undertakes and of the supporting evidence. Regulation 19 requires its policies, controls and procedures to be applied and maintained. The common thread is evidence: a supervisor's interest is rarely in whether the firm reached the correct conclusion, but in whether it can show what it considered, when, and on what basis.

Every exchange is written to the practice's own log: the question, its classification, the authorities relied on, the corpus version in force, the outcome and the timestamp. The log is per-practice and isolated - no query path reaches another firm's entries - and it is append-only, because an editable log is not evidence.

The verbatim question and answer are encrypted with a key the practice holds. WhyAML holds ciphertext it cannot read, which is also the position if WhyAML is ever compelled to produce it. What WhyAML can report on is that a question was asked, how it was routed and what was relied on - not the words.

One field is editable: the practitioner's own note. “Discussed with supervisor 14/10, agreed approach” is frequently the entry a supervisor most wants to see. Edits to it are themselves versioned.

The log can be exported as a PDF per entry and as a full CSV. A record that cannot be handed to a supervisor, an insurer or a file is not evidence, so export is a product requirement rather than a convenience. On termination the practice takes its log.

If the log cannot be written, the question is refused. An answer given but not recorded is an answer the practice cannot evidence, and evidence with gaps that correlate with incidents is worse than no evidence at all.

The record supports the firm's position. It does not substitute for the firm having considered the question properly, and it is not evidence that the conclusion reached was correct.