What crosses the boundary, why doing this yourself does not last, how a question moves through it, and where its limits are.
Every question lands in your compliance log like this.
Five things on the screen, and what each one means for you.
Not masked. Not tokenised. Not at any stage. There is no code path that sends it - which is a different claim from “we redact before sending”, and a much easier one to hold true over time.
“My client Acme Ltd is a Romanian construction company and the director lives in Dubai - do I need enhanced due diligence?”
“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?”
The classification does not cross either. The generation layer is told what it may and may not produce. It is not told that the question was assessed as disclosing a suspicion - that is a conclusion about a named party’s conduct, and it stays on our side.
Every threshold, period and date comes from a maintained table with its provision attached. Where a value cannot be retrieved, the answer states the gap or escalates. A stated gap is recoverable; a generated number is not, because nothing about it tells you to check it.
Source material is held as immutable snapshots, and every answer records the version in force when it was given. Superseded snapshots are kept for as long as any log entry refers to them, so the version in your record always points at something.
If classification or retrieval is unavailable, the question is refused with an explicit message and escalation is offered. Guessing a route would breach the route-up invariant in exactly the case where nothing is known about the question.
Regulation 19(4)(c) requires you to assess the risk of adopting it and record that in writing. It is the obligation most often missed - not through carelessness, but because nobody consciously decided to adopt anything. A Technological Due Diligence document is supplied for Companion.
None of it is technically hard. A policy on what may be asked. A standing context document. A redaction step. A log against every client. A review each time guidance moves.
Every minute of it is overhead. It competes with client work for the attention of the person least able to spare it, and it loses that argument quietly, every time.
Not a project with an end. A discipline applied to every question, by every person, indefinitely. A deadline lands, somebody skips the log, and nothing happens - which is precisely the problem.
Controls that depend on sustained manual effort do not fail loudly. They erode, unevenly, and stay eroded until something is audited. By then the question is how many months are missing.
None of this is expensive to prevent. All of it is expensive to repair - and parts of it cannot be repaired at any price.
No prompt discipline to learn and no new system to adopt. Companion sits inside the practice portal you already use for verification and payment.
In the portal, against a client or matter where there is one. No special phrasing required.
Type and signals, before anything is composed. Identifying detail is removed before the question travels.
An answer, an answer with a flag, a narrowing question, a route to your nominated officer, or a decline with where it belongs.
Question, classification, response and date - filed, whether anyone remembers to log it or not.
Where a question needs a named professional, an escalation can be placed with an approved compliance consultant through the portal, under that consultant’s own indemnity, and billed per engagement. There is no retainer behind it.
A compliance service claiming no limits should be read carefully. These are the edges, stated plainly.
The lowest-risk way to evaluate this is not to change anything. Take the next compliance question that comes up - the one that would otherwise go to a chatbot, a colleague, or Tuesday’s call - and ask it in the portal instead.